A lightweight reliability- and confidence-aware selective ECG framework for degraded remote monitoring

Remote electrocardiogram (ECG) monitoring in Internet of Medical Things (IoMT) systems is vulnerable to sensor noise, missing leads, packet loss, and temporal dropout. The proposed Lightweight Reliability-Aware Internet of Medical Things Electrocardiogram Network (LR-IoMT-ECGNet) is a lightweight reliability-aware framework that couples degradation-aware representation learning with selective decision support. Physikalisch-Technische Bundesanstalt eXtended Large (PTB-XL) electrocardiography dataset version 1.0.3 was used with official folds for training (n = 17,418), validation (n = 2183), and independent testing (n = 2198). Robustness was evaluated under clean, mild, moderate, and severe simulated degradation. A hybrid gate combining predictive confidence and signal reliability was tuned exclusively on validation data and locked before test evaluation. The 24,429-parameter model achieved an area under the receiver operating characteristic curve (AUC) of 0.9188 and F1-score of 0.8449 under clean conditions and an AUC of 0.8683, sensitivity of 0.7753, specificity of 0.8004, and F1-score of 0.8089 under severe degradation. By contrast, clean-trained CNN and ResNet-1D baselines showed near-zero severe specificity (0.0252 and 0.0143). On mixed-quality test data, the hybrid gate retained 90.2% of records and achieved accuracy of 0.8412, sensitivity of 0.8026, specificity of 0.8973, and F1-score of 0.8569 among accepted cases. Confidence-only referral was marginally stronger on some metrics, indicating that predictive confidence is the dominant selective signal while input reliability provides complementary information. The results support a compact, degradation-aware selective ECG framework for edge- or gateway-assisted monitoring, while external wearable validation remains necessary.