Robust and Practical Machine Learning Solutions for Wearable Biomedical Edge Devices
OPEN ACCESS
Loading...
Author / Producer
Date
2025
Publication Type
Doctoral Thesis
ETH Bibliography
yes
Citations
Altmetric
OPEN ACCESS
Data
Rights / License
Abstract
The growing need for real-time, continuous biosignal processing outside clinical settings necessitates performing analysis directly on wearable edge devices for improved health monitoring, user privacy, and low latency. This trend impacts clinical applications as well as human-machine interfaces. This thesis develops novel machine learning (ML) frameworks optimized for energy-efficient biosignal processing on such devices, focusing on practical, real-life operation.
We first establish efficient ML foundations for edge devices, developing the ECG-TCN framework for cardiac arrhythmia detection (94.2% accuracy, 0.1 mJ/inference) and addressing noisy electroencephalography (EEG) signals via artifact handling techniques that reduce false positives in seizure detection by up to 96%. To enhance robustness, our BrainFuseNet framework fuses heterogeneous sensor data (EEG, photoplethysmography (PPG), accelerometer (ACC)) and employs a custom sensitivity-specificity weighted cross-entropy (SSWCE) loss, achieving highly specific seizure detection (<1 FP/day at 0.11 mJ/inference). Furthermore, we address signal variability through on-device continual learning frameworks capable of adapting to new data (<0.5 mJ/update) while mitigating catastrophic forgetting.
Building on these findings, we investigate advanced deep learning (DL) architectures for biosignal analysis. We demonstrate lightweight attention-based models (EEGformer) optimized for microcontroller deployment, achieving low false positive rates (e.g., 0.15 FP/h) for seizure detection. We then explore the potential of large-scale foundation models pre-trained on unlabeled data, introducing efficient architectures like CEReBrO (alternating attention) and FEMBA (state-space model) to learn universal representations. We demonstrate the adaptability of these models through cross-modal transfer (EEG pre-training applied to electrocardiogram (ECG)/PPG for blood pressure estimation) and topology-invariance techniques (LUNA), significantly reducing computational complexity while maintaining high accuracy on diverse downstream tasks.
This thesis demonstrates the feasibility and effectiveness of deploying advanced, adaptive, and energy-efficient ML models on edge devices for real-time biosignal classification. The developed frameworks, including contributions in foundation models and on-device learning, pave the way for next-generation intelligent wearables with enhanced diagnostic capabilities and personalization.
Permanent link
Publication status
published
External links
Editor
Contributors
Examiner: Benini, Luca
Examiner : Mangia , Mauro
Book title
Journal / series
Volume
Pages / Article No.
Publisher
ETH Zurich
Event
Edition / version
Methods
Software
Geographic location
Date collected
Date created
Subject
Foundation models; Machine learning; Deep learning; Embedded systems; Deployment; Seizure detection; Brain-computer interfaces
Organisational unit
03996 - Benini, Luca / Benini, Luca
Notes
Funding
193813 - PEDESITE: Personalized Detection of Epileptic Seizure in the Internet of Things (IoT) Era (SNF)