Trust fusion: a weighted multi-perspective framework for ECG anomaly detection in IoMT

Continuous electrocardiogram (ECG) monitoring in Internet of Medical Things (IoMT) environments faces a fundamental precision–sensitivity trade–off. Standalone reconstruction–based detectors achieve high specificity but miss subtle anomalies, while geometric detectors capture rare patterns but generate excessive false alarms. This paper presents a hybrid Trust Fusion framework that resolves this trade–off through linearly weighted integration of an LSTM–Autoencoder, a 6–band Isolation Forest, and Z–score statistical validation, augmented by a clinically motivated Label–Free Adaptive Recovery mechanism that reclassifies borderline missed arrhythmias closest to the decision boundary. Evaluated on the MIT–BIH Arrhythmia Database under a strict inter–patient protocol (25 training records, 10 held–out test patients, 406,210 windows), the proposed method achieves 93.56% Accuracy, 86.96% Sensitivity, 96.15% Specificity, 89.86% Precision, 88.39% F1–Score, and an AUC–ROC of 0.9640, reducing missed arrhythmias by 74.5% compared to the standalone LSTM–AE baseline while maintaining a clinically acceptable False Positive Rate of 3.85%. The core detectors are trained without labeled data; only threshold calibration and recovery configuration use a small validation subset offline. Computationally, the framework operates at 7.93 ms per window with a throughput of 3,697 windows/s and a compact 350–400 KB footprint, making it well–suited for real–time, resource–constrained IoMT deployment.