Abstracts

Rationale: Machine learning methods for epileptic seizure detection and forecasting have been reported to show excellent performance. However, the broader implementation of these algorithms to populations of patients remains unclear and may compromise their clinical utility. The vast majority of the proposed algorithms are trained using a patient and acquisition system specific approach that does not take advantage of, and is not optimized for, a large-scale deployment in a population. In order to bridge this gap, we developed a seizure detection algorithm that is trained on previously acquired data from multiple subjects and immediately deployable on a new patient's data. The seizure detector works in real-time, independent of acquisition system, and can be re-iteratively re-trained and optimized if necessary.  Methods: We designed and implemented a real-time seizure detection system based on a recurrent Long-Short-Term-Memory (LSTM) neural network utilizing convolutional layers extracting time-frequency features of spectrograms. The model exhibits acquisition system invariance (sampling frequency, dynamic range), which allows for a large-scale, real-time deployment in subjects implanted with different devices. We used long-term recordings from canines implanted with the Neurovista Inc. system (400Hz sampling rate, 16 contacts) to train the model. The seizure detection algorithm was then deployed and prospectively tested in real-time in freely behaving canines with naturally occurring epilepsy using a different device (Medtronic Summit RC+S, 250Hz sampling rate, 16 contacts) and electrode configurations. The method was setup to operate for an arbitrary, configurable, number of electrodes and sampling frequency.  Results: The model was trained on NeuroVista device data from 5 canines. We then tested the algorithm prospectively over a six-month period. We captured 24 seizures in 2 canines implanted with the RC+S system recording from electrodes in anterior nuclei of the thalamus and hippocampus. We measured an average sensitivity of 100%, positive predictive value of 84%, and false positive rate of 0.03/day (one false positive detection in less than a month).  Conclusions: We implemented and tested a generalized automated method for seizure detection that does not require initial training, but can be further optimized with subsequent recordings as needed. The approach is independent of acquisition system and recording configuration (device, electrodes, montages, filters etc.). We verified and tested the method prospectively in pet dogs with epilepsy using data from a new generation implantable device with real-time data streaming capability. The approach should be useful for emerging implantable systems and enable creation of precise seizure diaries and to help guide adaptive stimulation.  Funding: This work was funded in part by NIH NINDS grants UH2-NS95495 & R01-NS92882-03 (GW) and in part by a Mayo Clinic Czech Republic collaboration funded by project no. LQ1605 from the National Program of Sustainability II (MEYS CR), Ministry of Youth and Sports of the Czech Republic project no. LH15047 (KONTAKT II), project LO1212 (MEYS CR), by the Czech Science Foundation under Grant 1720480S, in part by the Temporal Context in Analysis of LongTerm Non-Stationary Multidimensional Signal, and the European Regional Development Fund through the Project FNUSA-ICRC under Grant CZ.1.05/ 1.1.00/02.0123.