Abstracts

Rationale: Seizure prediction is a means to improve epilepsy patients’ control over their everyday lives and to deploy personalized preventive treatment. The usefulness of a seizure prediction system for an individual patient will depend on their individual preferences regarding sensitivity and accumulated alarm time [1]. Using long-term intracranial electroencephalography (iEEG) data [2], deep-learning technology, and a mobile neuromorphic processor [3], we have developed an early-stage proof-of-concept demonstration of a tunable, ultra-low power, fully-automated seizure-prediction system. Methods: We have demonstrated the potential of a mobile, patient-tunable, always-on, ultra-low power seizure prediction system by building an early-stage proof-of-concept implementation of such system (Figure 1). Performance was evaluated through a long-term pseudoprospective study using iEEG data from 10 patients [2]. IEEG data were converted into spectrograms and pre-ictal signals were classified as either inter-ictal or pre-ictal in real-time by a deep convolutional neural network (CNN) [4]. The output of the CNN was integrated using a exponentially weighted moving average filter. A threshold crossing triggered a pre-ictal alarm indicating an upcoming seizure. To account for the non-stationarity of the iEEG signal over time, the CNN was re-trained and redeployed at the end of each month. Finally, we demonstrated that the entire system can be run on a mobile processor. Results: Our system achieved a mean sensitivity of 68.6%, with a mean time in warning of 26.9% (averaged over the entire duration of the study, and all 10 patients). Seizure prediction performance was significantly above chance level for all patients (mean of 42.3% improvement over chance, p < 0.01). When tuning the alarm sensitivity versus time in warning for an example patient (see Figure 2), sensitivity could be increased from 83.1% to 91.8%, and time in warning could be decreased from 42.6% to 19.3%. Conclusions: This work provides a proof-of-concept for deploying the developed prediction models onto a mobile, ultra-low power system. The system architecture allows for instantaneous reconfiguration of the system. This tunability allows to adjust the relative importance of alarm sensitivity versus time in warning, catering to the preferences of different patients. Funding: This work was funded by the National Health & Medical Research Council (APP1065638) and by IBM.1. Elger, CE et al., “Seizure prediction and documentation—two important problems”, The Lancet Neurology, Volume 12, Issue 6, pp. 531 – 532, 2013.2. Cook, MJ et al., “Prediction of seizure likelihood with a long-term, implanted seizure advisory system in patients with drug-resistant epilepsy: a first-in-man study”, The Lancet Neurology, Volume 6, Issue 6, pp. 563-571, 2013.3. Merolla PA, Arthur JV, Alvarez-Icaza R, et al. “A million spiking-neuron integrated circuit with a scalable communication network and interface”, Science 345, 668–73, 2014.4. LeCun, Y. et al., “Deep Learning”, Nature 521, pp. 436-444, Feb. 2015.