Masked Generative Distillation for Robust Multi‐Lead‐to‐Single‐Lead ECG Classification on IoMT Devices
- 1School of Engineering University of Wollongong in Dubai Dubai UAE
- 2Faculty of Information Technology Majan University College Muscat Oman
Abstract
Wearable Internet of Medical Things (IoMT) devices typically rely on single‐lead electrocardiogram (ECG) acquisition, which supports ambulatory monitoring but provides less spatial information than 12‐lead clinical ECG and is more susceptible to signal corruption. Deploying accurate deep learning models on wearable hardware is also constrained by limited memory, computation and energy. This study investigates whether diagnostic knowledge from a multi‐lead ECG model can be transferred to a lightweight single‐lead model while improving robustness under wearable‐relevant noise. A teacher–student framework is proposed in which a 12‐lead 1D ResNet‐18 teacher transfers knowledge to a compact single‐lead convolutional neural network using masked generative distillation (MGD). The framework is evaluated on PTB‐XL through a controlled ablation study covering supervised learning, conventional knowledge distillation, MGD and noise augmentation. All student variants achieve comparable macro F1 performance on clean data, indicating that the main benefit lies in robustness rather than clean‐condition classification. Combining MGD with noise augmentation provides the strongest robustness under synthetic noise. Simulation‐based profiling demonstrates a 45‐fold parameter reduction, lower inference latency and minimal performance change after INT8 quantization. On the PTB‐XL five‐superclass task, the model achieves a macro F1‐score of 0.6449 versus 0.626 for MVKT‐ECG, although differences in lead selection and evaluation protocols prevent strict comparison.
Related
- Knowledge distillation-based lightweight generative large language models for standardized quality control of Chinese radiology reports
- Generative Artificial Intelligence in Periodontology: Application of OpenAI Models for the Classification of Periodontitis and Risk Assessment in Patients Undergoing Supportive Periodontal Care
- Efficient Alzheimer’s disease classification with robust rank pooling
- Privacy-preserving collaborative transfer learning with generative adversarial networks (CTL-GAN) for classification of dry eye disease
- Unmasking adversarial attacks using a robust XAI-driven approach for secure medical image classification
BibTeX
@article{shivakumar2026masked,
title = {{Masked Generative Distillation for Robust Multi‐Lead‐to‐Single‐Lead ECG Classification on IoMT Devices}},
author = {Anandu Shivakumar and Obada Al‐Khatib and Ghalia Nassreddine and Abeer ElKhouly},
journal = {Healthcare Technology Letters},
year = {2026},
month = oct,
doi = {10.1049/htl2.70108},
url = {https://doi.org/10.1049/htl2.70108},
note = {CC BY-NC-ND 4.0},
}
RIS
TY - JOUR
TI - Masked Generative Distillation for Robust Multi‐Lead‐to‐Single‐Lead ECG Classification on IoMT Devices
AU - Anandu Shivakumar
AU - Obada Al‐Khatib
AU - Ghalia Nassreddine
AU - Abeer ElKhouly
PY - 2026
DA - 2026/10/30
JO - Healthcare Technology Letters
DO - 10.1049/htl2.70108
UR - https://doi.org/10.1049/htl2.70108
AB - Wearable Internet of Medical Things (IoMT) devices typically rely on single‐lead electrocardiogram (ECG) acquisition, which supports ambulatory monitoring but provides less spatial information than 12‐lead clinical ECG and is more susceptible to signal corruption. Deploying accurate deep learning models on wearable hardware is also constrained by limited memory, computation and energy. This study investigates whether diagnostic knowledge from a multi‐lead ECG model can be transferred to a lightweight single‐lead model while improving robustness under wearable‐relevant noise. A teacher–student framework is proposed in which a 12‐lead 1D ResNet‐18 teacher transfers knowledge to a compact single‐lead convolutional neural network using masked generative distillation (MGD). The framework is evaluated on PTB‐XL through a controlled ablation study covering supervised learning, conventional knowledge distillation, MGD and noise augmentation. All student variants achieve comparable macro F1 performance on clean data, indicating that the main benefit lies in robustness rather than clean‐condition classification. Combining MGD with noise augmentation provides the strongest robustness under synthetic noise. Simulation‐based profiling demonstrates a 45‐fold parameter reduction, lower inference latency and minimal performance change after INT8 quantization. On the PTB‐XL five‐superclass task, the model achieves a macro F1‐score of 0.6449 versus 0.626 for MVKT‐ECG, although differences in lead selection and evaluation protocols prevent strict comparison.
SN - 2053-3713
ER -