Memory optimized random forest classifier for EEG seizure detection in implantable monitoring and closed-loop...

Objective. Up to one third of epilepsy patients do not achieve satisfactory seizure control and may benefit from implantable devices for responsive neurostimulation or online seizure monitoring. Beyond energy efficiency, the limited memory capacity in these devices, imposes significant constraints to algorithmic design of seizure detection models. This study aims to evaluate the performance of cross-patient random forest (RF) models optimized for low-power microcontroller applications by assessing various channel integration strategies and measuring their memory requirements.Approach. Fifty patients undergoing electroencephalographic monitoring with 362 seizures were included in the analysis, with approximately one hour of signal for each seizure. One central and four peripheral electrodes over the epileptogenic focus were selected to resemble the layout of a novel neurostimulation device. Fifteen features were extracted from 2 s non-overlapping segments. RF models comprised either 500 or 125 trees, with varying depths. Three early channel integration (EI) strategies were compared with late integration (LI), using three channel fusion methods. A leave-one-patient-out cross-validation approach was used for evaluation, and memory requirements, alongside with inference energy and latency for 8-bit integer and 32-bit floating point models were computed on a microcontroller.Main results.The performance of EI feature sorting and LI were comparable. LI was favored by the 32-bit floating point format and more complex models, with the median channel fusion achieving a median area under the receiver operating characteristic curve score of 0.925. Feature sorting performed best with medium-sized models and was largely unaffected by the 8-bit integer format. Following causal output post-processing, false stimulations per hour were reduced to 5.5 at 100% sensitivity and fell below 3 at∼80% sensitivity.Significance. Our findings suggest that RF models with minimal energy and memory requirements can achieve state-of-the-art performance, making them well-suited for embedded applications in implantable devices. The complex interplay of the investigated factors is critical to performance, and along hardware specifications, should guide algorithmic design.