An electronic neuromorphic system for real-time detection of High Frequency Oscillations (HFOs) in intracranial EEG

1 Introduction

• Growing sensory-data volumes call for low-power embedded systems that process signals locally, and SNNs are proposed for detecting complex spatiotemporal patterns.Neuromorphic circuits can integrate silicon neurons into compact, low-power systems for real-time processing.Source: §1 · p. 2
• Earlier neuromorphic systems required external bio-signal recording, frontend devices, and data-conversion interfaces.Clinical headstages prioritize few channels and high SNR, unlike experimental systems that emphasize high-density recordings and small circuit area.Source: §1 · p. 2
• The proposed system combines a neural recording headstage, signal-to-spike conversion, and a multi-core SNN on one die for epilepsy iEEG biomarker detection.The paper compares HFO detection with the Morphology Detector on medial temporal lobe iEEG.Source: §1 · p. 2
• HFOs are proposed iEEG biomarkers for delineating the epileptogenic zone, and online detection could provide information during surgery.The paper describes HFO co-occurrence as enabling optimal prediction of postsurgical seizure freedom in prior work.Source: §1 · p. 2

2 Results

• The processing chain filters prerecorded iEEG into Ripple and Fast Ripple bands, converts the signals to spikes, and feeds them to the SNN, whose output spikes signal HFO detection.The stages were first simulated in software to select parameters and then validated with hardware components.Source: §2 · p. 2

2.1 The neuromorphic system

• The chip integrates 8 recording headstages with four neuromorphic cores of 256 neurons each; its total area is 99 mm^2 and detection-task power is 614.3 μW.The four SNN cores occupy 77.2 mm^2.Source: §2.1 · p. 3

2.2 Analog headstage circuit measurements

• The LNA noise was below prerecorded iEEG power across the spectrum, and the filters were configured for 80–250 Hz Ripple and 250–500 Hz Fast Ripple bands.Each filter consumed 0.9 μW.Source: §2.2 · pp. 6–7
• Each ADM consumed 1.5 nJ per spike and had 96 nW static power dissipation; its spike rate varied from hundreds of hertz to hundreds of kilohertz with circuit settings.The refractory period was set to 300 μs.Source: §2.2 · p. 7

2.3 System performance

• The 256-neuron network uses fixed weights and varied synaptic time constants and neuron leak parameters; hardware device mismatch naturally produces response diversity.The ensemble detects an HFO when at least one second-layer neuron spikes within a 15 ms interval.Source: §2.3 · pp. 8–9
• Test-retest reliability ranged from 0.59 to 0.97 across patients, with a median of 0.91; Patient 1’s reliability score was 0.95.The analysis quantified consistency of HFO-rate spatial distributions across recording intervals.Source: §2.4 · p. 9, §2.3 · p. 9

2.4 Predicting seizure outcome

• 6 of 9 patients achieved seizure freedom; patients 7 and 9 were false negatives, while patient 8 was a true positive under the study’s classifications.Patients 1–6 had their HFO areas fully included in the resection area and achieved seizure freedom.Source: §2.4 · p. 10
• Overall postsurgical outcome-prediction accuracy was comparable to the Morphology Detector, and both detectors achieved 100% specificity.Table 2 compares outcome prediction between the detectors.Source: §2.4 · p. 10, Table 2 · p. 10

3 Discussion

• The integrated device demonstrates neuromorphic computing for compact, low-power edge applications that cannot rely on remote computers or cloud processing.The same chip integrates the signal-acquisition headstage and multi-core neuromorphic processor.Source: §3 · p. 10
• The headstage design adapts to different use cases by changing gain factors and band selections on the same input channel.The authors distinguish this capability from headstages optimized for particular application domains.Source: §3 · pp. 10–11
• The SNN approach uses mixed-signal circuits to emulate neuronal dynamics and exploits silicon-neuron variability to tune randomly distributed network parameters for HFO detection.Simulation also helped set headstage hyperparameters, while mismatch generated parameter distributions without manually specifying time constants or extra memory.Source: §3 · p. 11
• 100% specificity held at the individual-patient level, which the authors identify as a prerequisite for surgery aiming at seizure freedom.Overall accuracy was compatible with state-of-the-art software algorithms running on powerful computers.Source: §3 · p. 11

• Real-time, on-chip identification of relevant features in intracranial human data using event-based processors and spiking neural networks is presented as a feasibility study.Source: §4 · p. 11
• The proposed approach could lead to systems that detect HFO areas directly in the operating room.Source: §4 · p. 11
• The authors propose that such systems could improve seizure outcomes after epilepsy surgery.Source: §4 · p. 11

5.1 Design and setup of the hardware device

• The device was packaged on a ceramic 240-pin board assembly, programmed and debugged through an FPGA interface, and tested by feeding prerecorded iEEG to the chip.Circuit design and layout used Cadence tools, with Calibre used for design-rule checks, layout-versus-schematic checks, and post-layout extraction.Source: §5.1 · p. 11

5.2 Simulation and validation of the neural architecture

• The neural architecture was simulated in Brian2 with a custom toolbox, tuned using snippets around Morphology Detector-marked HFOs, and validated on the previous-generation DYNAP-SE processor.Source: §5.2 · p. 11

5.3 Patient data

• The study used medial temporal lobe iEEG from nine patients, focusing on the three most mesial bipolar channels and intervals recorded during interictal slow-wave sleep.Patients underwent resective surgery, were followed for more than a year, and postsurgical outcome was classified using ILAE categories.Source: §5.3 · pp. 11–12

5.4 HFO detection

• HFOs were detected per channel and five-minute interval after filtering, baseline detection, and ADM conversion, with thresholds adapted using the lowest quartile of first-second window maxima.Figure 6 contrasts the channel-wise optimal threshold with higher thresholds that produce too few spikes and lower thresholds that produce too many.Source: §5.4 · p. 12, Figure 6 · p. 13
• The SNN was tuned using spike-train ISIs for inputs with and without HFOs, then signaled detections from consecutive 15 ms output snippets.Output-layer neuron time constants averaged 15 ms; excitatory synapses used 3–6 ms and inhibitory synapses 0.1–1 ms.Source: §5.4 · p. 12

5.5 Post-surgical outcome prediction

• For retrospective outcome prediction, channel HFO rates formed interval-level vectors; the HFO area comprised electrodes above the 95th percentile of average rates and was compared with resection and postsurgical outcome.Source: §5.5 · pp. 12–13