Closed-Loop Neurostimulation Protocols for Sub-Second Seizure Prediction in Wearable Arrays

Closed-loop neurostimulation for seizure prediction promises a new frontier in epilepsy management: intervene before the storm hits. But building a system that predicts seizures in under a second using wearable arrays—while keeping false alarms low enough for daily use—is a formidable engineering challenge. This guide is for teams already familiar with EEG signal processing and neuromodulation; we skip the basics and focus on the trade-offs, failure modes, and design decisions that separate a lab prototype from a viable therapeutic protocol.

We'll walk through who needs sub-second prediction, what prerequisites you must settle first, a core workflow, tooling realities, variations for different constraints, debugging when things go wrong, and a final FAQ. Each section assumes you have hands-on experience with neural signal acquisition and basic control theory.

Who Needs Sub-Second Prediction and What Goes Wrong Without It

Patients with drug-resistant epilepsy who experience frequent, unpredictable seizures are the primary candidates. But not all seizure types benefit equally from ultra-fast prediction. Focal aware seizures with a gradual onset may tolerate a longer prediction window, while generalized tonic-clonic seizures require sub-second response to abort the ictal phase. Without closed-loop control, patients rely on rescue medication or manual stimulation, which often arrives too late. The consequences range from injury to status epilepticus.

In practice, many teams start with offline seizure detection and then attempt to close the loop. The first failure point is latency: even a 500-millisecond delay can render stimulation ineffective if the seizure propagates too quickly. The second is specificity: a system that triggers multiple false stimulations per day will be abandoned by patients. Without sub-second prediction, the trade-off between sensitivity and false-alarm rate becomes untenable. For example, a detector that catches 90% of seizures but produces one false alarm per hour would cause dozens of unnecessary stimulations daily, eroding trust and potentially causing adverse effects.

Another hidden issue is electrode drift. Wearable arrays shift subtly during daily activities, altering the spatial pattern of neural activity. A model trained on static electrode positions degrades over hours, leading to missed seizures or increased false alarms. Closed-loop protocols must adapt in real time, but naive adaptation can lock onto artifact patterns. We have seen projects where the system learned to predict head movement instead of epileptiform activity, triggering stimulations during normal activity—a dangerous failure.

Finally, without sub-second prediction, the loop cannot leverage early ictal EEG changes that occur 0.5–2 seconds before clinical onset. These micro-seizures or pre-ictal spikes are low-amplitude and easily masked by muscle artifact. A system that waits for clear electrographic seizure onset has already lost the therapeutic window. The goal is to detect the transition from interictal to pre-ictal state, not the seizure itself.

Prerequisites and Context to Settle First

Before designing a closed-loop protocol, you need a clear definition of the seizure type and its EEG signature. This requires annotated long-term recordings from the patient—ideally with simultaneous video and clinical onset markers. Without a labeled dataset of at least 10 seizures per patient, machine learning models will overfit to noise. We recommend a minimum of 50 hours of continuous wearable array data with at least 20 seizure events for initial model development.

Hardware prerequisites include a wearable array with at least 16 channels, sampling at 256 Hz or higher, and a low-latency wireless link to a processing unit (edge or cloud). The electrodes must be dry or semi-dry for long-term wear; wet electrodes dry out and increase impedance over hours. For sub-second prediction, the processing pipeline must complete feature extraction and classification in under 200 milliseconds to leave room for stimulation delivery and biological response time.

On the algorithmic side, you need a feature set that captures pre-ictal dynamics. Common choices include spectral power ratios (e.g., delta/theta, alpha/beta), phase-amplitude coupling, and cross-channel coherence. But these features must be computed over short windows (0.5–1 second) to achieve sub-second latency. This creates a trade-off: longer windows improve feature stability but increase delay. We have found that a sliding window of 0.8 seconds with 0.4-second overlap works well for most focal seizures, but you should optimize this for your specific patient population.

Another prerequisite is a stimulation modality that can be delivered rapidly and safely. Options include transcranial electrical stimulation (tES), vagus nerve stimulation (VNS), or focused ultrasound. tES is the easiest to integrate with wearable arrays but has limited depth penetration. VNS requires an implanted device, which adds surgical risk. Focused ultrasound is non-invasive but still experimental for seizure abort. The choice of stimulator dictates the control signal: a binary on/off pulse, a graded intensity ramp, or a patterned train. The protocol must specify the stimulation parameters (amplitude, frequency, duty cycle) and the safety constraints (maximum charge density, duty cycle limits).

Finally, you need a validation framework that goes beyond offline accuracy. Closed-loop systems must be tested in a mock ambulatory setting with realistic artifacts (walking, chewing, eye movements). We recommend a staged approach: first, offline replay of prerecorded data; then, real-time simulation with synthetic artifacts; then, human-in-the-loop testing with healthy volunteers performing daily activities; and finally, a small pilot study with patients under clinical supervision. Each stage should measure prediction latency, false-alarm rate, and stimulation efficacy (e.g., seizure abort rate or reduction in seizure duration).

Core Workflow: Sequential Steps for Sub-Second Closed-Loop Neurostimulation

The core workflow consists of five stages: (1) signal acquisition and preprocessing, (2) feature extraction, (3) classification and prediction, (4) stimulation decision and delivery, and (5) feedback adaptation. Each stage must complete within the sub-second budget.

Stage 1: Acquisition and Preprocessing

Wearable array signals are sampled at 256 Hz, bandpass filtered from 0.5 to 100 Hz, and notch filtered at 50/60 Hz. Artifact rejection is critical: we use a threshold-based rejection (e.g., amplitude > 500 µV) combined with a template subtraction for eye blinks and jaw clenches. The preprocessing must be causal (no future samples) and computationally light—a finite impulse response (FIR) filter with 50 taps is usually sufficient. We recommend implementing preprocessing on the wearable's microcontroller to reduce wireless bandwidth.

Stage 2: Feature Extraction

From the preprocessed signals, compute a feature vector every 200 milliseconds (overlapping windows). For each channel, extract: log-power in delta (0.5–4 Hz), theta (4–8 Hz), alpha (8–13 Hz), beta (13–30 Hz), gamma (30–50 Hz); pairwise phase-locking value (PLV) between channels; and the ratio of gamma to delta power. This yields a feature vector of dimension (16 channels × 5 bands) + (120 channel pairs) + 1 ratio = 201 features per window. To reduce dimensionality, apply a patient-specific principal component analysis (PCA) trained on interictal data, keeping components that explain 95% of variance. The PCA transform is fixed after training to avoid latency from online updates.

Stage 3: Classification and Prediction

We use a lightweight gradient-boosted tree (e.g., LightGBM) trained on feature windows labeled as pre-ictal (the 10 seconds before seizure onset) vs. interictal (all other windows). The model outputs a probability of pre-ictal state. To reduce false alarms, we apply a persistence filter: the prediction must exceed a threshold (e.g., 0.7) for three consecutive windows (600 ms) before triggering stimulation. This introduces a 600 ms delay, but the total pipeline latency (preprocessing + feature extraction + classification) should be under 400 ms, keeping total prediction-to-stimulation under one second. The threshold and persistence length are tunable per patient; we start with a threshold that gives 90% sensitivity on training data and adjust based on false-alarm rate in validation.

Stage 4: Stimulation Decision and Delivery

When the prediction triggers, the control unit sends a stimulation command to the stimulator. For tES, we deliver a 2 mA, 200 µs biphasic pulse train at 100 Hz for 2 seconds. The stimulation is delivered to the electrode pair closest to the epileptic focus (identified from prior EEG source localization). If the seizure is not aborted after 2 seconds, a second stimulation can be delivered, but with a minimum 5-second gap to avoid tissue damage. The stimulator must have a latency under 10 ms from command to delivery.

Stage 5: Feedback Adaptation

After each stimulation, the system monitors the EEG for 30 seconds. If no electrographic seizure appears, the stimulation is considered successful, and the prediction threshold is slightly lowered (e.g., by 0.02) to increase sensitivity. If a seizure occurs despite stimulation, the threshold is raised by 0.05 to reduce false alarms. This online adaptation must be slow to avoid oscillation; we use a multiplicative decay factor (0.95) so that thresholds drift back to baseline over hours.

Tools, Setup, and Environment Realities

The hardware stack typically includes a wearable EEG headband or cap (e.g., with dry electrodes), a microcontroller (e.g., Nordic nRF52840) for on-board preprocessing, a smartphone or edge device (e.g., Raspberry Pi 4) for feature extraction and classification, and a stimulator (e.g., a programmable tES device like the Starstim). For development, we use Python with scikit-learn and LightGBM, but for deployment, the model is converted to C or TensorFlow Lite for edge inference. The communication between devices uses Bluetooth Low Energy (BLE) for low power, but BLE latency can be 30–100 ms; we recommend a dedicated 2.4 GHz radio for the stimulation command to keep latency under 10 ms.

Environment setup requires a controlled testing space with a Faraday cage to minimize external interference, though for ambulatory testing, a shielded room is impractical. In practice, the system must be robust to Wi-Fi, Bluetooth, and cellular signals. We have found that using differential amplifiers and twisted-pair cables for electrode connections reduces common-mode noise. The wearable array should be secured with a cap or adhesive to minimize movement artifacts; even a 1 mm shift can change impedance by 10 kΩ.

Software tools include a real-time visualization dashboard (e.g., using Plotly Dash) for monitoring signal quality, feature trends, and prediction outputs. The dashboard is essential during debugging to catch artifact contamination or model drift. For data logging, we use a circular buffer that stores the last 10 minutes of raw EEG, features, and predictions; this helps reconstruct false alarms or missed seizures post-hoc. The logging must be low-overhead; we write to a microSD card in batches every 30 seconds to avoid file system corruption.

A major environment reality is battery life. The wearable array, processing unit, and stimulator must run for at least 12 hours on a single charge. The microcontroller should sleep between processing cycles; the BLE radio should be duty-cycled. We achieve 10-hour battery life with a 2000 mAh battery for the processing unit and a 500 mAh battery for the stimulator. The patient should charge the system overnight.

Variations for Different Constraints

Not every use case needs the same design. Here we outline three common variations and their trade-offs.

Low-Power Wearable for Long-Term Monitoring

For patients who need continuous monitoring for weeks, power consumption is paramount. The prediction model is simplified to a linear classifier (e.g., logistic regression) on a reduced feature set (e.g., only band power from 4 channels). The stimulation threshold is fixed after initial calibration to avoid online adaptation. This reduces sensitivity (maybe 70% seizure detection) but extends battery life to 24 hours. The trade-off is more missed seizures; this variant is suitable for patients with slow-progressing focal seizures where the risk of injury is low.

High-Accuracy Clinical System

In a hospital setting, power is not a constraint, and the system can use a full 32-channel array with a deep learning model (e.g., a small convolutional neural network) that runs on a laptop. The prediction window can be longer (1.5 seconds) because stimulation is delivered by a nurse. This achieves >95% sensitivity with <1 false alarm per day. The downside is the patient is tethered to a cart, limiting mobility. This variant is used for pre-surgical evaluation or for patients with very frequent seizures.

Implantable Closed-Loop System

For patients with a known focal epilepsy, an implantable system (like the RNS) offers direct cortical recording and stimulation. The prediction algorithm runs on an implanted processor with severe power and memory constraints. Features are limited to line length and band power from 4 channels. The stimulation is delivered directly to the seizure focus. This system has the lowest latency (under 100 ms) and highest specificity, but requires surgery. It is the gold standard for patients with refractory focal epilepsy who are candidates for resection.

Pitfalls, Debugging, and What to Check When It Fails

Even with a well-designed protocol, closed-loop systems fail in predictable ways. Here are the most common pitfalls and how to diagnose them.

False Alarms from Artifacts

The most frequent failure is the system triggering stimulation during normal activities like chewing, walking, or scratching. To debug, review the raw EEG traces around false alarms. Look for high-frequency spikes that match muscle activity (e.g., 20–60 Hz bursts). If the artifact is consistent, add a template subtraction or a notch filter at the artifact frequency. If the artifact is variable, consider using a separate accelerometer channel to detect movement and inhibit stimulation during high-motion periods.

Missed Seizures Due to Model Overfitting

A model that performs well on training data but misses seizures in the real world is usually overfit to the specific recording conditions (e.g., electrode placement, time of day). To diagnose, compare the feature distributions between training and deployment. If they differ, retrain the model with more varied data (including different electrode placements and times). Another cause is concept drift: the patient's seizure pattern changes over days or weeks. Implement a periodic retraining schedule (e.g., every night) using the most recent 24 hours of data, but only if the system can reliably label true positives (e.g., via patient diary).

Latency Bottlenecks

If the prediction-to-stimulation time exceeds one second, trace each stage: preprocessing, feature extraction, classification, communication, stimulation delivery. Use timestamps at each stage. Common bottlenecks include: software filtering (use hardware filters instead), BLE transmission (switch to a wired connection or a faster radio), and model inference (quantize the model to int8). We have seen cases where the Python interpreter introduced 200 ms jitter; moving inference to C reduced latency by 80%.

Stimulation Inefficacy

If the stimulation does not abort the seizure, check the stimulation parameters: is the current sufficient? Is the electrode pair over the seizure focus? Use EEG source imaging to verify electrode placement. Also check the stimulation timing: if the prediction is too late (e.g., after the seizure has already started), the stimulation may be ineffective. Shorten the prediction window or lower the threshold. Finally, consider that not all seizures are abortable; some patients may need a different stimulation modality (e.g., VNS instead of tES).

When debugging, keep a log of every stimulation event with pre- and post-stimulation EEG. This log is invaluable for root cause analysis. We recommend a standardized debugging checklist: (1) verify raw signal quality, (2) check feature extraction, (3) examine model output, (4) confirm stimulation command, (5) verify stimulator delivery, (6) compare with patient diary. Most issues are caught at step 1 or 2.

FAQ: Common Questions from Experienced Teams

How do we handle multiple seizure foci? If a patient has more than one focus, train separate models for each focus, or use a single model that outputs a probability per focus. The stimulation can be delivered to the focus with the highest probability, or to all foci simultaneously if the stimulator has multiple output channels. Be aware that stimulating multiple sites may cause unexpected interactions; test in a controlled setting first.

Can we use the same features for all patients? No. While band power is universal, the optimal feature set varies by seizure type and individual. We recommend starting with a large feature set and using automated feature selection (e.g., recursive feature elimination) on each patient's data. Expect that 20–30 features will be sufficient for most patients.

How do we validate the system without causing harm? Start with offline replay of recorded data to measure sensitivity and false-alarm rate. Then, test in a simulated environment with healthy subjects to ensure the stimulation is safe and tolerable. Finally, conduct a pilot study in a hospital with continuous monitoring by a neurologist. The study should have clear stopping rules: if the false-alarm rate exceeds 5 per hour or if any adverse event occurs, the system is paused.

What is the minimum dataset size for training? We recommend at least 20 seizures per patient, with at least 50 hours of interictal data. If fewer seizures are available, consider transfer learning from a larger dataset of similar seizure types, but be cautious about domain shift. Data augmentation (e.g., adding synthetic noise) can help, but it does not replace real seizure data.

How do we ensure patient safety during stimulation? Follow established safety guidelines: maximum charge density of 25 µC/cm² per phase, maximum current of 10 mA, and maximum duty cycle of 10%. Use a hardware current limiter and a watchdog timer that shuts off stimulation if the system freezes. The patient should have an emergency stop button. All stimulation parameters must be reviewed by a qualified medical professional before use.

Can the system run entirely on the wearable? Currently, the computational demands of real-time classification and adaptation exceed the capabilities of most microcontrollers. A hybrid approach with a smartphone or edge device is more practical. As hardware improves, fully on-wearable processing may become feasible in 3–5 years.

What is the most common mistake teams make? Underestimating the impact of artifacts. Many teams optimize on clean, lab-recorded data and then fail in ambulatory settings. We recommend spending 30% of development time on artifact detection and rejection. The second most common mistake is neglecting feedback adaptation: a static threshold will drift out of calibration as the patient's EEG changes over days.

This guide provides a foundation, but every patient is unique. The next step is to implement a minimal viable system on your hardware, test it with your own data, and iterate. Start with the core workflow, add variations as needed, and always validate under realistic conditions.