Detection Of Epileptic Seizure Using Stft And Statistical Analysis PDF Download

Over 65 million people worldwide suffer from epilepsy, and one third of these people have refractory epilepsy, which means typical epilepsy medications have not worked for them. Many patients who suffer from epilepsy are unaware of their seizures, yet doctors rely on reports by patients and families to assess the effectiveness of treatment. Studies have shown that patients do not remember more than half of their seizures, yet seizure journals are the most common method used to track seizure activity in outpatient settings. In epilepsy monitoring units, patients are monitored using combined video and electroencephalography (EEG) recordings. This is considered the gold standard of epilepsy monitoring, but is impractical for use outside of a hospital setting. Several studies have investigated using one or several other methods to track specific types of seizures, but not have been able to reliably detect all types of seizures. This thesis investigates the use of a suite of non-cerebral, i.e., non-EEG sensors that provide time series data with supervised machine learning to detect seizures. Data were recorded from fifteen patients (7 male, 5 female, 3 unrecorded, mean age 36.17), five of which had a total of seven seizures. The patients were monitored in the inpatient Epilepsy Monitoring Unit at the Dartmouth Hitchcock Medical Center, while wearing sensors including electrocardiography, electrodermal activity, electromyography, accelerometry, and audio data. Post-processing of the data shows that seizure data are separable from non-seizure data by using linear discriminant analysis on features derived from the signals recorded. The mean area under the ROC curve calculated using linear discriminant analysis for each patient was .9682. A detailed study was conducted to identify the features and sensors that contribute most significantly to separability of data acquired during seizures from other data. Subsequently, a real-time approach was evaluated using a discriminative classifier. A support vector machine was able to classify seizure and non-seizure data with high accuracy. The results presented show that this multimodal approach to seizure detection is promising. Moving forward, more data will be collected and other machine learning techniques will be investigated to increase accuracy in real-time seizure detection.