RATIONALE: Driving safety is consistently identified as a top concern for people with epilepsy. In young adults who outgrow clinical seizures, the question of whether they should be granted driving privileges remains a challenge for clinicians. We introduce an EEG-based machine learning approach to predict with a minimum false discovery rate whether or not generalized 3-4Hz spike-wave discharges (SWDs) produce impaired behavioral responsiveness.
METHODS: Time and frequency domain features were extracted from each EEG channel, in addition to features derived from the spatial distributions of discharges, as elucidated by the common spatial pattern (CSP) algorithm. Using features extracted from the pre-ictal and ictal periods, support vector machines and linear discriminant analysis were employed to classify seizures as spared or impaired in accordance with behavioral testing of responsiveness conducted in previous studies. In constructing our classifier, we prioritized the minimization of the false discovery rate, with the goal of avoiding at all costs classifying patients as potentially able to drive (spared) when in fact they are not. This was done by training our classifier on three labelled datasets from absence epilepsy patients where all SWD carried “spared” or “impaired” labels based on behavioral testing. We used the labelled data sets to fine tune the model parameters, prioritizing a minimal false discovery rate. Subsequently, the optimized classifiers were validated on a novel unlabeled dataset of patients with absence epilepsy on medication treatment which simply classified patients overall as having clinical seizures or not.
RESULTS: By utilizing the three labelled absence epilepsy datasets containing SWDs with corresponding behavioral testing, it was found that the optimal predictive period for behavioral impairment was the 1000ms leading up to seizure onset, in addition to the first 500ms of the discharge. This yielded a false discovery rate of zero with corresponding sensitivity of 92.86%. We found that classic features such as SWD power and duration, as well as other less intuitive features and the CSP algorithm were capable of greatly enhancing the classification. When validating the model on a fourth unlabeled dataset to assess patient “impairedness”, a false discovery rate of zero was achieved but with a reduced sensitivity of 34.78%, corresponding to 8 patients out of 23 being correctly classified as free of clinical seizures, but no patient being classified as potentially safe to drive if they were not.
CONCLUSIONS: The proposed approach demonstrates the feasibility and reliability of machine learning-based systems in clinical practice for the purposes of more finely evaluating patients with absence epilepsy. With further validation, this method could lead to significant improvements in quality of life by helping assess a patient’s ability to drive. In particular, we have shown that the synthesis of time, frequency, and spatial domain features is highly accurate in assessing subclinical spike-wave discharges and, as such, can be expanded on in future studies with larger cohorts for generalizability to more novel datasets.
FUNDING: This work was supported by the Betsy and Jonathan Blattmachr Family and the Loughridge Williams Foundation.