LAUSR

Introduction In recent years, multi-modal approaches have emerged integrating functional near-infrared spectroscopy (fNIRS) with electroencephalography (EEG) to offer dual hemodynamic and electro-potential characterization of a seizure event. Herein, we employ deep learning methods such as the recurrent neural network (RNN) with long short- term memory (LSTM) cell as it is well suited for sequential data to propose a novel means of seizure detection. Methods We designed a deep RNN-LSTM and used as input multi-modal data from 40 patients between the ages of 22–70 years of age suffering from recalcitrant bilateral temporal lobe epilepsy. Prior consent was obtained before recordings were performed in the epilepsy monitoring unit of Notre Dame Hospital, Montreal, Canada. For each recording, distinct seizure and non-seizure classes were partitioned. Initially, EEG data only was used as input into the RNN-LSTM, followed by multi-modal data. Results Through extensive hyper-parameter optimization and data regularization techniques, we show that multi-modal EEG-fNIRS data provides superior performance metrics in a seizure detection task with low generalization error and loss with sensitivity and specificity of 90%, 96% respectively as compared to EEG alone, with sensitivity and specificity of 82% and 90% respectively. The hemodynamic response to seizure non-routinely appeared prior to electrical discharges; thereby indicating cerebral oxygenation parameters are possible seizure harbinger. Conclusion These results exemplify the enhanced detection value of multi-modal neuroimaging, particularly fNIRS, in epileptic patients. Furthermore, the neural network models proposed and characterized herein offer a promising framework for future multi-modal EEG-fNIRS investigations in precocious seizure detection.