Computationally efficient neural network classifiers for next generation closed loop neuromodulation therapy - a case study in epilepsy

Abstract

This work explores the potential utility of neural network classifiers for real- time classification of field-potential based biomarkers in next-generation responsive neuromodulation systems. Compared to classical filter-based classifiers, neural networks offer an ease of patient-specific parameter tuning, promising to reduce the burden of programming on clinicians. The paper explores a compact, feed - forward neural network architecture of only dozens of units for seizure-state classification in refractory epilepsy. The proposed classifier offers comparable accuracy to filter- classifiers on clinician-labeled data, while reducing detection latency. As a trade-off to classical methods, the paper focuses on keeping the complexity of the architecture minimal, to accommodate the on-board computational constraints of implantable pulse generator systems. Clinical relevance-A neural network-based classifier is presented for responsive neurostimulation, with comparable accuracy to classical methods at reduced latency.

Fig. 1. Classifier architectures.

Top: classical filter-based spectral power detector [6]. Bottom: the multi-layer perceptron architecture evaluated in this paper. Through training on labeled data, the neural network is expected to assume an overall transfer function similar to the hand-crafted filter topology.

Fig. 2. ROCs for different classifiers.

Note that performance converges towards high TPR and FPR, which is the desirable operating point of seizure detectors as FNs pose significantly greater risk of harm to patients than FPs.

Fig. 3. Tuning the MLP classifier.

Grid search on the two main model hyperparameters: the input vector length and the number of hidden layer units. Loss is represented as binary cross entropy across all samples.

Fig. 4. Detailed performance of classifiers.

Top: histogram of classification latency. Bottom: histogram showing the percentage of overlap between positive classifier output and clinician-labeled event.

Fig. 5. Frequency response of classifiers to input signals of different magnitudes.

Left: filter chain classifier. Right: MLP classifier with average of 3 windows. Accuracy is presented as the mean classifier output for a 1 second sinusoidal test tone over 10 repeats.