A Comparison of Energy-Efficient Seizure Detectors for Implantable Neurostimulation Devices

Abstract

Introduction: About 30% of epilepsy patients are resistant to treatment with antiepileptic drugs, and only a minority of these are surgical candidates. A recent therapeutic approach is the application of electrical stimulation in the early phases of a seizure to interrupt its spread across the brain. To accomplish this, energy-efficient seizure detectors are required that are able to detect a seizure in its early stages.

Methods: Three patient-specific, energy-efficient seizure detectors are proposed in this study: (i) random forest (RF); (ii) long short-term memory (LSTM) recurrent neural network (RNN); and (iii) convolutional neural network (CNN). Performance evaluation was based on EEG data (n = 40 patients) derived from a selected set of surface EEG electrodes, which mimic the electrode layout of an implantable neurostimulation system. As for the RF input, 16 features in the time- and frequency-domains were selected. Raw EEG data were used for both CNN and RNN. Energy consumption was estimated by a platform-independent model based on the number of arithmetic operations (AOs) and memory accesses (MAs). To validate the estimated energy consumption, the RNN classifier was implemented on an ultra-low-power microcontroller.

Results: The RNN seizure detector achieved a slightly better level of performance, with a median area under the precision-recall curve score of 0.49, compared to 0.47 for CNN and 0.46 for RF. In terms of energy consumption, RF was the most efficient algorithm, with a total of 67k AOs and 67k MAs per classification. This was followed by CNN (488k AOs and 963k MAs) and RNN (772k AOs and 978k MAs), whereby MAs contributed more to total energy consumption. Measurements derived from the hardware implementation of the RNN algorithm demonstrated a significant correlation between estimations and actual measurements.

Discussion: All three proposed seizure detection algorithms were shown to be suitable for application in implantable devices. The applied methodology for a platform-independent energy estimation was proven to be accurate by way of hardware implementation of the RNN algorithm. These findings show that seizure detection can be achieved using just a few channels with limited spatial distribution. The methodology proposed in this study can therefore be applied when designing new models for responsive neurostimulation.

Keywords: convolutional neural network; low-power hardware implementation; random forest; recurrent neural network; responsive neurostimulation; seizure detection.

Figure 1

A minimally invasive electrode setup as a part of an implantable system for focal epilepsy (Copyright © Precisis AG, Heidelberg, Germany).

Figure 2

Schematic outline of the study design.

Figure 3

Comparison of the three classifiers across 40 patients using the AUC-ROC score as the performance metric.

Figure 4

Comparison of the three classifiers across 40 patients using AUC-PR score as the performance metric.

Figure 5

Comparison of the seizure detectors across 40 patients using sensitivity, FDR (per hour), and average detection delay (s) as the performance metrics.

Figure 6

Estimated number of arithmetic operations, memory accesses, and energy using the proposed method: (A) RF, (B) CNN, and (C) RNN.

Figure 7

Comparison of the classification energy consumption and the number of operations for the proposed seizure detection classifiers.

Figure 8

Measured classification energy over the calculated energy and its linear regression curve. The energies are determined by measuring the energy of an RNN implementation on an Apollo 3 Blue ARM Cortex-M4F microcontroller unit under the variation of the number of LSTM cells from 2 to 20.