ABSTRACT - ISIP



AUTOMATED IDENTIFICATION OF ABNORMAL ADULT EEGs______________________________________________________________________________A Thesis Proposal______________________________________________________________________________in Partial Fulfillmentof the Requirements for the Degree ofMaster of Science in Electrical and Computer EngineeringBySilvia López de DiegoDecember, 2016_________________________________ Dr. Joseph PiconeThesis Advisor__________________________________________________________________ Dr. Pallavi Chitturi Dr. Iyad Obeid Professor of StatisticalProfessor of Electrical & Science Computer Engineering Committee Member Committee MemberABSTRACTThe interpretation of electroencephalograms (EEGs) is a process that is still dependent on the subjective analysis of the examiners. Though inter-rater agreement on critical events such as seizures can be high, it is much lower on subtler events (e.g., when there are benign variants). The focus of this study is to automatically classify normal and abnormal EEGs to provide neurologists with real-time decision support.A demographically balanced subset of the TUH EEG Corpus was used to evaluate performance. The data, comprised of 200 abnormal EEGs and 202 normal EEGs was manually selected. This subset was partitioned into a training set (82 normal/80 abnormal) and an evaluation set (51 normal/55 abnormal). Principal Components Analysis (PCA) was used to reduce the dimensionality of the data. Two baseline classification algorithms were explored: k-Nearest Neighbor (kNN), Random Forest Ensemble Learning (RF). kNN achieved a 41.8% detection error rate while RF achieved an error rate of 31.7%. These error rates are significantly lower than those obtained by random guessing based on priors (49.5%). These algorithms were then compared to a Hidden Markov Models (HMM) based approach, which reduced the error rate to 17.0%, which is approaching human performance. Several deep learning architectures will also be explored in this thesis.TABLE OF CONTENTS TOC \o "1-3" \h \z \u ABSTRACT PAGEREF _Toc468487701 \h iiTABLE OF CONTENTS PAGEREF _Toc468487702 \h iiiLIST OF TABLES PAGEREF _Toc468487703 \h vLIST OF FIGURES PAGEREF _Toc468487704 \h viCHAPTER 1 PAGEREF _Toc468487705 \h 1INTRODUCTION PAGEREF _Toc468487706 \h 1The Normal Adult EEG PAGEREF _Toc468487707 \h 2Visual Analysis of EEGs PAGEREF _Toc468487708 \h 4Automatic Abnormal EEG Classification PAGEREF _Toc468487709 \h 5Thesis Overview PAGEREF _Toc468487710 \h 6CHAPTER 2 PAGEREF _Toc468487711 \h 7CLASSIFICATION OF SEQUENTIAL DATA PAGEREF _Toc468487712 \h 72.1 Markov Models PAGEREF _Toc468487713 \h 72.1.1 Hidden Markov Models (HMMs) PAGEREF _Toc468487714 \h 82.1.2 Gaussian Mixture Models (GMMs) PAGEREF _Toc468487715 \h 102.2 Performance of GMMs-HMMs Compared with Deep Neural Networks (DNNs) PAGEREF _Toc468487716 \h 11CHAPTER 3 PAGEREF _Toc468487717 \h 12DATA & EXPERIMENTS PAGEREF _Toc468487718 \h 123.1 Data PAGEREF _Toc468487719 \h 123.1.1 Data Subset Selection PAGEREF _Toc468487720 \h 123.1.2 Feature Extraction PAGEREF _Toc468487721 \h 133.1.3 Dimensionality Reduction PAGEREF _Toc468487722 \h 153.2 Experiments PAGEREF _Toc468487723 \h 153.2.1 Baseline Experiments PAGEREF _Toc468487724 \h 163.1.2 HMM Experiments PAGEREF _Toc468487725 \h 17CHAPTER 4 PAGEREF _Toc468487726 \h 18PRELIMINARY EXPERIMENTS AND RESULTS PAGEREF _Toc468487727 \h 184.1 Baseline Results PAGEREF _Toc468487728 \h 184.1 GMM-HMM Results PAGEREF _Toc468487729 \h 21CHAPTER 5 PAGEREF _Toc468487730 \h 24EXPECTATIONS AND FUTURE WORK PAGEREF _Toc468487731 \h 245.1 Expected Outcomes PAGEREF _Toc468487732 \h 245.2Timeline for Future Work PAGEREF _Toc468487733 \h 24REFERENCES PAGEREF _Toc468487734 \h 26LIST OF TABLES TOC \c "Table" Table 1. Summary of word error rates for a subspace Gaussian Mixture Model and a Deep Neural Network. PAGEREF _Toc468487735 \h 11Table 2. Comparison of the performance obtained with the two baseline systems PAGEREF _Toc468487736 \h 21Table 3. Confusion matrix for the kNN system PAGEREF _Toc468487737 \h 21Table 4. GMM-HMM correct detection rate for various numbers of Gaussian PAGEREF _Toc468487738 \h 22Table 5. Correct detection rate for different signal input lengths PAGEREF _Toc468487739 \h 22Table 6. Correct detection rate for different channels PAGEREF _Toc468487740 \h 23Table 8. Confusion matrix for the best GMM-HMM system PAGEREF _Toc468487741 \h 23Table 7. Summary of the performance for all the evaluated systems PAGEREF _Toc468487742 \h 23LIST OF FIGURES TOC \c "Figure" Figure 1. Summary of the common steps that are followed for a clinical EEG recording and interpretation PAGEREF _Toc468487743 \h 2Figure 2. Decission tree that shows the process that is generally followed for the abnormal identification of EEGs PAGEREF _Toc468487744 \h 5Figure 3. Temporal Evolution of a seizure in the T4-A2 channel of an EEG. The top of the figure shows the spectrogram of the signal, while the bottom panel shows the signal in the time domain. PAGEREF _Toc468487745 \h 7Figure 4. Example of a basic Markov model with states ωi and transition probabilities aij PAGEREF _Toc468487746 \h 8Figure 5. Example of a Hidden Markov model with states ωi, transition probabilities aij, emission probabilities bjk and visible stated vk. PAGEREF _Toc468487747 \h 9Figure 6. HMM based phone model with transition probabilities aij and observation distributions bj ( ) PAGEREF _Toc468487748 \h 9Figure 7. Distribution of the patients’ ages and genders for the selected dataset PAGEREF _Toc468487749 \h 12Figure 8. Illustration of the base feature extraction process. PAGEREF _Toc468487750 \h 14Figure 9. Location of studied channels in the 10-20 standard system of electrode placement for the TCP montage PAGEREF _Toc468487751 \h 17Figure 10. Normal/abnormal classification error rate as a function of number of (trees Nt) PAGEREF _Toc468487752 \h 18Figure 11. Error rate of the kNN and RF algorithms as a function of the PCA dimension PAGEREF _Toc468487753 \h 19Figure 12. Error rate as a function of the number of neighbors k for PCA dimension of 20 and 86 PAGEREF _Toc468487754 \h 20Figure 13. Classification error rate (for kNN) for a fronto-central (F4-C4) and a temporal-occipital (T5-O1) channel PAGEREF _Toc468487755 \h 20CHAPTER 1 INTRODUCTIONThe recording of the neurons’ electrical activity along the scalp, known as electroencephalography (EEG), has been widely used for the diagnosis and management of conditions such as sleep disorders and epilepsy in the past 30 years. Despite the emergence of new technologies, such as Magnetic Resonance Imaging (MRI), the noninvasiveness and relative inexpensiveness of EEGs make this technique a popular choice as a diagnostics tool among physicians ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "DOI" : "10.1136/jnnp.2005.069245", "ISSN" : "0022-3050", "author" : [ { "dropping-particle" : "", "family" : "Smith", "given" : "S", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "container-title" : "Journal of Neurology, Neurosurgery, and Psychiatry", "id" : "ITEM-1", "issue" : "Suppl 2", "issued" : { "date-parts" : [ [ "2005", "6" ] ] }, "page" : "ii2-ii7", "publisher" : "BMJ Group", "title" : "EEG in the diagnosis, classification, and management of patients with epilepsy", "type" : "article-journal", "volume" : "76" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Smith, 2005)", "plainTextFormattedCitation" : "(Smith, 2005)", "previouslyFormattedCitation" : "(Smith, 2005)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Smith, 2005). A typical routine EEG has a duration of about 20 minutes, duration that is not always enough to record ictal (or interictal activity) in patients with seizure disorders. As a matter of fact, only 50% of patients with epilepsy show interictal epileptiform discharges (IED) in their first recording, and increasing the amount of recordings up to four seems to increase the yield of the tests ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "DOI" : "10.1136/jnnp.2005.069245", "ISSN" : "0022-3050", "author" : [ { "dropping-particle" : "", "family" : "Smith", "given" : "S", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "container-title" : "Journal of Neurology, Neurosurgery, and Psychiatry", "id" : "ITEM-1", "issue" : "Suppl 2", "issued" : { "date-parts" : [ [ "2005", "6" ] ] }, "page" : "ii2-ii7", "publisher" : "BMJ Group", "title" : "EEG in the diagnosis, classification, and management of patients with epilepsy", "type" : "article-journal", "volume" : "76" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Smith, 2005)", "plainTextFormattedCitation" : "(Smith, 2005)", "previouslyFormattedCitation" : "(Smith, 2005)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Smith, 2005). In this sense, the diagnosis and characterization of epilepsy usually require more than one routine EEG and/or long term monitoring (LTM) EEG recordings, which last for more extended periods of time (hours to days). EEG records are interpreted by board certified physicians, process that, because of the demand for EEGs and the time that takes for their reading, can introduce a lag time, which ranges from days to weeks, to the diagnosis and decision making process. Additionally, reading EEGs depends heavily on the subjective judgement of the reader, which can lead to missed events or misdiagnosis of the patient ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "DOI" : "10.1046/j.1440-1819.2003.01152.x", "ISSN" : "1440-1819", "author" : [ { "dropping-particle" : "", "family" : "Azuma", "given" : "Hideki", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Hori", "given" : "Shiro", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Nakanishi", "given" : "Masao", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Fujimoto", "given" : "Shinji", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Ichikawa", "given" : "Norimasa", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Furukawa", "given" : "Toshiaki A", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "container-title" : "Psychiatry and Clinical Neurosciences", "id" : "ITEM-1", "issue" : "5", "issued" : { "date-parts" : [ [ "2003", "10", "1" ] ] }, "page" : "485-489", "publisher" : "Blackwell Science Pty", "title" : "An intervention to improve the interrater reliability of clinical EEG interpretations", "type" : "article-journal", "volume" : "57" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Azuma et al., 2003)", "plainTextFormattedCitation" : "(Azuma et al., 2003)", "previouslyFormattedCitation" : "(Azuma et al., 2003)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Azuma et al., 2003). Introducing a certain level of automation to the EEG interpretation task could potentially serve as an aid for the neurologists to speed up the reading process and ease some of the pressure that results from the high demand of EEGs by patients that are in the need of a diagnosis or the management of their conditions. The EEG recording workflow involves the placement of the electrodes on the patient’s scalp (for scalp EEG) by the EEG technicians, the acquisition of the EEG data, the interpretation of the signals by a certified neurologists and the generation of the report that is presented to the patient ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "DOI" : "", "author" : [ { "dropping-particle" : "", "family" : "Harati", "given" : "Amir", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Lopez", "given" : "Silvia", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Obeid", "given" : "Iyad", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Jacobson", "given" : "Mercedes", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Tobochnik", "given" : "Steven", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Picone", "given" : "Joseph", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "container-title" : "Proceedings of the IEEE Signal Processing in Medicine and Biology Symposium", "id" : "ITEM-1", "issued" : { "date-parts" : [ [ "2014" ] ] }, "page" : "1-5", "publisher-place" : "Philadelphia, Pennsylvania, USA", "title" : "The TUH EEG Corpus: A Big Data Resource for Automated EEG Interpretation", "type" : "paper-conference" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Amir Harati et al., 2014a)", "plainTextFormattedCitation" : "(Amir Harati et al., 2014a)", "previouslyFormattedCitation" : "(Amir Harati et al., 2014a)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Amir Harati et al., 2014a). The EEG report contains a combination of the history, medications, description of the record and interesting findings. One portion of the report, however, contains the impression of the record, which shows whether the EEG is normal or abnormal given the EEG activity recorded in the session. REF _Ref465286254 \h \* MERGEFORMAT Figure 1 summarizes the right1492250Figure SEQ Figure \* ARABIC 1. Summary of the common steps that are followed for a clinical EEG recording and interpretation00Figure SEQ Figure \* ARABIC 1. Summary of the common steps that are followed for a clinical EEG recording and interpretationprocedure for a typical EEG recording.The medical report that is produced for each EEG session describes the record, the recording conditions and summarizes the findings. One decision that is also shown in the report is whether the characteristics of the EEG was found to be within the normal limits for patients in a similar group of age and gender. The main objective for this study is to utilize machine learning techniques in order to generate the Normal/Abnormal decision automatically. If the proposed technology reaches clinically accepted performance, it could potentially serve as an aid to neurologists during the EEG interpretation task and reduce the lag time between EEG recording and reporting to the patients, establishing a more efficient workflow.The Normal Adult EEGThe EEG interpretation task can be broken down into two different parts: the analysis of the EEG background and the recognition of the transients ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "DOI" : "10.1016/j.clinph.2012.07.003", "PMID" : "22858178", "abstract" : "Investigations of (sub-)acute ischaemic stroke (IS) employing quantitative electroencephalographic (QEEG) methods, in concert with other assessments, are reviewed. Numerous outcomes from hundreds of patients collectively indicate that (sub-)acute QEEG indices from standard systems can uniquely inform clinical management, particularly prognostication of outcomes from IS. Two classes of QEEG indices have proven particularly informative. The first quantifies the power of abnormal, slow activity relative to that of faster activity and the second, interhemispheric voltage asymmetry (broadband). Both have been identified as statistically significant predictors of outcomes assessed (via routine clinical scales) in the weeks and months following IS. Furthermore both have demonstrated higher predictive value than concomitant neurological assessments and scales, and to improve upon outcome prediction afforded by neuroimaging alone. These indices also may continuously provide unique, real-time insights into the efficacy of thrombolytic therapy, prior to clinical changes. Two key applications of QEEG which should prove valuable for future clinical management of IS are: (1) continuous, acute monitoring to inform about the efficacy of thrombolysis and decisions about potential additional interventions, and; (2) brief, subacute recording to inform outcome prognostication and clinical decisions about, for example, rehabilitation strategies. Ongoing research and technological developments will continue to facilitate clinical translation of QEEG investigations reviewed herein.", "author" : [ { "dropping-particle" : "", "family" : "Finnigan", "given" : "Simon", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Putten", "given" : "Michel J a M", "non-dropping-particle" : "van", "parse-names" : false, "suffix" : "" } ], "container-title" : "Clinical neurophysiology : official journal of the International Federation of Clinical Neurophysiology", "id" : "ITEM-1", "issue" : "1", "issued" : { "date-parts" : [ [ "2013", "1" ] ] }, "page" : "10-19", "publisher" : "International Federation of Clinical Neurophysiology", "title" : "EEG in ischaemic stroke: quantitative EEG can uniquely inform (sub-)acute prognoses and clinical management.", "type" : "article-journal", "volume" : "124" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Finnigan & van Putten, 2013)", "plainTextFormattedCitation" : "(Finnigan & van Putten, 2013)", "previouslyFormattedCitation" : "(Finnigan & van Putten, 2013)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Finnigan & van Putten, 2013). The background pattern refers to the general characteristics of an EEG, which include the features that neurologists observe when making a normal/abnormal decision about the record. Some remarkable examples of the background pattern are the posterior dominant rhythm (PDR) and the frequency distributions of the signals throughout the scalp ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "DOI" : "10.1016/j.clinph.2012.07.007", "ISSN" : "13882457", "PMID" : "22917580", "abstract" : "Objective: Visual interpretation of EEG is time-consuming and not always consistent between reviewers. Our objective is to improve this by introducing guidelines and algorithms to quantify various properties, focussing on the background pattern in adult EEGs. Methods: Five common properties were evaluated: (i) alpha rhythm frequency; (ii) reactivity; (iii) anterio-posterior gradients; (iv) asymmetries; and (v) diffuse slow-wave activity. A formal description was found for each together with a guideline and proposed quantitative algorithm. All five features were automatically extracted from routine EEG recordings. Modified time-frequency plots were calculated to summarize spectral and spatial characteristics. Visual analysis scores were obtained from diagnostic reports. Results: Automated feature extraction was applied to 384 routine EEGs. Inter-rater agreement was calculated between visual and quantitative analysis using Fleiss' kappa: ??={(i) 0.60; (ii) 0.35; (iii) 0.19; (iv) 0.12; (v) 0.76}. The method is further illustrated with three representative examples of automated reports. Conclusions: Automated feature extraction of several background EEG properties seems feasible. Inter-rater agreement differed between various features, ranging from slight to substantial. This may be related to the nature of various guidelines and inconsistencies in visual interpretation. Significance: Formal descriptions, standardized terminology, and quantitative analysis may improve inter-rater reliability in reporting of the EEG background pattern and contribute to more efficient and consistent interpretations. ?? 2012 International Federation of Clinical Neurophysiology.", "author" : [ { "dropping-particle" : "", "family" : "Lodder", "given" : "Shaun S.", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Putten", "given" : "Michel J A M", "non-dropping-particle" : "van", "parse-names" : false, "suffix" : "" } ], "container-title" : "Clinical Neurophysiology", "id" : "ITEM-1", "issue" : "2", "issued" : { "date-parts" : [ [ "2013" ] ] }, "page" : "228-237", "title" : "Quantification of the adult EEG background pattern", "type" : "article-journal", "volume" : "124" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Lodder & van Putten, 2013)", "plainTextFormattedCitation" : "(Lodder & van Putten, 2013)", "previouslyFormattedCitation" : "(Lodder & van Putten, 2013)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Lodder & van Putten, 2013). The transient patterns, on the other hand, refer to rarer events that include pathological and physiological waveforms, such as spikes and sharp waves discharges.The Characterization of the normal adult EEG has been based on a specific description of the background pattern and the presence—or lack thereof—of certain transient waveforms given the patient’s state of consciousness (awake EEG is different from drowsy EEG or comatose EEG). The main background characteristics of the normal adult EEG can be summarized as follows ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "ISBN" : "978-1451131956", "author" : [ { "dropping-particle" : "", "family" : "Ebersole", "given" : "J S", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Pedley", "given" : "T A", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "edition" : "4th", "id" : "ITEM-1", "issued" : { "date-parts" : [ [ "2014" ] ] }, "number-of-pages" : "664", "publisher" : "Wolters Kluwer", "publisher-place" : "Philadelphia, Pennsylvania, USA", "title" : "Current practice of clinical electroencephalography", "type" : "book" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Ebersole & Pedley, 2014)", "plainTextFormattedCitation" : "(Ebersole & Pedley, 2014)", "previouslyFormattedCitation" : "(Ebersole & Pedley, 2014)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Ebersole & Pedley, 2014):Reactivity: Refers to the response to certain physiological changes. This changes could be eye opening and closing, sensory stimulation, etc.Alpha Rhythm: This rhythm is the starting point for the visual analysis of EEGs. The presence, characteristics according to age and reactivity of this feature play an important role in the normal/abnormal classification of the EEG. The Alpha waves originate (predominantly) in the occipital lobe and are between 8-13 Hz in frequency and 15 to 45 μV in amplitude.Mu Rhythm: It is a central rhythm of frequencies between 8 to 10 Hz with amplitudes comparable to the alpha rhythm. This rhythm is suppressed unilaterally by the movement of the opposite extremity. This rhythm, however, is also suppressed by conditions such as fatigue, somatosensory and sensorimotor stimulation. In this sense, the Mu rhythm is not always detectable.Beta Activity: Rhythm with frequencies of 18-25 Hz, 14-16 Hz and 35-40 Hz, with amplitudes between 5 and 20 mV. It is important to note, however, that it is rare to see activity higher than 25 Hz on scalp EEGs.Theta Activity: Normal adults tend to show traces of less than 15 μV 6-7 Hz activity in the frontal and frontocentral regions and occasionally in the midline central region. This rhythm, called Theta activity, usually becomes sustained and higher in voltage with the onset of drowsiness. These features provide a description of the characteristics that are systematically observed by neurologists when evaluating an EEG. Characteristics such as the state and the age of the patient are also important considerations that may alter the characteristics presented above. These characteristics are mainly observed in normal adult EEGs ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "ISBN" : "978-1451131956", "author" : [ { "dropping-particle" : "", "family" : "Ebersole", "given" : "J S", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Pedley", "given" : "T A", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "edition" : "4th", "id" : "ITEM-1", "issued" : { "date-parts" : [ [ "2014" ] ] }, "number-of-pages" : "664", "publisher" : "Wolters Kluwer", "publisher-place" : "Philadelphia, Pennsylvania, USA", "title" : "Current practice of clinical electroencephalography", "type" : "book" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Ebersole & Pedley, 2014)", "plainTextFormattedCitation" : "(Ebersole & Pedley, 2014)", "previouslyFormattedCitation" : "(Ebersole & Pedley, 2014)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Ebersole & Pedley, 2014). Visual Analysis of EEGsThe previous section described the general features that characterize a normal EEG. The presence of these features does not necessarily guarantee the normality of the record. As it was explained before, these characteristics comprise what is defined as the background EEG, without taking into account transient patterns that could additionally be present in the record.Neurologists analyze the records by evaluating the background EEG and determining whether the patient presents normal characteristics according to his or her state. If the patient did not present abnormal transients during the recording, and the background EEG was within normal limits, the recording is considered normal. The analysis of the background is broken down in steps that allow to take all of the characteristics into account in a systematic way. The analysis steps that ultimately lead to a decision about the normality of the record involve the observation of the following characteristics: frequency, voltage, waveform, regulation (make sure the alpha rhythm does not vary more than ±0.5 Hz ), locus, reactivity and interhemispheric coherence ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "ISBN" : "978-1451131956", "author" : [ { "dropping-particle" : "", "family" : "Ebersole", "given" : "J S", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Pedley", "given" : "T A", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "edition" : "4th", "id" : "ITEM-1", "issued" : { "date-parts" : [ [ "2014" ] ] }, "number-of-pages" : "664", "publisher" : "Wolters Kluwer", "publisher-place" : "Philadelphia, Pennsylvania, USA", "title" : "Current practice of clinical electroencephalography", "type" : "book" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Ebersole & Pedley, 2014)", "plainTextFormattedCitation" : "(Ebersole & Pedley, 2014)", "previouslyFormattedCitation" : "(Ebersole & Pedley, 2014)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Ebersole & Pedley, 2014). The systematic visual analysis of EEGs usually starts with the evaluation of the occipital alpha rhythm, also called the Posterior Dominant Rhythm. This activity emerges in the occipital region when the eyes are closed, and fades as the patients enter a state of drowsiness. In this sense, the evaluation of the reactivity for the emergence of the PDR is one of the main features that are evaluated in order to make a decision about the normality of the record ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "ISBN" : "978-1451131956", "author" : [ { "dropping-particle" : "", "family" : "Ebersole", "given" : "J S", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Pedley", "given" : "T A", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "edition" : "4th", "id" : "ITEM-1", "issued" : { "date-parts" : [ [ "2014" ] ] }, "number-of-pages" : "664", "publisher" : "Wolters Kluwer", "publisher-place" : "Philadelphia, Pennsylvania, USA", "title" : "Current practice of clinical electroencephalography", "type" : "book" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Ebersole & Pedley, 2014)", "plainTextFormattedCitation" : "(Ebersole & Pedley, 2014)", "previouslyFormattedCitation" : "(Ebersole & Pedley, 2014)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Ebersole & Pedley, 2014). A decision tree for the evaluation of the normality of an EEG record is presented in REF _Ref466135461 \* MERGEFORMAT Figure 2 ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "DOI" : "", "author" : [ { "dropping-particle" : "", "family" : "Lopez", "given" : "Silvia", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Suarez", "given" : "Gabriella", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Jungries", "given" : "Dave", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Obeid", "given" : "Iyad", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Picone", "given" : "Joseph", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "container-title" : "IEEE Signal Processing in Medicine and Biology Symposium", "id" : "ITEM-1", "issued" : { "date-parts" : [ [ "2015" ] ] }, "page" : "1-4", "publisher-place" : "Philadelphia, Pennsylvania, USA", "title" : "Automated Identification of Abnormal EEGs", "type" : "paper-conference" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Lopez, Suarez, Jungries, Obeid, & Picone, 2015)", "plainTextFormattedCitation" : "(Lopez, Suarez, Jungries, Obeid, & Picone, 2015)", "previouslyFormattedCitation" : "(Lopez, Suarez, Jungries, Obeid, & Picone, 2015)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Lopez, Suarez, Jungries, Obeid, & Picone, 2015).right0Figure SEQ Figure \* ARABIC 2. Decission tree that shows the process that is generally followed for the abnormal identification of EEGs00Figure SEQ Figure \* ARABIC 2. Decission tree that shows the process that is generally followed for the abnormal identification of EEGsIn essence, a common characteristic that is evaluated in adults is the PDR and the reactivity of its occipital emergence. Throughout this study, the characteristics of the background EEG and, especially the PDR, are taken into consideration in order to make a normal/abnormal classification with the implementation of state-of-the-art machine learning techniques.Automatic Abnormal EEG ClassificationA generalized algorithm or method for the classification of clinical abnormal EEGs is a task that has not been yet explored. While some work has been done in the identification of EEG abnormalities specific to certain pathological or physiological conditions, the study of the general background EEG as a resource for the classification of normal and abnormal records has not been investigated. For instance, studies have been done in order to classify athletes with residual functional deficits after the occurrence of a concussion with the help of EEG data and Support Vector Machines (SVMs) ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "DOI" : "10.1109/TNSRE.2008.918422", "ISBN" : "1534-4320", "ISSN" : "1558-0210", "PMID" : "18701381", "abstract" : "There is a growing body of knowledge indicating long-lasting residual electroencephalography (EEG) abnormalities in concussed athletes that may persist up to 10-year postinjury. Most often, these abnormalities are initially overlooked using traditional concussion assessment tools. Accordingly, premature return to sport participation may lead to recurrent episodes of concussion, increasing the risk of recurrent concussions with more severe consequences. Sixty-one athletes at high risk for concussion (i.e., collegiate rugby and football players) were recruited and underwent EEG baseline assessment. Thirty of these athletes suffered from concussion and were retested at day 30 postinjury. A number of task-related EEG recordings were conducted. A novel classification algorithm, the support vector machine (SVM), was applied as a classifier to identify residual functional abnormalities in athletes suffering from concussion using a multichannel EEG data set. The total accuracy of the classifier using the 10 features was 77.1%. The classifier has a high sensitivity of 96.7% (linear SVM), 80.0% (nonlinear SVM), and a relatively lower but acceptable selectivity of 69.1% (linear SVM) and 75.0% (nonlinear SVM). The major findings of this report are as follows: 1) discriminative features were observed at theta, alpha, and beta frequency bands, 2) the minimal redundancy relevance method was identified as being superior to the univariate t -test method in selecting features for the model calculation, 3) the EEG features selected for the classification model are linked to temporal and occipital areas, and 4) postural parameters influence EEG data set and can be used as discriminative features for the classification model. Overall, this report provides sufficient evidence that 10 EEG features selected for final analysis and SVM may be potentially used in clinical practice for automatic classification of athletes with residual brain functional abnormalities following a concussion episode.", "author" : [ { "dropping-particle" : "", "family" : "Cao", "given" : "Cheng", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Tutwiler", "given" : "Richard Laurence", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Slobounov", "given" : "Semyon", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "container-title" : "IEEE Transactions on Neural Systems & Rehabilitation Engineering", "id" : "ITEM-1", "issue" : "4", "issued" : { "date-parts" : [ [ "2008" ] ] }, "page" : "327-335", "title" : "Automatic classification of athletes with residual functional deficits following concussion by means of EEG signal using support vector machine", "type" : "article-journal", "volume" : "16" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Cao, Tutwiler, & Slobounov, 2008)", "plainTextFormattedCitation" : "(Cao, Tutwiler, & Slobounov, 2008)", "previouslyFormattedCitation" : "(Cao, Tutwiler, & Slobounov, 2008)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Cao, Tutwiler, & Slobounov, 2008). These study, however, did not rely on clinical EEG data, and the classifier was designed to train a very specific condition. In this work, baselines for the classification of normal and abnormal clinical EEG records are established. Two non-parametric algorithms, k-Nearest Neighbor and Random Forest Ensemble Learning, explained in more depth in Chapter 2, were used for the establishment of the baseline systems. Hidden Markov Models, also explained in Chapter 2, were then utilized for classification and comparison with the baselines.Thesis OverviewThis study presents the results of the classification of clinical EEGs obtained through a series of pilot studies, where the normal/abnormal classification problem was approached with K-nearest neighbors (KNN), and Random Forest Ensemble learning (RF). In addition, these preliminary studies are compared to a parametric approach based on Hidden Markov Models (HMMs). In Chapter 2, a description of the HMM utilized in this study will be presented in more depth. Chapter 3 will introduce the dataset utilized for this classification study, along with the subset selection process and the experimental setup. In Chapter 4, the results of the preliminary experiments will be presented and discussed. Finally, Chapter 5 will offer an idea of the expected outcomes of this study and a timeline in which the future work will be completed.CHAPTER 2CLASSIFICATION OF SEQUENTIAL DATA right3137535Figure SEQ Figure \* ARABIC 3. Temporal Evolution of a seizure in the T4-A2 channel of an EEG. The top of the figure shows the spectrogram of the signal, while the bottom panel shows the signal in the time domain.00Figure SEQ Figure \* ARABIC 3. Temporal Evolution of a seizure in the T4-A2 channel of an EEG. The top of the figure shows the spectrogram of the signal, while the bottom panel shows the signal in the time domain.Electroencephalography signals, like speech signals, are the product of a physiological process that unfolds in time. REF _Ref467705588 \* MERGEFORMAT Figure 3, for instance, shows the temporal evolution of a seizure in one EEG channel and its respective spectrogram. In this sense, machine learning approaches that treat the observations in the data as independent and identically distributed (i.i.d.) would not successfully exploit the sequential nature of the data ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "ISBN" : "978-0387310732", "author" : [ { "dropping-particle" : "", "family" : "Bishop", "given" : "Christopher", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "edition" : "2nd", "id" : "ITEM-1", "issued" : { "date-parts" : [ [ "2011" ] ] }, "number-of-pages" : "738", "publisher" : "Springer", "publisher-place" : "New York, New York, USA", "title" : "Pattern Recognition and Machine Learning", "type" : "book" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Bishop, 2011)", "plainTextFormattedCitation" : "(Bishop, 2011)", "previouslyFormattedCitation" : "(Bishop, 2011)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Bishop, 2011). The inherent temporality of EEGs and the success that HMMs have shown in the area of speech recognition ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "DOI" : "10.1109/53.54527", "ISSN" : "0740-7467", "author" : [ { "dropping-particle" : "", "family" : "Picone", "given" : "Joseph", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "container-title" : "IEEE ASSP Magazine", "id" : "ITEM-1", "issue" : "3", "issued" : { "date-parts" : [ [ "1990", "7" ] ] }, "page" : "26-41", "title" : "Continuous Speech Recognition Using Hidden Markov Models", "type" : "article-journal", "volume" : "7" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Picone, 1990)", "plainTextFormattedCitation" : "(Picone, 1990)", "previouslyFormattedCitation" : "(Picone, 1990)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Picone, 1990) serve as a motivation to select these models for the decoding and classification of EEG signals. Accordingly, the remainder of this chapter will offer the reader the necessary theoretical background for HMMs and a brief explanation of how these models have been implemented for continuous speech recognition. The following sections mostly follow theoretical explanations that can be found in ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "DOI" : "10.1007/BF01237942", "ISBN" : "0471056693", "ISSN" : "1433-7541", "PMID" : "2630878", "abstract" : "Introduction to Mathematical Techniques in Pattern Recognition by Harry C. Andrews This volume is one of the first cohesive treatments of the use of mathematics for studying interactions between various recognition environments. It brings together techniques previously scattered throughout the literature and provides a concise common notation that will facilitate the understanding and comparison of the many aspects of mathematical pattern recognition. The contents of this volume are divided into five interrelated subject areas: Feature Selection, Distribution Free Classification, Statistical Classification, Nonsupervised Learning, and Sequential Learning. Appendices describing specific aspects of feature selection and extensive reference and bibliographies are included. 1972 253 pp. Threshold Logic and its Applications by Saburo Muroga This is the first in-depth exposition of threshold logic and its applications using linear programming and integer programming as optimization tools. It presents threshold logic as a unified theory of conventional simple gates, threshold gates and their networks. This unified viewpoint explicitly reveals many important properties that were formerly concealed in the framework of conventional switching theory (based essentially on and, or and not gates). 1971 478 pp. Knowing and Guessing A Quantitative Study of Inference and Information By Satosi Watanabe This volume presents a coherent theoretical view of a field now split into different disciplines: philosophy, information science, cybernetics, psychology, electrical engineering, and physics. The target of investigation is the cognitive process of knowing and guessing. In contrast to traditional philosophy, the approach is quantitative rather than qualitative. The study is formal in the sense that the author is not interested in the contents of knowledge or the physiological mechanism of the process of knowing.", "author" : [ { "dropping-particle" : "", "family" : "Duda", "given" : "Richard O.", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Hart", "given" : "Peter E.", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Stork", "given" : "David G.", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "container-title" : "New York: John Wiley, Section", "id" : "ITEM-1", "issued" : { "date-parts" : [ [ "2001" ] ] }, "number-of-pages" : "680", "title" : "Pattern Classification", "type" : "book" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Duda, Hart, & Stork, 2001)", "plainTextFormattedCitation" : "(Duda, Hart, & Stork, 2001)", "previouslyFormattedCitation" : "(Duda, Hart, & Stork, 2001)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Duda, Hart, & Stork, 2001), ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "DOI" : "10.1109/5.18626", "ISSN" : "00189219", "author" : [ { "dropping-particle" : "", "family" : "Rabiner", "given" : "LawrenceR.", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "container-title" : "Proceedings of the IEEE", "id" : "ITEM-1", "issue" : "2", "issued" : { "date-parts" : [ [ "1989" ] ] }, "page" : "257-286", "title" : "A Tutorial on Hidden Markov Models and Selected Applications in Speech Recognition", "type" : "article-journal", "volume" : "77" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Rabiner, 1989)", "plainTextFormattedCitation" : "(Rabiner, 1989)", "previouslyFormattedCitation" : "(Rabiner, 1989)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Rabiner, 1989) and ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "ISBN" : "978-0387310732", "author" : [ { "dropping-particle" : "", "family" : "Bishop", "given" : "Christopher", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "edition" : "2nd", "id" : "ITEM-1", "issued" : { "date-parts" : [ [ "2011" ] ] }, "number-of-pages" : "738", "publisher" : "Springer", "publisher-place" : "New York, New York, USA", "title" : "Pattern Recognition and Machine Learning", "type" : "book" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Bishop, 2011)", "plainTextFormattedCitation" : "(Bishop, 2011)", "previouslyFormattedCitation" : "(Bishop, 2011)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Bishop, 2011).2.1 Markov ModelsIf a sequence of states at subsequent times t are considered, the state at a time t is denoted as ω(t). The description of the model for a specific sequence ωT (where T represents the length of the sequence) is then given by:Pωjt+1ωit=aij( SEQ equation \* MERGEFORMAT 1)right1906270Figure SEQ Figure \* ARABIC 4. Example of a basic Markov model with states ωi and transition probabilities aij00Figure SEQ Figure \* ARABIC 4. Example of a basic Markov model with states ωi and transition probabilities aijwhere aij represents a transition probability, or the probability of being in state ωj at t+1 given that the state at t is ωi ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "DOI" : "10.1007/BF01237942", "ISBN" : "0471056693", "ISSN" : "1433-7541", "PMID" : "2630878", "abstract" : "Introduction to Mathematical Techniques in Pattern Recognition by Harry C. Andrews This volume is one of the first cohesive treatments of the use of mathematics for studying interactions between various recognition environments. It brings together techniques previously scattered throughout the literature and provides a concise common notation that will facilitate the understanding and comparison of the many aspects of mathematical pattern recognition. The contents of this volume are divided into five interrelated subject areas: Feature Selection, Distribution Free Classification, Statistical Classification, Nonsupervised Learning, and Sequential Learning. Appendices describing specific aspects of feature selection and extensive reference and bibliographies are included. 1972 253 pp. Threshold Logic and its Applications by Saburo Muroga This is the first in-depth exposition of threshold logic and its applications using linear programming and integer programming as optimization tools. It presents threshold logic as a unified theory of conventional simple gates, threshold gates and their networks. This unified viewpoint explicitly reveals many important properties that were formerly concealed in the framework of conventional switching theory (based essentially on and, or and not gates). 1971 478 pp. Knowing and Guessing A Quantitative Study of Inference and Information By Satosi Watanabe This volume presents a coherent theoretical view of a field now split into different disciplines: philosophy, information science, cybernetics, psychology, electrical engineering, and physics. The target of investigation is the cognitive process of knowing and guessing. In contrast to traditional philosophy, the approach is quantitative rather than qualitative. The study is formal in the sense that the author is not interested in the contents of knowledge or the physiological mechanism of the process of knowing.", "author" : [ { "dropping-particle" : "", "family" : "Duda", "given" : "Richard O.", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Hart", "given" : "Peter E.", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Stork", "given" : "David G.", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "container-title" : "New York: John Wiley, Section", "id" : "ITEM-1", "issued" : { "date-parts" : [ [ "2001" ] ] }, "number-of-pages" : "680", "title" : "Pattern Classification", "type" : "book" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Duda et al., 2001)", "plainTextFormattedCitation" : "(Duda et al., 2001)", "previouslyFormattedCitation" : "(Duda et al., 2001)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Duda et al., 2001). The state at step t+1 in a first-order Markov model is a function that only depends on t. Higher order Markov chains allow to consider states at earlier steps. So far, an observable Markov model, in which each step corresponds to an observable event, has been described. REF _Ref467778074 /*mergeformat \* MERGEFORMAT Figure 4 shows an illustration of a three-state Markov model, with its respective states ωi represented by nodes and the transition probabilities aij represented by links.So far, the states of the model that has been described correspond to an observable event, which constitutes a restrictive model to be applied to problems such as speech recognition ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "DOI" : "10.1109/5.18626", "ISSN" : "00189219", "author" : [ { "dropping-particle" : "", "family" : "Rabiner", "given" : "LawrenceR.", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "container-title" : "Proceedings of the IEEE", "id" : "ITEM-1", "issue" : "2", "issued" : { "date-parts" : [ [ "1989" ] ] }, "page" : "257-286", "title" : "A Tutorial on Hidden Markov Models and Selected Applications in Speech Recognition", "type" : "article-journal", "volume" : "77" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Rabiner, 1989)", "plainTextFormattedCitation" : "(Rabiner, 1989)", "previouslyFormattedCitation" : "(Rabiner, 1989)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Rabiner, 1989). As a matter of fact, in speech recognition systems, the perceiver does not have access to the Markov model states. On the contrary, spectral properties of the emitted sounds are measured and the outputs are analyzed in visible states v(t), which represent a new set of stochastic processes that produce a sequence of directly accessible observations. This resulting augmented model describes a hidden Markov Model (HMM).2.1.1 Hidden Markov Models (HMMs)Hidden Markov models can be considered an extension, or augmentation, of the models that have been described to this point. In fact, in the case of HMMs, the visible observations v(t) are given by a probabilistic function of the state ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "DOI" : "10.1109/5.18626", "ISSN" : "00189219", "author" : [ { "dropping-particle" : "", "family" : "Rabiner", "given" : "LawrenceR.", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "container-title" : "Proceedings of the IEEE", "id" : "ITEM-1", "issue" : "2", "issued" : { "date-parts" : [ [ "1989" ] ] }, "page" : "257-286", "title" : "A Tutorial on Hidden Markov Models and Selected Applications in Speech Recognition", "type" : "article-journal", "volume" : "77" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Rabiner, 1989)", "plainTextFormattedCitation" : "(Rabiner, 1989)", "previouslyFormattedCitation" : "(Rabiner, 1989)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Rabiner, 1989). In this augmented model, pictorially represented in REF _Ref467791579 /*mergeformat \* MERGEFORMAT Figure 5, the assumption that at every single time t the system is at state ω(t) is kept. However, for HMMs, the assumption that the system also emits a visible observation or symbol v(t) is also made. In this way (assuming a discrete symbol is emitted at each state), a probability of emitting a specific visible state vk(t) is given by ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "DOI" : "10.1007/BF01237942", "ISBN" : "0471056693", "ISSN" : "1433-7541", "PMID" : "2630878", "abstract" : "Introduction to Mathematical Techniques in Pattern Recognition by Harry C. Andrews This volume is one of the first cohesive treatments of the use of mathematics for studying interactions between various recognition environments. It brings together techniques previously scattered throughout the literature and provides a concise common notation that will facilitate the understanding and comparison of the many aspects of mathematical pattern recognition. The contents of this volume are divided into five interrelated subject areas: Feature Selection, Distribution Free Classification, Statistical Classification, Nonsupervised Learning, and Sequential Learning. Appendices describing specific aspects of feature selection and extensive reference and bibliographies are included. 1972 253 pp. Threshold Logic and its Applications by Saburo Muroga This is the first in-depth exposition of threshold logic and its applications using linear programming and integer programming as optimization tools. It presents threshold logic as a unified theory of conventional simple gates, threshold gates and their networks. This unified viewpoint explicitly reveals many important properties that were formerly concealed in the framework of conventional switching theory (based essentially on and, or and not gates). 1971 478 pp. Knowing and Guessing A Quantitative Study of Inference and Information By Satosi Watanabe This volume presents a coherent theoretical view of a field now split into different disciplines: philosophy, information science, cybernetics, psychology, electrical engineering, and physics. The target of investigation is the cognitive process of knowing and guessing. In contrast to traditional philosophy, the approach is quantitative rather than qualitative. The study is formal in the sense that the author is not interested in the contents of knowledge or the physiological mechanism of the process of knowing.", "author" : [ { "dropping-particle" : "", "family" : "Duda", "given" : "Richard O.", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Hart", "given" : "Peter E.", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Stork", "given" : "David G.", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "container-title" : "New York: John Wiley, Section", "id" : "ITEM-1", "issued" : { "date-parts" : [ [ "2001" ] ] }, "number-of-pages" : "680", "title" : "Pattern Classification", "type" : "book" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Duda et al., 2001)", "plainTextFormattedCitation" : "(Duda et al., 2001)", "previouslyFormattedCitation" : "(Duda et al., 2001)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Duda et al., 2001): right565150Figure SEQ Figure \* ARABIC 5. Example of a Hidden Markov model with states ωi, transition probabilities aij, emission probabilities bjk and visible stated vk.00Figure SEQ Figure \* ARABIC 5. Example of a Hidden Markov model with states ωi, transition probabilities aij, emission probabilities bjk and visible stated vk.Pvktωjt=bjk( SEQ equation \* MERGEFORMAT 2)In speech recognition, each spoken word w is decomposed into a sequence of Kw sounds (or base phones), which have pronunciation sequences q1:Kw(w)=q1,…,qKw. The likelihood of a word w given an acoustic feature vector Y is given by:w=argmaxw{P(w|Y)}( SEQ equation \* MERGEFORMAT 3)Given the modeling difficulty for P(w|Y), Bayes Rule can be used to transform this equation into the equation specified by right552421Figure SEQ Figure \* ARABIC 6. HMM based phone model with transition probabilities aij and observation distributions bj ( ) 00Figure SEQ Figure \* ARABIC 6. HMM based phone model with transition probabilities aij and observation distributions bj ( ) w=argmaxw{P(Y|w)P(w)}( SEQ equation \* MERGEFORMAT 4)where P(Y|w) represents the acoustic model and P(w) represents the language model ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "DOI" : "10.1561/2000000004", "ISSN" : "1932-8346", "author" : [ { "dropping-particle" : "", "family" : "Gales", "given" : "Mark", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Young", "given" : "Steve", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "container-title" : "Foundations and Trends\u00ae in Signal Processing", "id" : "ITEM-1", "issue" : "3", "issued" : { "date-parts" : [ [ "2007" ] ] }, "page" : "195-304", "title" : "The Application of Hidden Markov Models in Speech Recognition", "type" : "article-journal", "volume" : "1" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Gales & Young, 2007)", "plainTextFormattedCitation" : "(Gales & Young, 2007)", "previouslyFormattedCitation" : "(Gales & Young, 2007)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Gales & Young, 2007). For the explanation of pertinent concepts, the focus will be centered around the acoustic model.If the decoding of the word “bat” is considered, for example, each of the valid pronunciations for the phones that comprise the word (/b/, /ae/ and /t/) would be represented by a continuous density HMM of the form shown by REF _Ref467846853 \* MERGEFORMAT Figure 6 ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "DOI" : "10.1561/2000000004", "ISSN" : "1932-8346", "author" : [ { "dropping-particle" : "", "family" : "Gales", "given" : "Mark", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Young", "given" : "Steve", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "container-title" : "Foundations and Trends\u00ae in Signal Processing", "id" : "ITEM-1", "issue" : "3", "issued" : { "date-parts" : [ [ "2007" ] ] }, "page" : "195-304", "title" : "The Application of Hidden Markov Models in Speech Recognition", "type" : "article-journal", "volume" : "1" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Gales & Young, 2007)", "plainTextFormattedCitation" : "(Gales & Young, 2007)", "previouslyFormattedCitation" : "(Gales & Young, 2007)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Gales & Young, 2007), and the likelihood P(Y|w) would be given by P(Y|w)=Qp(Y|Q)P(Q|w)( SEQ equation \* MERGEFORMAT 5)where Q represents a sequence of valid pronunciations. If the assumption of a single multivariate Gaussian is made for the output distribution, then bjy would be given by:bjy=N(y;μj,(j))( SEQ equation \* MERGEFORMAT 6)where μj is the mean of state ωj and Σ(j) represents its covariance. In this sense, the acoustic likelihood is described as follows:pYQ=θpθ,YQ=aθ0θ1t=1Tbθtytaθtθt+1 ( SEQ equation \* MERGEFORMAT 7)where θ=θ0,…,θT+1 represents a state sequence through the model ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "DOI" : "10.1561/2000000004", "ISSN" : "1932-8346", "author" : [ { "dropping-particle" : "", "family" : "Gales", "given" : "Mark", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Young", "given" : "Steve", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "container-title" : "Foundations and Trends\u00ae in Signal Processing", "id" : "ITEM-1", "issue" : "3", "issued" : { "date-parts" : [ [ "2007" ] ] }, "page" : "195-304", "title" : "The Application of Hidden Markov Models in Speech Recognition", "type" : "article-journal", "volume" : "1" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Gales & Young, 2007)", "plainTextFormattedCitation" : "(Gales & Young, 2007)", "previouslyFormattedCitation" : "(Gales & Young, 2007)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Gales & Young, 2007).The parameters for the acoustic model are commonly estimated through the forward-backward algorithm as explained in ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "DOI" : "10.1214/aoms/1177697196", "ISSN" : "0003-4851", "author" : [ { "dropping-particle" : "", "family" : "Baum", "given" : "Leonard E.", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Petrie", "given" : "Ted", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Soules", "given" : "George", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Weiss", "given" : "Norman", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "container-title" : "The Annals of Mathematical Statistics", "id" : "ITEM-1", "issue" : "1", "issued" : { "date-parts" : [ [ "1970", "2" ] ] }, "page" : "164-171", "title" : "A Maximization Technique Occurring in the Statistical Analysis of Probabilistic Functions of Markov Chains", "type" : "article-journal", "volume" : "41" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Baum, Petrie, Soules, & Weiss, 1970)", "plainTextFormattedCitation" : "(Baum, Petrie, Soules, & Weiss, 1970)", "previouslyFormattedCitation" : "(Baum, Petrie, Soules, & Weiss, 1970)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Baum, Petrie, Soules, & Weiss, 1970). In general, this approach, which is a generalized instance of the Expectation Maximization (EM) algorithm, updates the weights of the system to better explain the observed training sequences. 2.1.2 Gaussian Mixture Models (GMMs)The explanations that have been presented so far rely on the fact that a single Gaussian distribution models the state—output distributions. In problems like speech recognition, or electroencephalography in the present case, the utilization of a single Gaussian distribution is not necessary accurate, since this implies the assumption that the feature vectors are symmetric and unimodal. Variations such as speaker identity, accent, gender and others make this assumption rarely possible in practice ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "DOI" : "10.1561/2000000004", "ISSN" : "1932-8346", "author" : [ { "dropping-particle" : "", "family" : "Gales", "given" : "Mark", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Young", "given" : "Steve", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "container-title" : "Foundations and Trends\u00ae in Signal Processing", "id" : "ITEM-1", "issue" : "3", "issued" : { "date-parts" : [ [ "2007" ] ] }, "page" : "195-304", "title" : "The Application of Hidden Markov Models in Speech Recognition", "type" : "article-journal", "volume" : "1" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Gales & Young, 2007)", "plainTextFormattedCitation" : "(Gales & Young, 2007)", "previouslyFormattedCitation" : "(Gales & Young, 2007)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Gales & Young, 2007). To overcome this issue, several systems have successfully implemented mixtures of Gaussians ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "DOI" : "10.1109/79.536824", "ISSN" : "1053-5888", "author" : [ { "dropping-particle" : "", "family" : "Steve Young", "given" : "", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "container-title" : "IEEE Signal Processing Magazine", "id" : "ITEM-1", "issue" : "5", "issued" : { "date-parts" : [ [ "1996", "9" ] ] }, "page" : "45", "title" : "A review of large-vocabulary continuous-speech", "type" : "article-journal", "volume" : "13" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Steve Young, 1996)", "plainTextFormattedCitation" : "(Steve Young, 1996)", "previouslyFormattedCitation" : "(Steve Young, 1996)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Steve Young, 1996), which are able to properly, and more accurately, model multi-modal data. If Gaussian mixture models are implemented, then the value for bj(y) would be given by:bjy=m=1McjmN(y;μjm,(jm))( SEQ equation \* MERGEFORMAT 8)where cjm represents the prior probability for mixture component m of state ωj. Commonly, the number of Gaussian mixtures is selected through the testing of models with different number of components and their evaluation in a held out set in order to find the optimal model ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "DOI" : "10.1561/2000000004", "ISSN" : "1932-8346", "author" : [ { "dropping-particle" : "", "family" : "Gales", "given" : "Mark", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Young", "given" : "Steve", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "container-title" : "Foundations and Trends\u00ae in Signal Processing", "id" : "ITEM-1", "issue" : "3", "issued" : { "date-parts" : [ [ "2007" ] ] }, "page" : "195-304", "title" : "The Application of Hidden Markov Models in Speech Recognition", "type" : "article-journal", "volume" : "1" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Gales & Young, 2007)", "plainTextFormattedCitation" : "(Gales & Young, 2007)", "previouslyFormattedCitation" : "(Gales & Young, 2007)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Gales & Young, 2007). 2.2 Performance of GMMs-HMMs Compared with Deep Neural Networks (DNNs)right2489200CorpusTraining SpeechSGMM WERDNN WERBABEL Pashto10 hours69.20%67.60%BABEL Pashto80 hours50.20%42.30%Fisher English2000 hours15.40%10.30%Table SEQ Table \* ARABIC 1. Summary of word error rates for a subspace Gaussian Mixture Model and a Deep Neural Network.00CorpusTraining SpeechSGMM WERDNN WERBABEL Pashto10 hours69.20%67.60%BABEL Pashto80 hours50.20%42.30%Fisher English2000 hours15.40%10.30%Table SEQ Table \* ARABIC 1. Summary of word error rates for a subspace Gaussian Mixture Model and a Deep Neural Network.Over the last decade, advances in computer hardware, machine learning, and deep learning algorithms have facilitated the faster and more accurate training of Deep Neural Networks (DNN) ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "DOI" : "10.1109/MSP.2012.2205597", "ISSN" : "1053-5888", "author" : [ { "dropping-particle" : "", "family" : "Hinton", "given" : "Geoffrey", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Deng", "given" : "Li", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Yu", "given" : "Dong", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Dahl", "given" : "George", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Mohamed", "given" : "Abdel-rahman", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Jaitly", "given" : "Navdeep", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Senior", "given" : "Andrew", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Vanhoucke", "given" : "Vincent", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Nguyen", "given" : "Patrick", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Sainath", "given" : "Tara", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Kingsbury", "given" : "Brian", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "container-title" : "IEEE Signal Processing Magazine", "id" : "ITEM-1", "issue" : "6", "issued" : { "date-parts" : [ [ "2012", "11" ] ] }, "page" : "82-97", "title" : "Deep Neural Networks for Acoustic Modeling in Speech Recognition: The Shared Views of Four Research Groups", "type" : "article-journal", "volume" : "29" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Hinton et al., 2012)", "plainTextFormattedCitation" : "(Hinton et al., 2012)", "previouslyFormattedCitation" : "(Hinton et al., 2012)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Hinton et al., 2012). This technology has made a series of breakthroughs in the area of Automatic Speech Recognition (ASR) in the past few years, outperforming systems based on approaches such as HMM and GMM-HMM. The performance gap, however, gets smaller as the amount of training data decreases. This observation is evident from the results that have been obtained with the Kaldi Speech Recognition Toolkit ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "author" : [ { "dropping-particle" : "", "family" : "Povey", "given" : "Daniel", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Ghoshal", "given" : "Arnab", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Boulianne", "given" : "Gilles", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Burget", "given" : "Lukas", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Glembek", "given" : "Ondrej", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Goel", "given" : "Nagendra", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Hannemann", "given" : "Mirko", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Motlicek", "given" : "Petr", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Qian", "given" : "Yanmin", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Schwarz", "given" : "Petr", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Silovsky", "given" : "Jan", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Stemmer", "given" : "Georg", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Vesely", "given" : "Karel", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "container-title" : "IEEE 2011 Workshop on Automatic Speech Recognition and Understanding", "id" : "ITEM-1", "issued" : { "date-parts" : [ [ "2011", "12" ] ] }, "note" : "IEEE Catalog No.: CFP11SRW-USB", "publisher" : "IEEE Signal Processing Society", "title" : "The Kaldi Speech Recognition Toolkit", "type" : "paper-conference" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Povey et al., 2011)", "plainTextFormattedCitation" : "(Povey et al., 2011)", "previouslyFormattedCitation" : "(Povey et al., 2011)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Povey et al., 2011) on the Intelligence Advanced Research Projects Activity (IARPA) provided database, BABEL, and the Fisher English Corpus. These results are summarized in REF _Ref468033205 \* MERGEFORMAT Table 1.The results that are shown in REF _Ref468034083 \* MERGEFORMAT Table 1 show that, indeed, DNNs are capable of achieving significant improvements in the performance of a speech recognition system. However, the difference is not as significant as when the number of training observations is not large enough.CHAPTER 3DATA & EXPERIMENTSIn this section, the data that was used for the experiments is characterized and the subset selection is described. Additionally, the setup for the baseline experiments, based on k-nearest neighbors and Random Forest, is presented. Finally, the experiments related to the establishment of a GMM-HMM based system for the abnormal identification of EEGs are described in detail.3.1 DataThis study utilized a subset of the Temple University Hospital EEG (TUH EEG) data corpus, which represents the largest publicly available database of clinical EEGs ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "DOI" : "", "author" : [ { "dropping-particle" : "", "family" : "Harati", "given" : "Amir", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Lopez", "given" : "Silvia", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Obeid", "given" : "Iyad", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Jacobson", "given" : "Mercedes", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Tobochnik", "given" : "Steven", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Picone", "given" : "Joseph", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "container-title" : "Proceedings of the IEEE Signal Processing in Medicine and Biology Symposium", "id" : "ITEM-1", "issued" : { "date-parts" : [ [ "2014" ] ] }, "page" : "1-5", "publisher-place" : "Philadelphia, Pennsylvania, USA", "title" : "THE TUH EEG CORPUS: A Big Data Resource for Automated EEG Interpretation", "type" : "paper-conference" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Amir Harati et al., 2014b)", "plainTextFormattedCitation" : "(Amir Harati et al., 2014b)", "previouslyFormattedCitation" : "(Amir Harati et al., 2014b)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Amir Harati et al., 2014b). The database is currently comprised of more than 30,000 records from over 18,000 unique patients. Given the nature of this study, it is additionally important to point out the fact that around 75% of the records present in TUH EEG are abnormal. 3.1.1 Data Subset Selectionright1263650Figure SEQ Figure \* ARABIC 7. Distribution of the patients’ ages and genders for the selected dataset00Figure SEQ Figure \* ARABIC 7. Distribution of the patients’ ages and genders for the selected datasetFor the purposes of the study, a demographically balanced subset of the TUH EEG database was selected. The age and gender of the patients were considered for the selection of the data, and because pediatric EEGs are essentially very different in nature than adult EEGs ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "ISBN" : "978-1451131956", "author" : [ { "dropping-particle" : "", "family" : "Ebersole", "given" : "J S", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Pedley", "given" : "T A", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "edition" : "4th", "id" : "ITEM-1", "issued" : { "date-parts" : [ [ "2014" ] ] }, "number-of-pages" : "664", "publisher" : "Wolters Kluwer", "publisher-place" : "Philadelphia, Pennsylvania, USA", "title" : "Current practice of clinical electroencephalography", "type" : "book" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Ebersole & Pedley, 2014)", "plainTextFormattedCitation" : "(Ebersole & Pedley, 2014)", "previouslyFormattedCitation" : "(Ebersole & Pedley, 2014)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Ebersole & Pedley, 2014), the majority of the records utilized were obtained from patients that were older than 20 years old. REF _Ref468228834 \* MERGEFORMAT Figure 7 shows the histograms of ages for the training and evaluation sets respectively. It is possible to see that, excluding two outliers in the datasets, all of the patients in the age range of 20-90, with a mean of 45.72 and a standard deviation of 14.91. The genders of the patients, as it can be also seen in REF _Ref468228834 \* MERGEFORMAT Figure 7, were also kept balanced. The final data subset that was divided in two sets: a training, which contained 80 abnormal and 82 normal EEGs, and an evaluation set, which contained 55 abnormal and 51 normal EEGs. From these recordings, only one channel was taken into consideration for the final analysis. The utilized channel was selected through experimental results, in a process that will be shown in a later section. 3.1.2 Feature ExtractionFeature extraction was performed on the EEG data in a pre-processing step. The feature extraction approach followed techniques that are similar to the ones based on Mel Frequency Cepstral Coefficients (MFCCs) that have been used for speech recognition ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "DOI" : "10.1109/53.54527", "ISSN" : "0740-7467", "author" : [ { "dropping-particle" : "", "family" : "Picone", "given" : "Joseph", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "container-title" : "IEEE ASSP Magazine", "id" : "ITEM-1", "issue" : "3", "issued" : { "date-parts" : [ [ "1990", "7" ] ] }, "page" : "26-41", "title" : "Continuous Speech Recognition Using Hidden Markov Models", "type" : "article-journal", "volume" : "7" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Picone, 1990)", "plainTextFormattedCitation" : "(Picone, 1990)", "previouslyFormattedCitation" : "(Picone, 1990)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Picone, 1990). MFCCs are normally calculated through the computation of a high resolution Fast Fourier Transform and down-sampling the results with an oversampling approach that uses overlapping bandpass filters. The results obtained from this process are then transformed to the cepstral domain through a cosine transform ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "ISBN" : "978-0130226167", "author" : [ { "dropping-particle" : "", "family" : "Huang", "given" : "Xuedong", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Acero", "given" : "Alex", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Hon", "given" : "Hsiao-Wuen", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "id" : "ITEM-1", "issued" : { "date-parts" : [ [ "2001" ] ] }, "number-of-pages" : "1008", "publisher" : "Prentice Hall", "publisher-place" : "Upper Saddle River, New Jersey, USA", "title" : "Spoken Language Processing: A Guide to Theory, Algorithm and System Development", "type" : "book" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Huang, Acero, & Hon, 2001)", "plainTextFormattedCitation" : "(Huang, Acero, & Hon, 2001)", "previouslyFormattedCitation" : "(Huang, Acero, & Hon, 2001)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Huang, Acero, & Hon, 2001). When extracting the cepstral coefficients from EEG signals, a very similar approach is followed, with the exception that the filterbanks are linearly spaced, rather than Mel-spaced, as they usually are for speech recognition approaches. The first eight cepstral coefficients were kept, and the rest were discarded. Following the cepstral coefficients extraction, the frequency energy of the signal was calculated and used to replace the 0th cepstral coefficient. The calculation of the frequency energy is given by:Ef=log?(kN-1Xk2 )( SEQ equation \* MERGEFORMAT 9)It is important to note that the frame and window duration for this portion of the feature extraction is 0.1 seconds and 0.2 seconds respectively ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "DOI" : "10.1109/SPMB.2015.7405421", "ISBN" : "978-1-5090-1350-0", "author" : [ { "dropping-particle" : "", "family" : "Harati", "given" : "A.", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Golmohammadi", "given" : "M.", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Lopez", "given" : "S.", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Obeid", "given" : "I.", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Picone", "given" : "J.", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "container-title" : "2015 IEEE Signal Processing in Medicine and Biology Symposium (SPMB)", "id" : "ITEM-1", "issued" : { "date-parts" : [ [ "2015", "12" ] ] }, "page" : "1-4", "publisher" : "IEEE", "publisher-place" : "Philadelphia", "title" : "Improved EEG event classification using differential energy", "type" : "paper-conference" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(A. Harati, Golmohammadi, Lopez, Obeid, & Picone, 2015)", "plainTextFormattedCitation" : "(A. Harati, Golmohammadi, Lopez, Obeid, & Picone, 2015)", "previouslyFormattedCitation" : "(A. Harati, Golmohammadi, Lopez, Obeid, & Picone, 2015)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(A. Harati, Golmohammadi, Lopez, Obeid, & Picone, 2015). right58Figure SEQ Figure \* ARABIC 8. Illustration of the base feature extraction process.00Figure SEQ Figure \* ARABIC 8. Illustration of the base feature extraction process.The extraction of the frequency energy and the cepstral coefficients is followed by the calculation of another type of energy: the differential energy (Ed). The differential energy is a feature derived from the features that have been described to this point, and it is given by the difference between the largest and the smallest sample in a 0.9 seconds window. This feature is described as:Ed=maxmEfm-minmEfm ?( SEQ equation \* MERGEFORMAT 10)Here, M represents the number of frames. REF _Ref468311978 \* MERGEFORMAT Figure 8 shows an illustration of the feature extraction process that has been explained up to this point.The first and second derivatives (differential and acceleration coefficients) of the base features explained to this point, are computed. These features represent the trajectory of the base features ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "ISBN" : "978-0130226167", "author" : [ { "dropping-particle" : "", "family" : "Huang", "given" : "Xuedong", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Acero", "given" : "Alex", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Hon", "given" : "Hsiao-Wuen", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "id" : "ITEM-1", "issued" : { "date-parts" : [ [ "2001" ] ] }, "number-of-pages" : "1008", "publisher" : "Prentice Hall", "publisher-place" : "Upper Saddle River, New Jersey, USA", "title" : "Spoken Language Processing: A Guide to Theory, Algorithm and System Development", "type" : "book" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Huang et al., 2001)", "plainTextFormattedCitation" : "(Huang et al., 2001)", "previouslyFormattedCitation" : "(Huang et al., 2001)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Huang et al., 2001), and are calculated as follows:dt=n=1Nnct+n-ct-n2n=1Nn2( SEQ equation \* MERGEFORMAT 11)In other words, dt represents a delta coefficient calculated for frame t in terms of the static coefficients ct+n to ct-n. Similar to the calculation of the Ed feature, the window N used for the first and second derivatives is set to 0.9 seconds in this study. ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "DOI" : "10.1109/SPMB.2015.7405421", "ISBN" : "978-1-5090-1350-0", "author" : [ { "dropping-particle" : "", "family" : "Harati", "given" : "A.", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Golmohammadi", "given" : "M.", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Lopez", "given" : "S.", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Obeid", "given" : "I.", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Picone", "given" : "J.", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "container-title" : "2015 IEEE Signal Processing in Medicine and Biology Symposium (SPMB)", "id" : "ITEM-1", "issued" : { "date-parts" : [ [ "2015", "12" ] ] }, "page" : "1-4", "publisher" : "IEEE", "publisher-place" : "Philadelphia", "title" : "Improved EEG event classification using differential energy", "type" : "paper-conference" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(A. Harati et al., 2015)", "plainTextFormattedCitation" : "(A. Harati et al., 2015)", "previouslyFormattedCitation" : "(A. Harati et al., 2015)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(A. Harati et al., 2015)In summary, the features that are extracted from the EEG signals are the frequency energy (1 feature), 7 cepstral coefficients (7 features), a differential energy term (1 feature), and the first and second derivatives (18 features) of the base features. In this sense, each feature vector for each frame of data in every channel is represented by a 27 dimensional feature vector. 3.1.3 Dimensionality ReductionAs it was mentioned before, only one channel was utilized for the experiments. Only the beginning of the recording was used for the abnormal identifications if the EEGs. This experimental paradigm was majorly based on the fact that neurologists are reportedly able to distinguish a normal recording from a normal one by looking at the first few seconds of the files. To establish a baseline for the project, the first 60 seconds of the signal were considered, and the features across all the frames corresponding to this time were stacked together, forming a 16,200 (600×27) ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "DOI" : "", "author" : [ { "dropping-particle" : "", "family" : "Lopez", "given" : "Silvia", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Suarez", "given" : "Gabriella", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Jungries", "given" : "Dave", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Obeid", "given" : "Iyad", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Picone", "given" : "Joseph", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "container-title" : "IEEE Signal Processing in Medicine and Biology Symposium", "id" : "ITEM-1", "issued" : { "date-parts" : [ [ "2015" ] ] }, "page" : "1-4", "publisher-place" : "Philadelphia, Pennsylvania, USA", "title" : "Automated Identification of Abnormal EEGs", "type" : "paper-conference" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Lopez et al., 2015)", "plainTextFormattedCitation" : "(Lopez et al., 2015)", "previouslyFormattedCitation" : "(Lopez et al., 2015)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Lopez et al., 2015). The dimensionality of the mentioned feature vectors was reduced, and only the most relevant eigenvectors of the covariance matrix for each class were retained ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "ISBN" : "978-0-387-95442-4", "author" : [ { "dropping-particle" : "", "family" : "Jolliffe", "given" : "I.T.", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "edition" : "2nd", "id" : "ITEM-1", "issued" : { "date-parts" : [ [ "2002" ] ] }, "number-of-pages" : "100", "publisher" : "Springer-Verlag", "publisher-place" : "New York City, New York, USA", "title" : "Principal Component Analysis", "type" : "book" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Jolliffe, 2002)", "plainTextFormattedCitation" : "(Jolliffe, 2002)", "previouslyFormattedCitation" : "(Jolliffe, 2002)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Jolliffe, 2002).3.2 ExperimentsThe experiments are separated in two different sections: the establishment of a baseline through standard machine learning techniques, and the implementation and optimization of different GMM-HMM systems in order to solve the problem. This subsection goes over the details for the setup and execution of said experiments.3.2.1 Baseline ExperimentsIn order to establish a baseline for the problem, two standard machine learning algorithms were implemented and studied: k-Nearest Neighbors (kNN) and Random Forest Ensemble Learning (RF). In essence, the k-nearest neighbor algorithm assigns a class to an observation based on the distance measurements, Mahalanobis distance, in this case, between the observation and its k-nearest neighbors ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "DOI" : "10.1007/BF01237942", "ISBN" : "0471056693", "ISSN" : "1433-7541", "PMID" : "2630878", "abstract" : "Introduction to Mathematical Techniques in Pattern Recognition by Harry C. Andrews This volume is one of the first cohesive treatments of the use of mathematics for studying interactions between various recognition environments. It brings together techniques previously scattered throughout the literature and provides a concise common notation that will facilitate the understanding and comparison of the many aspects of mathematical pattern recognition. The contents of this volume are divided into five interrelated subject areas: Feature Selection, Distribution Free Classification, Statistical Classification, Nonsupervised Learning, and Sequential Learning. Appendices describing specific aspects of feature selection and extensive reference and bibliographies are included. 1972 253 pp. Threshold Logic and its Applications by Saburo Muroga This is the first in-depth exposition of threshold logic and its applications using linear programming and integer programming as optimization tools. It presents threshold logic as a unified theory of conventional simple gates, threshold gates and their networks. This unified viewpoint explicitly reveals many important properties that were formerly concealed in the framework of conventional switching theory (based essentially on and, or and not gates). 1971 478 pp. Knowing and Guessing A Quantitative Study of Inference and Information By Satosi Watanabe This volume presents a coherent theoretical view of a field now split into different disciplines: philosophy, information science, cybernetics, psychology, electrical engineering, and physics. The target of investigation is the cognitive process of knowing and guessing. In contrast to traditional philosophy, the approach is quantitative rather than qualitative. The study is formal in the sense that the author is not interested in the contents of knowledge or the physiological mechanism of the process of knowing.", "author" : [ { "dropping-particle" : "", "family" : "Duda", "given" : "Richard O.", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Hart", "given" : "Peter E.", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Stork", "given" : "David G.", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "container-title" : "New York: John Wiley, Section", "id" : "ITEM-1", "issued" : { "date-parts" : [ [ "2001" ] ] }, "number-of-pages" : "680", "title" : "Pattern Classification", "type" : "book" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Duda et al., 2001)", "plainTextFormattedCitation" : "(Duda et al., 2001)", "previouslyFormattedCitation" : "(Duda et al., 2001)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Duda et al., 2001). While the Random Forest Ensemble Learning, makes a classification decision by considering the decisions made by all of the decision trees in the ensemble of trees (forest), and picking the class that received the majority of the votes. The first set of experiments involved the variation of the dimensions of the features, in order to find the optimal dimension for the feature vectors. The following step was the individual optimization of the models by varying parameters specific to the algorithms. For kNN the number of nearest neighbors (k) was varied from 1 to 100 and for RF, the number of trees (Nt) was studied for Nt values that ranged from 1 to 100. Once the parameters were properly optimized, a study about the relevance of the different channels for the normal/abnormal problem was conducted. Basically, the optimized systems were tested for all the 22 channels in the transverse central parietal (TCP) montage ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "author" : [ { "dropping-particle" : "", "family" : "ACNS", "given" : "", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "id" : "ITEM-1", "issued" : { "date-parts" : [ [ "2006" ] ] }, "title" : "Guideline 6: A Proposal for Standard Montages to Be Used in Clinical EEG [White Paper]. Retrieved from", "type" : "report" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(ACNS, 2006)", "plainTextFormattedCitation" : "(ACNS, 2006)", "previouslyFormattedCitation" : "(ACNS, 2006)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(ACNS, 2006), which accentuates spike activity, and the best performance was selected.The results of these experiments helped to establish a baseline for the classification of normal and abnormal adult EEGs. The performance reported for the baselines was then compared to an HMM system, which was implemented with the hopes that the nature of the model and its reported usefulness on sequential data could help decrease the false alarm and detection rates.3.1.2 HMM Experimentsright3432637Figure SEQ Figure \* ARABIC 9. Location of studied channels in the 10-20 standard system of electrode placement for the TCP montage00Figure SEQ Figure \* ARABIC 9. Location of studied channels in the 10-20 standard system of electrode placement for the TCP montageSeveral experiments were conducted in order to optimize the HMM and find the proper number of Gaussian mixtures. The first step for these experiments was to find the optimal number of Gaussian mixtures and states for the HMM by running classification experiments with the full set of features and the first 10 minutes for each file. The features that were reduced through (PCA) were then used to test this model for a better comparison with the baseline. Once the system’s parameters were properly optimized and a pertinent comparison had been done with the baseline, the models were used to find the optimal amount of input time for the signal by varying the input time from 5 to 25 in steps of 5 minutes. Finally, the fully optimized model was implemented for different channels across the scalp. REF _Ref468395445 \h \* MERGEFORMAT Figure 9 shows the spatial information for the channels that were selected for the comparative study. It can be seen that approximately all the scalp regions (left side only, since symmetry is assumed) are represented by the channels selected for the study.The optimization for the HMM parameters was performed with the channel that showed the best performance for the baseline systems. The channel test with the HMM was conducted in order to verify whether the channel optimization could be generalized across the tested models.CHAPTER 4PRELIMINARY EXPERIMENTS AND RESULTSThis section presents the results of the experiments that were described in Chapter 3. First, the results for the baseline experiments, which involved the classification of the normal and abnormal EEGs with the kNN and the RF algorithms, are presented. The last part of this chapter, presents the preliminary results that were obtained through HMMs. 4.1 Baseline Resultsright1212215Figure SEQ Figure \* ARABIC 10. Normal/abnormal classification error rate as a function of number of (trees Nt)00Figure SEQ Figure \* ARABIC 10. Normal/abnormal classification error rate as a function of number of (trees Nt)The first step for the baseline experimental section was the selection of an optimal number of trees (Nt). The error rate for the normal/abnormal classification was computed as a function of Nt. REF _Ref468398360 \* MERGEFORMAT Figure 10 shows that the performance seems to saturate for Nt>50. This observation motivated the selection of Nt=50 in a compromise for the tradeoff between training time and performance ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "DOI" : "", "author" : [ { "dropping-particle" : "", "family" : "Lopez", "given" : "Silvia", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Suarez", "given" : "Gabriella", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Jungries", "given" : "Dave", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Obeid", "given" : "Iyad", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Picone", "given" : "Joseph", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "container-title" : "IEEE Signal Processing in Medicine and Biology Symposium", "id" : "ITEM-1", "issued" : { "date-parts" : [ [ "2015" ] ] }, "page" : "1-4", "publisher-place" : "Philadelphia, Pennsylvania, USA", "title" : "Automated Identification of Abnormal EEGs", "type" : "paper-conference" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Lopez et al., 2015)", "plainTextFormattedCitation" : "(Lopez et al., 2015)", "previouslyFormattedCitation" : "(Lopez et al., 2015)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Lopez et al., 2015). As it was discussed in chapter 3, the following experiments involved the study of the performance of both systems, kNN and RF, as a function of the PCA dimension of the features. REF _Ref468397581 \* MERGEFORMAT Figure 11 shows the results generated for each algorithm while varying the PCA dimension from 0 to 100. The previously left0Figure SEQ Figure \* ARABIC 11. Error rate of the kNN and RF algorithms as a function of the PCA dimension00Figure SEQ Figure \* ARABIC 11. Error rate of the kNN and RF algorithms as a function of the PCA dimensionselected number of trees Nt=50 was used for the RF implementation, while a value of k=1 was used for kNN ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "DOI" : "", "author" : [ { "dropping-particle" : "", "family" : "Lopez", "given" : "Silvia", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Suarez", "given" : "Gabriella", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Jungries", "given" : "Dave", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Obeid", "given" : "Iyad", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Picone", "given" : "Joseph", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "container-title" : "IEEE Signal Processing in Medicine and Biology Symposium", "id" : "ITEM-1", "issued" : { "date-parts" : [ [ "2015" ] ] }, "page" : "1-4", "publisher-place" : "Philadelphia, Pennsylvania, USA", "title" : "Automated Identification of Abnormal EEGs", "type" : "paper-conference" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Lopez et al., 2015)", "plainTextFormattedCitation" : "(Lopez et al., 2015)", "previouslyFormattedCitation" : "(Lopez et al., 2015)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Lopez et al., 2015). From REF _Ref468397581 \* MERGEFORMAT Figure 11 it can be inferred that, for kNN, the performance is not heavily impacted by PCA values larger than 20, while for RF, the trend of the error seems to decrease up to the point where the PCA dimension is 86. Taking this information into account, 86 was selected to be the input dimension. Considering the fact that the results obtained through kNN did not change much for PCA dimensions greater than 20, the optimization for the number of nearest neighbors (k) was studied separately for features that were reduced to a dimension of 20, and features that were reduced to a dimension of 86. REF _Ref468405805 \* MERGEFORMAT Figure 12 shows the results that were obtained through this experiment, and indicates that the performance has high variability, but considering its trend, the value of k that shows a reasonable performance in comparison with other values is k=20.Finally, the optimized models were used to conduct an analysis about the channel relevance for the classification. The 22 channels of the TCP montage were individually used for the classification. In this way, it was possible to understand which channels (and what regions of the scalp) contribute the most for right0Figure SEQ Figure \* ARABIC 12. Error rate as a function of the number of neighbors k for PCA dimension of 20 and 86. 00Figure SEQ Figure \* ARABIC 12. Error rate as a function of the number of neighbors k for PCA dimension of 20 and 86. the identification of abnormal EEGs. REF _Ref468448332 \* MERGEFORMAT Figure 13 shows the most relevant results from this study. This analysis was performed through the implementation of the of the optimized kNN system (k=20), since it right4221480Figure SEQ Figure \* ARABIC 13. Classification error rate (for kNN) for a fronto-central (F4-C4) and a temporal-occipital (T5-O1) channel00Figure SEQ Figure \* ARABIC 13. Classification error rate (for kNN) for a fronto-central (F4-C4) and a temporal-occipital (T5-O1) channelpresented less variance than the RF implementation ADDIN CSL_CITATION { "citationItems" : [ { "id" : "ITEM-1", "itemData" : { "DOI" : "", "author" : [ { "dropping-particle" : "", "family" : "Lopez", "given" : "Silvia", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Suarez", "given" : "Gabriella", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Jungries", "given" : "Dave", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Obeid", "given" : "Iyad", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" }, { "dropping-particle" : "", "family" : "Picone", "given" : "Joseph", "non-dropping-particle" : "", "parse-names" : false, "suffix" : "" } ], "container-title" : "IEEE Signal Processing in Medicine and Biology Symposium", "id" : "ITEM-1", "issued" : { "date-parts" : [ [ "2015" ] ] }, "page" : "1-4", "publisher-place" : "Philadelphia, Pennsylvania, USA", "title" : "Automated Identification of Abnormal EEGs", "type" : "paper-conference" }, "uris" : [ "" ] } ], "mendeley" : { "formattedCitation" : "(Lopez et al., 2015)", "plainTextFormattedCitation" : "(Lopez et al., 2015)", "previouslyFormattedCitation" : "(Lopez et al., 2015)" }, "properties" : { "noteIndex" : 0 }, "schema" : "" }(Lopez et al., 2015). The performance reported on REF _Ref468448332 \* MERGEFORMAT Figure 13 was shown for the channel that performed worse (F4-C4) and the one that showed the best performance (T5-O1). This observation is consistent with the way in which neurologists interpret the EEGs, which heavily involves the identification of abnormalities (slowing, lack of reactivity) in the posterior dominant rhythm, present in the posterior regions of the scalp (posterior, occipital-temporal channels in the TCP montage). right1225550No.System DescriptionError1kNN (k=20)41.80%2RF (Nt=50)31.70%Table SEQ Table \* ARABIC 2. Comparison of the performance obtained with the two baseline systems00No.System DescriptionError1kNN (k=20)41.80%2RF (Nt=50)31.70%Table SEQ Table \* ARABIC 2. Comparison of the performance obtained with the two baseline systemsright2178050Ref/Hyp?NormalAbnormalNormal50.50%49.50%Abnormal34.00%66.00%Table SEQ Table \* ARABIC 3. Confusion matrix for the kNN system00Ref/Hyp?NormalAbnormalNormal50.50%49.50%Abnormal34.00%66.00%Table SEQ Table \* ARABIC 3. Confusion matrix for the kNN systemTo summarize, the performance of the optimized version of both systems for input feature vectors of dimension 86 is presented in REF _Ref468449712 \* MERGEFORMAT Table 2. The RF algorithm showed considerably higher variance than the kNN system. In order to obtain a better comparison with the HMM system, a confusion matrix for kNN is also shown in REF _Ref468450180 \* MERGEFORMAT Table 3.The results for the optimized baselines were then compared with a GMM-HMM system. The optimization details and performance of this system are presented in next section. 4.1 GMM-HMM ResultsThe optimization of the GMM-HMM system for this classification problem involves the selection of parameters such as the number of Gaussian mixtures and the number of HMM states. In order to do this, the first 10 minutes of data (features) for the T5-O1 channel were used as an input to the system. REF _Ref468451812 \* MERGEFORMAT Table 4 shows the summary of the results that were obtained through the evaluation of a number of system parameters. The closed loop performance for the best system (#GM=3, #HMM States=3) reached a correct detection rate of 86.420%.right0# Gaussian Mixtures# HMM StatesCorrect Detection (%)1169.81%1265.09%1365.09%2176.42%2280.19%2377.36%3176.42%3282.08%3383.02%4182.08%4264.15%4377.36%Table SEQ Table \* ARABIC 4. GMM-HMM correct detection rate for various numbers of Gaussian 00# Gaussian Mixtures# HMM StatesCorrect Detection (%)1169.81%1265.09%1365.09%2176.42%2280.19%2377.36%3176.42%3282.08%3383.02%4182.08%4264.15%4377.36%Table SEQ Table \* ARABIC 4. GMM-HMM correct detection rate for various numbers of Gaussian The information shown in REF _Ref468451812 \* MERGEFORMAT Table 4 shows that the optimal classification is obtained when the number of Gaussian mixtures is 3 and the number of HMM states is also 3 (#GM=3, #HMM States=3). right1546860Input (min)#Gaussians/#HMM StatesCorrect Detection (%)53/380.19%103/383.02%153/380.19%203/379.25%253/376.42%Table SEQ Table \* ARABIC 5. Correct detection rate for different signal input lengths00Input (min)#Gaussians/#HMM StatesCorrect Detection (%)53/380.19%103/383.02%153/380.19%203/379.25%253/376.42%Table SEQ Table \* ARABIC 5. Correct detection rate for different signal input lengthsTo understand how much signal information would work better for the identification of abnormal EEGs, the optimized system was used to process different input lengths. REF _Ref468454951 \* MERGEFORMAT Table 5, shows this analysis, and reveals that the best performance can be obtained for an input time of 10 minutes. The length of the majority of the recordings in the dataset are less than 25 minutes, so the performance saturates for durations longer than 25 minutes.So far, the results that were presented were calculated with data from the T5-O1 channel, which was found to be optimal for the baseline systems. To make sure the channel selection could be generalized for the different systems an analysis was ran for several channels. REF _Ref468457277 \* MERGEFORMAT Table 6 shows the results of these experiments.right0Input (min)#Gaussians/#HMM StatesChannelCorrect Detection (%)53/3Fp1-F780.19%103/3T5-O183.02%153/3F7-T380.19%203/3C3-Cz79.25%253/3P3-O176.42%Table SEQ Table \* ARABIC 6. Correct detection rate for different channels00Input (min)#Gaussians/#HMM StatesChannelCorrect Detection (%)53/3Fp1-F780.19%103/3T5-O183.02%153/3F7-T380.19%203/3C3-Cz79.25%253/3P3-O176.42%Table SEQ Table \* ARABIC 6. Correct detection rate for different channelsIt is possible to observe that the channel that performed best for the GMM-HMM system is the same that was discovered through the baseline systems. In this sense, it can be said that this temporal-occipital channel has great relevance in the classification of abnormal EEGs. right3564890Ref/HypNormalAbnormalNormal78.18%21.82%Abnormal11.76%88.24%Table SEQ Table \* ARABIC 8. Confusion matrix for the best GMM-HMM system00Ref/HypNormalAbnormalNormal78.18%21.82%Abnormal11.76%88.24%Table SEQ Table \* ARABIC 8. Confusion matrix for the best GMM-HMM systemright2204720System DescriptionError (%)kNN (k=20)41.80%RF (Nt=50)31.70%PCA-HMM #GM = 3 #HMM States = 3)25.64%GMM-HMM (#GM = 3 #HMM States = 3)16.98%Table SEQ Table \* ARABIC 7. Summary of the performance for all the evaluated systems00System DescriptionError (%)kNN (k=20)41.80%RF (Nt=50)31.70%PCA-HMM #GM = 3 #HMM States = 3)25.64%GMM-HMM (#GM = 3 #HMM States = 3)16.98%Table SEQ Table \* ARABIC 7. Summary of the performance for all the evaluated systemsThe results that have been presented to this point, can be further summarized and compared to the baseline performance. REF _Ref468458564 \* MERGEFORMAT Table 7 shows the results of this comparison. The PCA-HMM experiment used the same exact inputs that were used for the baseline systems (for comparison) and the GMM-HMM classification system was use as the back-end. It can be seen that the best performance was achieved by both of the HMM systems, with the full feature system having the lowest overall error rate. REF _Ref468459570 \* MERGEFORMAT Table 8 shows the confusion matrix for the best reported system. It can be seen that the GMM-HMM system showed an improvement of 27.7% compared to the false alarm rate of the baseline kNN system. CHAPTER 5EXPECTATIONS AND FUTURE WORK5.1 Expected OutcomesThe classification experiments that have been presented in this study have shown that HMMs can be used in the abnormal identification of EEG signals, achieving performance that surpasses that of the other classification algorithms. Similarly, as it was discussed in Chapter 2, HMMs have been very successful in the area of automatic speech recognition and have recently been outperformed by deep learning models trained in sufficiently large amounts of data. The main goal for the remaining research on this topic involves the implementation of deep learning techniques in order to decrease the reported false alarm rate and increase the sensitivity of the system, as has been done in the speech domain. In addition, given previous research, this task would involve the expansion and validation of a larger normal/abnormal database. The implementation of a deep learning system would allow to integrate the spatial and temporal context in order to take advantage of certain aspects of the domain knowledge, such as the fact that when the eyes are closed the PDR emerges, and improve the classification error rates. Timeline for Future WorkThe timeline for the expected completion dates for the tasks that are necessary for the completion of this study are outlined as follows:December-January:Set up deep learning system for a second pass of deep learning after the GMM-HMM processing:Implement and optimize a Stacked Denoising Autoencoders (SdA) system for the classification and increase the number of channels that are taken into account for the classification decision.Expand and evaluate the normal/abnormal TUH database subset:Generate simple natural language processing (NLP) scripts to obtain EEG sessions that have been evaluated and classified by neurologists and form a larger, demographically balanced, subset of the data.FebruaryImplement a long short term memory system for the normal/abnormal classification of EEGs in order to compare to the HMM-SdA implementation.This system will be implemented with the Theano Python library for deep learning and evaluated in the expanded dataset.Evaluate the SdA implementation on the expanded dataset.Get results for the SdA implementation with the expanded dataset.March-MayComplete the writing of the thesis and work on publications.Defend this thesis.REFERENCESADDIN Mendeley Bibliography CSL_BIBLIOGRAPHY ACNS. (2006). Guideline 6: A Proposal for Standard Montages to Be Used in Clinical EEG [White Paper]. Retrieved from. Retrieved from , H., Hori, S., Nakanishi, M., Fujimoto, S., Ichikawa, N., & Furukawa, T. A. (2003). An intervention to improve the interrater reliability of clinical EEG interpretations. Psychiatry and Clinical Neurosciences, 57(5), 485–489. , L. E., Petrie, T., Soules, G., & Weiss, N. (1970). A Maximization Technique Occurring in the Statistical Analysis of Probabilistic Functions of Markov Chains. The Annals of Mathematical Statistics, 41(1), 164–171. , C. (2011). Pattern Recognition and Machine Learning (2nd ed.). New York, New York, USA: Springer. Retrieved from , C., Tutwiler, R. L., & Slobounov, S. (2008). Automatic classification of athletes with residual functional deficits following concussion by means of EEG signal using support vector machine. IEEE Transactions on Neural Systems & Rehabilitation Engineering, 16(4), 327–335. , R. O., Hart, P. E., & Stork, D. G. (2001). Pattern Classification. New York: John Wiley, Section. , J. S., & Pedley, T. A. (2014). Current practice of clinical electroencephalography (4th ed.). Philadelphia, Pennsylvania, USA: Wolters Kluwer. Retrieved from , S., & van Putten, M. J. a M. (2013). EEG in ischaemic stroke: quantitative EEG can uniquely inform (sub-)acute prognoses and clinical management. Clinical Neurophysiology?: Official Journal of the International Federation of Clinical Neurophysiology, 124(1), 10–19. , M., & Young, S. (2007). The Application of Hidden Markov Models in Speech Recognition. Foundations and Trends? in Signal Processing, 1(3), 195–304. , A., Golmohammadi, M., Lopez, S., Obeid, I., & Picone, J. (2015). Improved EEG event classification using differential energy. In 2015 IEEE Signal Processing in Medicine and Biology Symposium (SPMB) (pp. 1–4). Philadelphia: IEEE. , A., Lopez, S., Obeid, I., Jacobson, M., Tobochnik, S., & Picone, J. (2014a). The TUH EEG Corpus: A Big Data Resource for Automated EEG Interpretation. In Proceedings of the IEEE Signal Processing in Medicine and Biology Symposium (pp. 1–5). Philadelphia, Pennsylvania, USA. , A., Lopez, S., Obeid, I., Jacobson, M., Tobochnik, S., & Picone, J. (2014b). THE TUH EEG CORPUS: A Big Data Resource for Automated EEG Interpretation. In Proceedings of the IEEE Signal Processing in Medicine and Biology Symposium (pp. 1–5). Philadelphia, Pennsylvania, USA. , G., Deng, L., Yu, D., Dahl, G., Mohamed, A., Jaitly, N., … Kingsbury, B. (2012). Deep Neural Networks for Acoustic Modeling in Speech Recognition: The Shared Views of Four Research Groups. IEEE Signal Processing Magazine, 29(6), 82–97. , X., Acero, A., & Hon, H.-W. (2001). Spoken Language Processing: A Guide to Theory, Algorithm and System Development. Upper Saddle River, New Jersey, USA: Prentice Hall. Retrieved from , I. T. (2002). Principal Component Analysis (2nd ed.). New York City, New York, USA: Springer-Verlag. Retrieved from , S. S., & van Putten, M. J. A. M. (2013). Quantification of the adult EEG background pattern. Clinical Neurophysiology, 124(2), 228–237. , S., Suarez, G., Jungries, D., Obeid, I., & Picone, J. (2015). Automated Identification of Abnormal EEGs. In IEEE Signal Processing in Medicine and Biology Symposium (pp. 1–4). Philadelphia, Pennsylvania, USA. , J. (1990). Continuous Speech Recognition Using Hidden Markov Models. IEEE ASSP Magazine, 7(3), 26–41. , D., Ghoshal, A., Boulianne, G., Burget, L., Glembek, O., Goel, N., … Vesely, K. (2011). The Kaldi Speech Recognition Toolkit. In IEEE 2011 Workshop on Automatic Speech Recognition and Understanding. IEEE Signal Processing Society.Rabiner, L. (1989). A Tutorial on Hidden Markov Models and Selected Applications in Speech Recognition. Proceedings of the IEEE, 77(2), 257–286. , S. (2005). EEG in the diagnosis, classification, and management of patients with epilepsy. Journal of Neurology, Neurosurgery, and Psychiatry, 76(Suppl 2), ii2-ii7. Young. (1996). A review of large-vocabulary continuous-speech. IEEE Signal Processing Magazine, 13(5), 45. ................
................

In order to avoid copyright disputes, this page is only a partial summary.

Google Online Preview   Download