Hybrid Deep Learning with Late Fusion for Edge-Intelligent IoMT Hemodialysis Monitoring

Acute intra-dialytic arrhythmias and hypoxemia are common among hemodialysis patients and require continuous intelligent monitoring. Fully cloud-based platforms raise privacy and connectivity concerns, whereas single-modality edge solutions cannot capture complementary cardiac and respiratory information. This paper introduces a hybrid 1D-CNN and BiLSTM model with learnable late fusion combining electrocardiogram and oxygen saturation (SpO2). Following the External Edge Intelligence paradigm, the bedside ESP32 node performs only acquisition, filtering, and secure transmission, whereas inference of the INT8-quantized model executes on an in-ward host-cloud server via ONNX Runtime. Evaluated under a strict patient-level protocol with fully disjoint training, validation, and test partitions, the FP32 prototype achieves 96.8% accuracy, 0.968 weighted F1, and 0.991 area under the curve (AUC); the deployed INT8 model retains 96.5% accuracy (−0.3 percentage points) with a model-only latency of 13 ms per 4-s window. Clinical validation on 18 patients across 96 sessions (372 hours) at Al-Amal Dialysis Center, Kirkuk, Iraq, yielded 95.8% sensitivity, 96.5% specificity, and Cohen’s kappa of 0.92, enabling real-time, privacy-preserving Internet of Medical Things monitoring.