Epileptic Seizure Prediction with Stacked Auto-encoders: Lessons from the Evaluation on a Large and Collaborative Database



The seizure prediction performance of algorithms based in stacked auto-encoders deep-learning technique has been evaluated. The study is established on long-term electroencephalography (EEG) recordings of 103 patients suffering from drug-resistant epilepsy. The proposed patient-specific methodology consists of feature extraction, classification by machine learning techniques, post-classification alarm generation, and performance evaluation using long-term recordings in a quasi-prospective way. Multiple quantitative features were extracted from EEG recordings. The classifiers were trained to discriminate preictal and non-preictal states. The first part of the feature time series was considered for training, a second part for selection of the ``optimal'' predictors of each patient, while the remaining data was used for prospective out-of-sample validation. The performance was assessed based on sensitivity and false prediction rate per hour (FPR/h). The prediction performance was statistically evaluated using an analytical random predictor. The validation data consisted of approximately 1664 h of interictal data and 151 seizures, for the invasive patients, and approximately 4446 h of interictal data and 406 seizures for the scalp patients. For the patients with intracranial electrodes 18% of the seizures were correctly predicted (27), leading to an average sensitivity of 16.05% and average FPR/h of 0.27/h. For the patients with scalp electrodes 20.69% of the seizures (84) on the validation set were correctly predicted, leading to an average sensitivity of 17.49% and an average FPR/h of 0.88/h. The observed performances were considered statistically significant for 4/19 invasive patients (? 21%) and for 5/84 scalp patients (? 6%). The observed results evidence the fact that, when applied in realistic conditions, the auto-encoder based classifier shows limited performance for a larger number of patients. However, the results obtained for some patients point that, in some specific situations seizure prediction is possible, providing a ``proof-of-principle'' of the feasibility of a prospective alarming system.


Precision Medicine Powered by pHealth and Connected Health 2018

Cited by

No citations found