In-Situ Dehydration Monitoring via a Stable Diffusion-Aided Single-Lead ECG IoMT: ML/DL Models Shine While LLMs Hallucinate

• Published in: IEEE internet of things journal Vol. 12; no. 17; pp. 36617 - 36633
• Main Authors: Perzhilla, Levina, Siyoucef, Soumia, Al-Aslani, Rose, Rahman, Muhammad Mahboob Ur, Al-Naffouri, Tareq Y.
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 01.09.2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: 1-lead electrocardiogram (ECG); Accuracy; Artificial intelligence; Biomarkers; Classification; Data augmentation; Data collection; Data models; Datasets; Deep learning; deep learning(DL); Dehydration; Electrocardiography; Explainable artificial intelligence; explainable artificial intelligence (AI); Fasting; Hydration; Internet of medical things; Internet of Medical Things (IoMT); large language model(LLM); Large language models; Machine learning; machine learning (ML); Monitoring; noninvasive methods; Random noise; Segments; stable diffusion; Telemedicine; Testing; Training
• ISSN: 2327-4662; 2327-4662
• DOI: 10.1109/JIOT.2025.3583220

This study introduces a novel, noninvasive approach to monitor hydration status using single-lead electrocardiogram (ECG) signals. MAX86150 Internet of Medical Things (IoMT) module is utilized to collect raw 1-lead ECG data from 65 subjects under fasting and exercise conditions, creating a labeled ECG dataset. To augment the dataset, white Gaussian noise is added to the ECG segments. In addition, synthetic ECG data is generated using STABLE DIFFUSION models trained on the conditioned ECG segments. The augmented data is then fed into various machine learning (ML) and deep learning (DL) models as part of a baseline evaluation. Two classification tasks are performed: binary classification (hydrated versus dehydrated) and four-class classification, categorizing hydration level of subjects on a one to four scale. The models report a high accuracy that is up to 98.73% for binary classification, 97.41% for four-class classification in fasting subjects, and 98.32% for sportspeople, showing the potential of using single-lead ECG for hydration monitoring. Additionally, the model's decisions are interpreted using LIME-based explainable artificial intelligence (AI) technique, which identifies ECG features, such as the RR interval and QRS intervals as relevant biomarkers for dehydration. The study also investigates the use of large language models (LLMs) and large vision modelss (LVMs) to analyze sequential ECG data for hydration assessment. However, LLMs and LVMs struggle due to the time-series nature of the ECG data and fail to accurately interpret ECG graphs. While LLM and LVM results are not favorable due to hallucinations, this study provides valuable insights into the limitations of these models, paving the way for future research in this area.