Title: ArrythML: A TinyML Autoencoder Framework for On-Device Arrhythmia Detection in Resource-Constrained Embedded Systems

Abstract:

This study introduces a novel approach for real-time, on-device ECG segmentation and arrhythmia detection utilizing Tiny Machine Learning (TinyML) models. Designed for resource-constrained embedded systems, our method employs INT8-quantized autoencoder architectures characterized by minimal layer depth and parameter counts to facilitate efficient embedded deployment. The proposed models are rigorously assessed using a custom dataset generated from the MIT-BIH Arrhythmia Database, with validation performed across both PC-based simulations and actual on-device environments.

During the evaluation phase, more than 95,000 ECG segments were processed on an ESP32-S3 microcontroller executing the TensorFlow Lite Micro runtime. Following these tests, we conducted a comprehensive analysis, including record-wise and annotation-wise failure assessments, to characterize model performance across varied ECG morphologies and rhythm patterns, thereby elucidating the causes of missed detections. Our findings indicate that certain apparent misclassifications may actually reflect early or subtle anomaly patterns that were labeled as normal in the reference annotations, underscoring the model's heightened sensitivity.

After refining the evaluation by excluding ambiguous cases within the dataset, the top-performing DNN-based autoencoder demonstrated a recall of 84% and an F1-score of 79%. This model maintains a compact size of approximately 180 KB and achieves an on-device inference latency of just 9 ms. These outcomes confirm the viability of low-power, privacy-preserving wearable embedded systems capable of executing accurate arrhythmia detection entirely on-device.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC