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

Abstract

Epileptic seizure detection from electroencephalography (EEG) remains challenging in real-world clinical settings due to severe class imbalance, limited labeled ictal data, inter-patient variability, and the computational constraints of edge deployment. To address these challenges, this paper proposes an uncertainty-aware few-shot epileptic seizure detection framework for Internet of Medical Things edge devices based on TinyML and Kolmogorov–Arnold Networks (KANs). The proposed method combines a clinically motivated EEG preprocessing pipeline, compact handcrafted feature extraction, a lightweight KAN classifier, Monte Carlo dropout with temperature scaling for uncertainty-aware prediction, and a deployment-oriented compression strategy based on structured pruning, quantization-aware training, and post-training quantization. The framework is designed to jointly achieve reliable seizure discrimination, robustness under scarce ictal supervision, and practical compatibility with resource-constrained embedded hardware. Experiments on the CHB-MIT dataset are reported under two explicitly separated regimes. First, a full-data training regime is used as an upper-bound reference, where the full-precision KAN reaches 99.98% accuracy, 99.98% F1-score, and an AUC-ROC of 0.9998. Second, the proposed few-shot regime restricts the number of labeled ictal training windows through the subset \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$D_K^+$$\end{document}. Under the deployment-oriented few-shot setting with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$K=128$$\end{document}, the compressed embedded-ready KAN achieves \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$94.87 \pm 0.31\%$$\end{document} accuracy, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$94.90 \pm 0.31\%$$\end{document} F1-score, and an AUC-ROC of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0.904 \pm 0.005$$\end{document}, while reducing the model size to 11.76 kB and the inference time to 4.6 ms on a 32-bit ARM Cortex-M–class target. Additional auxiliary experiments on the Bangalore EEG Epilepsy Dataset and the Bonn EEG Dataset are reported under dataset-specific protocols. These experiments are used to assess whether the compression behavior of the proposed KAN framework remains stable on heterogeneous EEG benchmarks, rather than to claim direct clinical cross-dataset generalization from CHB-MIT. The results demonstrate that uncertainty-aware few-shot seizure detection with a lightweight KAN provides an effective compromise between predictive performance, decision reliability, and TinyML deployment feasibility for next-generation intelligent epilepsy monitoring systems.