Data Science in Medicine & Biology I
A1L-B: Data Science in Medicine & Biology ISession Type: Lecture
Session Code: A1L-B
Location: Room 2
Date & Time: Wednesday March 22, 2023 (09:00 - 10:00)
Chair: Adam Charles
|Feasibility of Regression Modeling and Biomarker Analysis for Epileptic Seizure Prediction
|Dominique Tanner, Michael Privitera, Marepalli Rao
|Epilepsy is a neurological disease that causes recurrent, spontaneous seizures, which can lead people to experience ephemeral neurological and physiological impairments that disrupt day-to-day living. To advance seizure prediction, this study focused on the feasibility of self-prediction by examining patient-specific morning and evening seizure diaries that consisted of possible seizure triggers, measurements of mood, and predictive symptoms. Prediction models were generated by employing logistic regression. Seizure triggers for the models were chosen using stepwise selection. Akaike Information Criterion was used to select ideal regression models that evaluated patients’ data. Biomarkers that were associated with seizure occurrences were analyzed and diagnostic tests were developed to determine seizure probability. Seizure triggers’ influence on patient’s seizure outcome was greater in the morning than in the evening, and performance accuracies for prediction models varied among patients. The correlation between the severity of seizure triggers and how biomarkers oscillated over time was identified. This research expanded efforts to further refine precision medicine and develop more dependable epilepsy-based healthcare treatments.
|Rebalancing Techniques for Asynchronously Distributed EEG Data to Improve Automatic Seizure Type Classification
|Niamh McCallan, Scot Davidson, Kok Yew Ng, Pardis Biglarbeigi, Dewar Finlay, Boon Leong Lan, James McLaughlin
|Epilepsy, a nervous system disorder, is characterised by unprovoked, unpredictable, and recurrent seizures. To diagnose epileptic seizures, electroencephalography (EEG) is frequently used in medical settings. Effective automated detection and classification strategies are needed because visual analysis and interpretation of EEG signals consume time and call for specialised expertise. The main objective of this paper is to examine the effectiveness of multiple rebalancing techniques to address the problem of asynchronously distributed data, specifically employing random resampling, synthetic minority oversampling technique (SMOTE), and adaptive synthetic sampling approach for imbalanced learning (ADASYN), for seizure type classification. The model utilises both frequency information using variational mode decomposition (VMD), and phase information by extracting the phase locking value (PLV) across 19 common EEG channels found in the Temple University Hospital EEG Seizure Corpus (TUSZ) v1.5.2 dataset. The random subspace k-nearest neighbour (RSkNN) ensemble classifier is used for seizure type classification of five classes --- complex partial seizures (CPSZ), simple partial seizures (SPSZ), absence seizures (ABSZ), tonic clonic seizures (TCSZ), and tonic seizures (TNSZ) --- to determine the performance of each rebalancing techniques, with the highest accuracy and weighted F1 score of 96.28% and 0.946, respectively using SMOTE with two nearest neighbours.
|Decoding EEG Signals with Visibility Graphs to Predict Varying Levels of Mental Workload
|Arya Teymourlouei, Rodolphe Gentili, James Reggia
|Recent work in predicting mental workload through EEG analysis has centered around features in the frequency domain. However, these features alone may not be enough to accurately predict mental workload. We propose a graph-based approach that filters EEG channels into five frequency bands. The time series data for each band is transformed into two types of visibility graphs. The natural visibility graph and horizontal visibility graph algorithms are used. Six graph-based features are then calculated which seek to distinguish between EEGs of low and high mental workload. Feature selection is evaluated with statistical tests. The features are fed as input data to two machine learning algorithms which are random forest and neural network. The accuracy of the random forest method is 90%, and the neural network has 86% accuracy. The graphical analysis showed that higher frequency ranges (alpha, beta, gamma) had a stronger ability to classify levels of mental workload. Unexpectedly, the natural visibility graph algorithm had better overall performance. Using the method presented here, accurate classification of MWL using EEG signals can enable the development of robust BCI.