Closed-loop responsive neurostimulation requires predicting electrographic seizure onset seconds before behavioral manifestation. Cloud latency makes remote inference clinically useless. This paper evaluates specialized low-power edge tensor processing units (Edge TPUs) analyzing 64-channel intracranial EEG in real time.
By treating intracranial electrode arrays as dynamic graph nodes, graph neural networks capture cross-channel phase synchrony, predicting seizure events up to 45 seconds prior to clinical onset.
The custom edge TPU consumes only 180 milliwatts during peak inference, preventing tissue heating while delivering over 1.2 TOPS of dedicated tensor compute.