TS-MedFuse: A Remote Monitoring Framework for Cardio-Respiratory Patients using Self-Supervised Temporal Signal Fusion and Cross-Attentional Transformers in IoMT

Remote monitoring for patients with cardiopulmonary conditions has recently attracted significant attention, for valid reasons. With IoMT devices, there's real potential to continuously monitor people in real time, without them having to be in a hospital. The primary challenge lies in extracting clinically useful information from heterogeneous data sources, including ECG, PPG, rPPG, and other physiological signals. The signals tend to be noisy, sensor readings are often misaligned, and labeled data is perpetually scarce. Most existing methods fail to capture the holistic context, particularly temporal dynamics and cross‑signal relationships. That usually means compromised accuracy and models that aren't robust enough for real-world use. To address this gap, we propose TS-MedFuse. It's a self-supervised framework that uses cross-attentional transformers to pull together data from multiple signal types and make sense of them as a whole. Because it learns from raw data, it doesn't need massive labeled datasets to get up to speed. This also enhances its resilience to noise, artifacts, and other common signal corruptions. We ran it through its paces on MIMIC-IV-ECG, a paired PPG-ECG dataset, and the MMPD dataset. In all cases, TS-MedFuse outperformed recent state-of-the-art methods, achieving gains of up to 5% in accuracy, precision, recall, F1-score, AUROC, and MCC.