TinyML Seizure Detector Balances Accuracy and Uncertainty on Edge Devices

The study, published in Discover Internet of Things, introduces an uncertainty-aware few-shot detection system that combines TinyML techniques with Kolmogorov–Arnold Networks (KANs). This approach addresses the significant challenges of severe class imbalance, limited labeled ictal data, and the strict computational constraints of embedded hardware. By integrating a lightweight classifier with explicit reliability estimation, the framework aims to provide a practical solution for continuous, privacy-preserving seizure monitoring in clinical and wearable settings.

The central challenge in EEG-based seizure detection is the scarcity of labeled seizure data compared to the abundance of non-seizure activity. Traditional deep learning models often require substantial memory and computational resources, making them unsuitable for deployment on microcontrollers. Furthermore, many existing systems output only binary predictions without quantifying predictive reliability, which is insufficient for clinical decision support where ambiguous signals can lead to overconfident errors. The proposed framework addresses these issues by using a compact handcrafted feature extraction pipeline and a shallow KAN classifier that operates on a fixed-dimensional input vector derived from harmonized EEG channels.

The methodology involves a clinically motivated preprocessing pipeline that transforms raw EEG recordings into standardized, denoised representations. This includes channel harmonization, baseline correction, band-pass filtering between 0.5 and 40 Hz, and wavelet-based denoising to remove transient artifacts. The signals are then segmented into overlapping four-second windows with 50% overlap. From each window, 15 descriptors per channel are extracted, including line length, root mean square, Hjorth parameters, spectral entropy, and wavelet sub-band energies. These features are concatenated to form a 270-dimensional vector, which serves as the input to the lightweight KAN classifier, balancing descriptive richness with computational efficiency for edge deployment.

The KAN architecture replaces fixed pointwise activations with learnable univariate nonlinear functions along network edges, allowing for adaptive modeling of feature interactions without deep structures. The classifier consists of two hidden KAN layers and a final dense sigmoid output layer. To provide uncertainty-aware predictions, the framework employs Monte Carlo dropout with temperature scaling. Instead of disabling dropout at inference, the model performs multiple stochastic forward passes to estimate predictive variance, entropy, and mutual information. This allows the system to flag ambiguous EEG windows for clinician review rather than forcing an overconfident binary decision, enhancing safety in a safety-sensitive application.

Experiments were conducted on the CHB-MIT scalp EEG dataset, which contains recordings from 23 pediatric subjects, including 21 with epilepsy and two healthy controls. The evaluation used a patient-wise splitting protocol to prevent data leakage, ensuring that no patient appeared in both training and test sets. The study distinguished between three training regimes: a full-data upper-bound reference, a few-shot regime with restricted ictal supervision, and a compressed deployment regime. The full-data regime served as an architectural upper bound, while the few-shot regime simulated the clinically realistic scenario where only a limited number of labeled seizure windows are available for training.

In the full-data training regime, the full-precision KAN achieved near-perfect performance, with an accuracy of 99.98%, an F1-score of 99.98%, and an AUC-ROC of 0.9998. However, the primary focus of the study was the few-shot setting, where the number of labeled ictal training windows was restricted. Under the deployment-oriented few-shot setting with K=128 seizure windows, the compressed embedded-ready KAN achieved an accuracy of 94.87% ± 0.31%, an F1-score of 94.90% ± 0.31%, and an AUC-ROC of 0.904 ± 0.005. These results demonstrate that the model maintains strong discriminative capability even when trained on a small fraction of the available ictal data, specifically approximately 3.33% of the full training pool.

To ensure practical compatibility with IoMT edge devices, the model underwent a deployment-oriented compression strategy involving structured pruning, quantization-aware training, and post-training quantization. Structured pruning removed 70% of the KAN edge functions, reducing the parameter count from 40,065 to 12,039. Subsequent quantization to INT8 precision further reduced the model size to 11.76 kB. On a 32-bit ARM Cortex-M–class target, the compressed classifier achieved a deterministic inference time of 4.6 ms per window. When uncertainty-aware inference was enabled with 10 Monte Carlo dropout passes, the classifier-only latency increased to approximately 46.8 ms, while the end-to-end compute budget, including preprocessing and feature extraction, was estimated at 87.5 ms per window.

The study also evaluated the framework on auxiliary datasets, including the Bangalore EEG Epilepsy Dataset and the Bonn EEG Dataset, to assess the stability of the compression pipeline across heterogeneous EEG benchmarks. These experiments were designed to test whether the KAN architecture and its compression strategy remained stable under dataset-specific protocols, rather than to claim direct clinical cross-dataset generalization. The results showed that the performance trend remained consistent, with the compressed model exhibiting a controlled reduction in accuracy compared to the full-precision variant, confirming the robustness of the design across different recording structures.

Uncertainty quantification analysis revealed that while compression introduced some degradation in probabilistic calibration, the uncertainty-aware rejection mechanism remained effective. The final compressed model exhibited higher Brier scores and Expected Calibration Error compared to the full-precision model, but the selective accuracy on retained samples stayed markedly higher than the raw classification accuracy. This indicates that the mutual-information-based filtering successfully rejected the most ambiguous cases, providing a reliability estimate that supports safer decision-making in clinical contexts where overconfident false predictions are particularly undesirable.

Limitations of the study include its reliance on benchmark EEG datasets, which may not fully capture the variability of real-world clinical environments, such as diverse electrode placements and comorbidities. The framework also depends on a fixed-dimensional handcrafted feature representation, which may constrain the model’s ability to capture subtle seizure dynamics that raw-signal representations might preserve. Additionally, the TinyML deployment analysis demonstrates feasibility on a representative microcontroller but does not constitute a full end-to-end wearable clinical deployment study. Future work will focus on broader validation across more heterogeneous cohorts and the development of subject-adaptive extensions to handle inter-patient variability more effectively.

The proposed framework offers a promising direction for next-generation intelligent epilepsy monitoring systems by balancing predictive performance, decision reliability, and TinyML deployment feasibility. By integrating few-shot learning, uncertainty quantification, and compression-aware design into a single edge-oriented framework, the study addresses the critical need for low-latency, privacy-preserving seizure detection in resource-constrained settings. The results suggest that lightweight KAN-based classifiers can provide clinically useful performance while satisfying the strict memory and latency constraints of IoMT devices, paving the way for more accessible and continuous seizure monitoring solutions.

The research team, including Chaymae Yahyati, Ismail Lamaakal, Khalid El Makkaoui, and Ibrahim Ouahbi, designed the system to operate under strict resource constraints typical of wearable medical devices. The study utilized the CHB-MIT scalp EEG dataset, which comprises recordings from 23 pediatric subjects, including 21 with epilepsy and two healthy controls. To ensure methodological rigor and prevent data leakage, the evaluation employed a patient-wise splitting protocol, assigning 65% of patients to training, 15% to validation, and 20% to testing. This separation is critical in seizure detection because EEG morphology is highly subject-dependent; allowing windows from the same patient to appear in both training and test sets would artificially inflate performance metrics by enabling the model to memorize subject-specific patterns rather than learning generalizable seizure signatures.

The preprocessing pipeline was engineered to handle the non-stationary nature of EEG signals while maintaining compatibility with edge hardware. After channel harmonization and baseline correction, the signals underwent band-pass filtering between 0.5 and 40 Hz using a zero-phase Butterworth filter to preserve waveform morphology without phase distortion. Wavelet-based denoising was applied to remove transient artifacts while retaining localized seizure-related structures.

The few-shot training strategy was designed to simulate clinically realistic scenarios where labeled ictal data is scarce. The model was evaluated across a unified shot grid, with K representing the number of labeled seizure windows retained from training patients. Performance improved consistently as K increased, with the full-precision few-shot model’s F1-score rising from 73.96% at K=1 to 96.36% at K=128. In the compressed few-shot setting, the F1-score improved from 71.75% to 94.90% over the same range. This trend indicates that the lightweight KAN architecture can efficiently exploit scarce positive supervision to refine its decision boundary. The deployment-oriented setting with K=128 utilized approximately 3.33% of the available ictal training windows, demonstrating that high performance can be maintained with a small fraction of the full data pool.

Event-level metrics were included to assess clinical utility beyond window-level accuracy. The study measured event sensitivity, false alarms per hour, and detection delay to evaluate how the system performs in continuous monitoring scenarios. A seizure event was considered detected if at least one predicted positive window overlapped the annotated interval or occurred within a short tolerance window after onset. The results showed that the embedded-ready KAN maintained stable clinical behavior across test patients, with some variability attributable to differences in seizure morphology and recording duration. This patient-wise aggregation ensures that pooled metrics do not mask lower performance in specific subjects, providing a more comprehensive view of the system’s reliability in diverse clinical contexts.

The auxiliary experiments on the Bangalore EEG Epilepsy Dataset and the Bonn EEG Dataset were conducted to assess the stability of the compression pipeline under different data structures. These datasets differ substantially from CHB-MIT in recording organization and label definition, so the same event-level protocol could not be directly transferred. The results confirmed that the KAN architecture and its compression strategy remained stable across these heterogeneous benchmarks, although the interpretation was limited to cross-benchmark consistency rather than direct clinical generalization. This distinction is important for understanding the scope of the framework’s applicability to different EEG acquisition environments.

Subject of Research: Development of an uncertainty-aware few-shot EEG seizure detection framework for TinyML edge devices using Kolmogorov–Arnold Networks.

Article Title: Uncertainty-aware few-shot epileptic seizure detection for IoMT edge devices using TinyML and Kolmogorov–Arnold networks

Article References: Yahyati, C., Lamaakal, I., El Makkaoui, K., & Ouahbi, I. (2026). Uncertainty-aware few-shot epileptic seizure detection for IoMT edge devices using TinyML and Kolmogorov–Arnold networks. Discover Internet of Things, 6(1), Article 187. https://doi.org/10.1007/s43926-026-00514-4

Image Credits: AI Generated

DOI: 10.1007/s43926-026-00514-4

Keywords: TinyML, Epileptic Seizure Detection, EEG, Kolmogorov–Arnold Networks, IoMT, Uncertainty Quantification, Few-Shot Learning, Edge Computing, Medical Devices, Uncertainty-aware, few-shot, epileptic