Distributed On-Node Seizure Detection and Network-Level Fusion for Closed-Loop Epilepsy Intervention (Mock R21)

Specific Aims

Drug-resistant temporal lobe epilepsy (TLE) affects over a third of the 50 million people worldwide suffering from epilepsy [1]. TLE is a network disorder characterized by repeated seizures, initiated in the hippocampus and propagating through the amygdala, thalamus, and temporal neocortex. However, the only FDA-approved, closed-loop neuromodulatory therapy that exists is the NeuroPace responsive neurostimulation (RNS) device, which consists of two electrode arrays with 4 contacts each. Although this system has been shown to provide a 75% seizure reduction at 9 years post-implantation, only 18% of patients achieve complete seizure freedom [2, 3]. We hypothesize that this is a result of a fundamental mismatch: TLE is a network-wide disorder, effective closed-loop intervention therefore requires simultaneous monitoring at multiple network nodes, which is a capability that focal architectures like NeuroPace RNS, limited to two electrode arrays, cannot provide. Recent advances in nano-scale, wireless, untethered neural interfaces have made distributed recording architectures feasible in vivo. For example, Lee et al developed "Neurograins", which demonstrated distributed neural recordings (48 nodes) and telemetry via radio frequency (RF) from the cortical surface to an external receiver [4]. Other groups have investigated the use of ultrasound-powered implants to extend wireless power delivery to depths exceeding 5 cm, which could be sufficient to reach the hippocampus [5]. However, these hardware platforms are fundamentally limited by strict power budgets (<1-10 μW per node after sensing and telemetry), and bandwidth constraints that shape how on-node detection algorithms must be designed. This project develops the computational decision layer between distributed recording and stimulation hardware through interface-algorithm co-design, where every algorithmic choice is derived from the power, thermal, and bandwidth constraints of implantable systems. Objective. Develop a computational decision layer for distributed, closed-loop seizure intervention, constrained by the realistic power, thermal, and bandwidth limitations of implantable hardware. This project involves no human subjects, animal subjects, or new data collection. Validation combines computational simulation, analysis of publicly available de-identified intracranial electroencephalogram (iEEG) datasets (iEEG.org, SWEC-ETHZ), and embedded hardware benchmarking using pre-recorded data.

Aim 1. Determine recording node computational requirements under realistic constraints.

Derive power, thermal, and telemetry budgets for on-node seizure detection, and compare existing hardware interfaces (RF vs ultrasound) against the depth requirements of hippocampus (5-7 cm), thalamic (6-7 cm), and neocortical targets. The power envelope derived here parameterizes Aim 2.

Aim 2. Develop and empirically validate on-node sequential detection algorithms in adherence to Aim 1 power budgets.

Design feature extraction and sequential probability ratio test (SPRT) algorithms for on-node seizure detection. Characterize the sensitivity, specificity, and latency per node as a function of power allocation to establish that individual detectors can serve as building blocks for network-level fusion. Validate detection performance on intracranial EEG against the RNS detection algorithm (line-length) and cumulative-sum (CUSUM) baselines. Implement the detection pipeline on a low-power embedded platform to empirically measure power consumption, latency, and bandwidth constraints. Single-node detection performance establishes the baseline that Aim 3's fusion must exceed, regardless of which sequential procedure proves optimal.

Aim 3. Design a distributed fusion-and-closed-loop intervention architecture.

Develop a multi-node fusion system to aggregate detection events from each node into a network-level seizure classification and neuromodulatory protocol, mapping the fusion output to stimulation commands compatible with the Lee et al. (2025) distributed microstimulator architecture [6]. RNS line-length detection achieves approximately 65-75% sensitivity at clinically acceptable FP rates; we target >90% to demonstrate network-level advantage.

Expected Impact

This project aims to produce the first complete computational specification for distributed, closed-loop seizure intervention, validated against clinical recordings and on embedded hardware. This will establish the algorithmic foundation for chronic animal testing under a subsequent R01.

Significance

Beyond seizure frequency, drug-resistant TLE carries risk of sudden unexpected death in epilepsy (SUDEP) at 6.3-9.3 per 1,000 person-years, alongside driving restrictions, employment limitations, and repeated hospitalizations that compound the disease burden [19]. TLE seizure propagation generally originates in the hippocampus and entorhinal cortex, then passes through the amygdala and anterior thalamus to the temporal and orbitofrontal neocortex; however, the propagation dynamics are far more complex than a linear spread. Functional network analysis of intracranial EEG has revealed that seizure networks evolve according to distinct topographical progressions, including initial cortical synchrony and fragmentation, with network hubs contributing significantly to both seizure propagation and termination [7]. Studies using graph-theoretically based models suggest that epileptic brains have atypical properties across networks, including changes in clustering, path length, and hub organization that extend beyond the onset area [8]. These network dynamics are discoverable before clinical seizure onset. Multi-day iEEG studies have demonstrated that degree (amount of influence within the network) and betweenness centrality (connectedness within the network) change in seizure-onset zone nodes up to seconds before seizure onset [9], and that spatial-temporal network reconfigurations precede seizures across macro-scale networks [10]. Recent sEEG studies have further shown that TLE involves many structures outside of the temporal lobe, including insular, opercular, and orbitofrontal regions [11], and that seizure propagation tends to follow well-defined white matter pathways, including the cingulate fasciculus, inferior longitudinal fasciculus, and corpus callosum, with distinct patterns of propagation that correspond to tract involvement [12]. The broader clinical community has acknowledged the network-based perspective of TLE, with the sEEG becoming the gold standard for pre-surgical evaluation at 92% of US epilepsy centers [13]. Effective closed-loop intervention, therefore, must match the architecture of the disorder and act across the system based on network-wide monitoring.

The Limitations of Current Interventions

Existing therapeutic approaches for drug-resistant TLE provide a focal solution for a network disease. Anterior temporal lobectomy achieves seizure freedom in 60-70% of pre-selected candidates at 1-2 years post-op, which declines to ~50% at 5 years and 47% at 10 years [14, 15, 16]. While surgery has been shown to be the strongest evidence-based intervention for TLE [14], it can cause irreversible damage, excludes bilateral TLE, and, by definition, removes only a single structure from a distributed network. Current RNS approaches demonstrate improvements from modest early reductions to 75% median seizure reduction after nine years post-implant, suggesting plasticity-mediated therapeutic benefits of chronic stimulation [2]. Real-world data confirm this pattern, with an 82% median reduction at ≥3 years across eight epilepsy centers; however, only 18% of patients achieved complete seizure freedom [3] — a ceiling we attribute to the 2-lead/8-contact focal system rather than to the closed-loop concept itself.

Deep-brain stimulation of the anterior nucleus of the thalamus has demonstrated, through the SANTE trial, a 56% median reduction in seizures at 2 years [17], improving to 69% at 5 years [18] and 75% at 7 years [19]. These results parallel the findings from RNS trials despite the differing structure, suggesting that both approaches modulate key nodes in the seizure network. Interestingly, the sudden unexpected death in epilepsy rate of 2 per 1000 people per year was below the 6.3-9.3 expected in drug-resistant epilepsy [19], emphasizing that even partial seizure control can have life-saving benefits. The therapeutic relevance of the anterior thalamus, a propagation hub rather than an onset zone, directly supports the rationale for a distributed architecture in closed-loop monitoring.

Technological Readiness

Three independent hardware advances have converged to enable a technically feasible distributed, network-level intervention. Neurograins achieved coordinated, 48-node cortical recording with an architecture that can scale up to 770 nodes [4]. Ultrasound power delivery has been demonstrated in neural tissue at >5 cm [5], approaching the depths required for hippocampal and thalamic targets. Distributed stimulation hardware with individually addressable, coordinated nodes has been demonstrated by the same group using a daisy-chain bitmap protocol [6]. Computational modeling tools have matured, with the Epileptor neural mass model [20], combined with advances in patient-specific structural connectivity from diffusion tractography [21] enabling computational simulation of seizures. Public iEEG datasets (iEEG.org) are also available for testing and validation. The recording hardware, stimulation hardware, simulation tools, and validation datasets exist as individual entities, but what is missing (and what this proposal provides) is the computational decision layer that connects them.

Innovation

Existing seizure detection literature takes an algorithm-first approach: groups will design an algorithm and then ask if it fits on hardware. This proposal seeks to invert the process, constraining the power, thermal, bandwidth, and chronic stability to that of existing hardware. Algorithm-first approaches risk designing solutions that exceed the power or bandwidth budgets of implantable hardware, requiring costly redesign. By deriving constraints first, every algorithmic choice is guaranteed to be deployable on the target platform. State-of-the-art deep learning-based models consume 87.4 μW (SNN on Xylo neuromorphic processor [22]), which is far greater than the <10 μW of power available on wireless recording nodes after sensing and telemetry. Our approach ensures that algorithmic design is physically realizable on the target class of hardware. The core architectural insight is that the detection performance for each node is a fundamentally different quantity from the system-level performance. Instead of prioritizing individual node accuracy, which demands a relatively large amount of power, we aim to design each node as a deliberately modest detector, with high sensitivity but moderate specificity. Network-level fusion at the hub will leverage spatial consistency, temporal propagation patterns, and cross-node independence to achieve the desired performance. This is similar to ensemble learning with weak classifiers, as the spatial distribution across distinct structures guarantees error independence. The fusion layer will integrate propagation features, such as the sequential activation of the hippocampus, then amygdala, then thalamus, which focal recording is unable to achieve. This may allow targeted stimulation to impede the propagation of the seizure. The sequential probability ratio test (SPRT) has been proven optimal for sequential binary hypothesis testing as it minimizes the expected number of samples to reach a decision for any given pair of error rates [23]. While cumulative sum (CUSUM) methods have been applied to seizure detection [25], SPRT’s optimality has not been explored in the distributed, power-constrained setting, such as the proposed work. The computation underlying SPRT relies only on sums and threshold comparisons, which are operations achievable at the nanowatt-scale power, making it well-suited for ultra-low-power devices. The neurograins recording platform [4] and stimulation platform [6] have both been demonstrated in vivo, yet lack a computational bridge. This proposal outlines such a bridge, with on-node algorithms, network fusion protocols, and intervention policies for a platform-agnostic decision layer. Because the proposed architecture relies on compressed detection events, it is compatible with both RF and ultrasound-powered recording architectures.

Approach

The proposed system is structured with three components: distributed devices to record neural data that relay gamma activity features, an external receiver that fuses the per-node detection events into a network-wide classification, and distributed stimulation nodes receiving instructions from the external transceiver and applying the targeted intervention. Both in-brain hardware components are drawn from demonstrated devices [4, 5, 6]. This proposal outlines the computational decision layer that connects them. Aim 1 defines the constraints, aim 2 develops the algorithms within those constraints, and aim 3 integrates detection with fusion and stimulation.

Aim 1: Determine Computational Requirements for Recording Devices Given Realistic Constraints

1.1 Power and Thermal Budget Analysis

We will model wireless power delivery to implanted devices for both RF-powered and ultrasound-powered platforms, mimicking devices such as Neurograins [4] and StimDust [5], respectively. RF platforms have demonstrated power transfer to the cortical surface with 10 Mbps of aggregate backscatter [4], while ultrasonic platforms have demonstrated power delivery to >5 cm depth in neural tissue in a pig model [5]. Neither platform has published results that show power transfer at depths of 5-7 cm through the skull and brain in vivo. This is acknowledged as a feasibility boundary that Aim 1 will characterize. For each platform/target combination, we will allocate power across three systems, including sensing (2.4-4.5 μW [26], ADC at 10-bit/1 kHz, 1-10 μW), computation (feature extraction and SPRT), and telemetry (uplink of compressed detection events). Thermal limits will be constrained to <1ºC tissue temperature rise in adherence with FDA recommendations. If power delivery proves insufficient at depth, a minimal threshold-only detection mode (~1 μW) can transmit raw crossings to the hub for centralized processing.

1.2 Sensing and Signal Specifications

The recorded signals will consist of local field potential (LFP) activity, with a focus on high-gamma band power (70-150 Hz), as these comprise the strongest biomarkers of ictal activity. Electrode geometry, expected signal-to-noise ratio (SNR), and sampling rate requirements will be specified based on electrode impedance and tissue characteristics. Chronic signal degradation due to glial encapsulation will be explicitly modeled as electrode impedance typically increases 2-3 times over the first 6 months, with SNR degrading ~30-50% [27]. These chronic conditions outline the signal constraints Aim 2 will operate under.

1.3 Communication Constraints

We will define the uplink bandwidth per device, specify a 22-byte detection event packet (e.g., address, timestamp, detection event, confidence, compressed features, signal quality, checksum), and design a multi-access protocol based on the spatial time division multiple access (TDMA) with event-triggered priority [4]. At the maximum detection rate of 10 events/second across 30 nodes, the total bandwidth is 6.6 kBps, which is well within demonstrated RF backscatter capacity [4] and ultrasonic uplink capability [5]. Latency from on-device detection to hub reception will also be modeled for each device type.

Aim 2: Develop and Empirically Validate On-device Sequential Detection Algorithms

2.1 Feature Extraction Under Power Constraints

We will design features that are computable within the Aim 1 power constraints, targeting <10 μW power consumption under continuous processing. Candidates include high-gamma bandpower (70-150 Hz), line length, and zero-crossing rate. Bandpower requires a bandpass filter and integration (~10-100 nW), line length, and zero-crossing rate require only comparisons and accumulation. Feature discriminability will be validated on iEEG datasets.

2.2 Sequential Decision Policy (SPRT)

Seizure detection will be designed as a sequential binary hypothesis test with H0 (baseline) vs H1 (seizure). At each time step, the device will compute the log-likelihood ratio:

$$\lambda(t)=\log\frac{P(\mathrm{data}|\mathrm{seizure})}{P(\mathrm{data}|\mathrm{baseline})}$$

And accumulate a running sum $\Lambda(t)$. When $\Lambda(t)$ exceeds an upper threshold, a seizure is declared, and when $\Lambda(t)$ passes a lower threshold, baseline is declared. SPRT minimizes the expected number of samples to decision for given error rates [23, 24]. We will derive analytical expressions for the detection delay and false positive rate. Additionally, we will compare SPRT against CUSUM [25] and Shiryaev-Roberts procedures. Finally, we will characterize sensitivity, specificity, and latency as functions of power allocation for each node. If SPRT's stationarity assumption is violated by seizure dynamics or electrode drift, CUSUM provides a distribution-free fallback. Poor single-node specificity is expected and motivates Aim 3's network-level fusion.

2.3 Adaptive Calibration for Chronic Signal Changes

To maintain stable, chronic detection despite signal drift, we will design an adaptive system that uses an exponential moving average of the feature distributions to recalculate the baseline. This system will be designed such that the time constant is slow enough that seizures are not tracked as baseline shifts, but fast enough to track changes on the order of weeks to months. Performance will be simulated under the constraints outlined in Aim 1, with the goal of achieving stable false positive rates.

2.4 Validation on Clinical iEEG Dataset

Algorithms will be validated on the SWEC-ETHZ dataset (116 annotated seizures, 2,656 hours continuous iEEG, 18 patients, 512 Hz) as the primary evaluation, with iEEG.org used for replication across electrode configurations. SPRT, CUSUM, and line-length detectors will be applied to the same seizure epochs. The primary outcome is sensitivity at a false positive rate of ≤ 1/hour. The secondary outcomes include detection latency and specificity. Paired sensitivity comparisons will use McNemar's test on per-seizure detection outcomes. With 116 seizures, a baseline sensitivity of ~70% (line-length), and a target of >90%, we expect approximately 23 discordant seizures favoring SPRT, yielding >95% power at α=0.05. To account for within-patient clustering (18 patients), we will perform a sensitivity analysis comparing per-patient median sensitivity across methods. Detection latency differences will be assessed using bootstrap confidence intervals on paired per-seizure latency.

2.5 Hardware-in-the-loop Validation

The entire detection pipeline will be implemented on an ARM Cortex-M0 embedded platform. Pre-recorded iEEG will be streamed in real time to determine real-world power consumption, memory footprint, and detection latency. These measurements will confirm adherence to Aim 1 power constraints and provide a demonstration that the proposed system is realizable on resource-limited embedded hardware.

Aim 3: Design a Distributed Fusion and Closed-loop Intervention System

3.1 Multi-node Fusion Under Telemetry Constraints

We will compare three fusion strategies across telemetry constraints (bandwidth, latency, packet loss) for both RF and ultrasonic uplinks: centralized likelihood aggregation, which transmits all log-likelihood ratios to the hub for Bayesian fusion but requires the largest bandwidth; distributed spatial voting, in which each node casts a binary decision; and hierarchical fusion, which groups adjacent nodes into clusters that fuse locally before forwarding to the hub. We will identify the Pareto-optimal tradeoff between detection accuracy and communication load for each platform.

3.2 Network-level Seizure Detection and Localization

We will simulate seizure propagation using the Epileptor neural mass model [20] with coupled Epileptor networks connected via slow permittivity coupling [28] and physiological connectivity from diffusion tractography [21]. Virtual nodes will process simulated LFP data and use the algorithms outlined in Aim 2 to calculate features, which will be fused at the hub. We will benchmark network-level fusion against single-node detection and the RNS line-length paradigm, which achieves approximately 65-75% sensitivity at clinically acceptable false-positive rates. To demonstrate a meaningful network-level advantage, we target >90% sensitivity at ≤1 false positive per hour with detection-to-intervention latency under 500 ms, within the seconds-scale window of pre-seizure network reconfiguration identified by Bröhl et al. [9]. For simulation-based comparisons, we will generate ≥200 seizure events per parameter configuration using the Epileptor model, providing >90% power to detect a 15 percentage-point sensitivity difference between fusion and single-node detection at α=0.05. We will sweep across degraded operating conditions, including packet loss (0-20%), SNR reduction (0-50%), and varied structural connectivity, to characterize robustness. Comparisons between fusion strategies and the focal baseline will use permutation tests on sensitivity and latency distributions.

3.3 Closed-loop Intervention Protocol

The intervention protocol will map the hub’s fusion output (seizure classification including localization data) to a stimulation bitmap similar to the daisy-chain protocol introduced by Lee et al. [6]. Stimulation will use biphasic charge-balanced pulses (100-200 μs/phase, 1-10 Hz burst frequency, ≤120 μA per the Lee 2025 hardware ceiling), constrained to satisfy the Shannon safety criterion (k ≤ 1.75) with charge density held below 30 μC/cm²/phase to ensure chronic tissue safety [29]. The downlink transmission will contain node-specific stimulation parameters through a single command. We will compare three strategies (reactive stimulation near the onset zone, preemptive stimulation triggered by preictal signals, and propagation disruption at downstream network nodes) on simulated seizure duration, probability of termination within 10 seconds of intervention onset, and total charge delivered per event. If propagation outpaces reactive stimulation, pre-computed intervention templates indexed by seizure type reduce real-time latency to template selection only, and preemptive mode can trigger during the preictal window [9, 10].

Timeline and Milestones

Milestones: Power budget table; per-node operating curves; embedded power data; full system specification.

References

[1] Asadi-Pooya AA, et al. (2017). Prevalence and Incidence of Drug-Resistant Mesial Temporal Lobe Epilepsy in the United States. World Neurosurg, 99, 662–666.

[2] Nair DR, et al. (2020). Nine-year prospective efficacy and safety of brain-responsive neurostimulation for focal epilepsy. Neurology, 95(9), e1244–e1256.

[3] Razavi B, Rao VR, Lin C, et al. (2020). Real-world experience with direct brain-responsive neurostimulation for focal onset seizures. Epilepsia, 61(8), 1749–1757.

[4] Lee J, et al. (2021). Neural recording and stimulation using wireless networks of microimplants. Nat Electron, 4, 604–614.

[5] Piech DK, et al. (2020). A wireless millimetre-scale implantable neural stimulator with ultrasonically powered bidirectional communication. Nat Biomed Eng, 4, 207–222.

[6] Lee J, et al. (2025). Wireless network of distributed implantable microelectronic neural stimulators. Nat Commun.

[7] Kramer MA & Cash SS. (2012). Epilepsy as a Disorder of Cortical Network Organization. The Neuroscientist, 18(4), 360–372.

[8] Bernhardt BC, et al. (2015). Network analysis for a network disorder: The emerging role of graph theory in the study of epilepsy. Epilepsy Behav, 50, 162–170.

[9] Bröhl T, et al. (2024). The time-evolving epileptic brain network: Concepts, definitions, accomplishments, perspectives. Front Netw Physiol, 3, 1338864.

[10] Fruengel R, et al. (2020). Reconfiguration of human evolving large-scale epileptic brain networks prior to seizures. Sci Rep, 10, 21921.

[11] Bartolomei F, et al. (2025). The different subtypes of temporal lobe seizures networks. Rev Neurol.

[12] Ahmadi N, et al. (2021). Patterns of seizure spread in temporal lobe epilepsy are associated with distinct white matter tracts. Front Neurol, 12.

[13] Delgado-Garcia, G., & Frauscher, B. (2022). Future of Neurology & Technology: Stereoelectroencephalography in Presurgical Epilepsy Evaluation. Neurology, 98(4).

[14] Engel J, et al. (2012). Early surgical therapy for drug-resistant temporal lobe epilepsy: a randomized trial. JAMA, 307(9), 922–930.

[15] de Tisi J, et al. (2011). The long-term outcome of adult epilepsy surgery, patterns of seizure remission, and relapse. The Lancet, 378(9800), 1388–1395.

[16] Mathon B, et al. (2022). Long-Term Outcome of Temporal Lobe Epilepsy Surgery in 621 Patients With Hippocampal Sclerosis. Front Neurol, 13, 833293.

[17] Fisher R, et al. (2010). Electrical stimulation of the anterior nucleus of thalamus for treatment of refractory epilepsy. Epilepsia, 51(5), 899–908.

[18] Salanova V, et al. (2015). Long-term efficacy and safety of thalamic stimulation for drug-resistant partial epilepsy. Neurology, 84(10), 1017–1025.

[19] Salanova V, et al. (2021). The SANTÉ study at 10 years of follow-up. Epilepsia, 62(6), 1306–1317.

[20] Jirsa VK, et al. (2014). On the nature of seizure dynamics. Brain, 137(8), 2210–2230.

[21] Proix T, et al. (2017). Individual brain structure and modelling predict seizure propagation. Brain, 140(3), 641–654.

[22] Li W, Zhao D, et al. (2024). Real-time sub-milliwatt epilepsy detection implemented on a spiking neural network edge inference processor. Comput Biol Med, 183, 109225.

[23] Wald A. (1945). Sequential Tests of Statistical Hypotheses. Ann Math Stat, 16(2), 117–186.

[24] Tartakovsky AG, et al. (2014). Sequential Analysis: Hypothesis Testing and Change Point Detection. CRC Press.

[25] Zandi AS, et al. (2010). Automated real-time epileptic seizure detection in scalp EEG recordings using an algorithm based on wavelet packet transform. IEEE TBME, 57(7), 1639–1651.

[26] Qian C, et al. (2013). A micropower ultra-low-noise neural recording front-end circuit for epileptic seizure detection. IEEE TBCAS, 7(4), 540–551.

[27] Salatino JW, Ludwig KA, Kozai TDY, Purcell EK. (2017). Glial responses to implanted electrodes in the brain. Nat Biomed Eng, 1, 862–877.

[28] Proix T, et al. (2014). Permittivity coupling across brain regions sets the timescale of seizure recruitment. J Neurosci, 34(45), 14769–14778.

[29] Shannon RV. (1992). "A model of safe levels for electrical stimulation." IEEE Transactions on Biomedical Engineering, 39(4), 424–426. DOI: 10.1109/10.126616.