Journal of Rare Cardiovascular Diseases

This paper presents a comprehensive framework for real-time cardiac arrest detection and heart-disease prediction that integrates Internet of Things (IoT) sensing, advanced signal preprocessing, and state-of-the-art deep learning models based on deep convolutional neural networks (DCNNs). The proposed system combines continuous ambulatory acquisition of physiological waveforms (single-lead and multi-lead ECG, photoplethysmography (PPG), respiration and accelerometry) through wearable IoT nodes with a hierarchical data-management pipeline for edge preprocessing, secure transmission, and cloud-based inference. Signal preprocessing applies artifact removal, beat segmentation, and time–frequency feature extraction; these engineered representations are fed to a hybrid DCNN–temporal network that fuses convolutional feature encodings with sequential modelling to capture both morphological and temporal dynamics relevant to acute cardiac events. The model performs two linked tasks: (1) early detection of cardiac arrest and life-threatening arrhythmias with low latency for automated alerting and dispatcher integration; and (2) longitudinal risk stratification for heart disease prediction using multimodal time-series and clinical metadata. Evaluation is performed on publicly available ECG/PPG benchmarks and on a clinical in-hospital dataset, using sensitivity, specificity, F1-score, area under the receiver operating characteristic curve (AUC-ROC), and time-to-detection as primary metrics. The system also addresses deployment considerations: energy-efficient edge inference, privacy-preserving transmission, and clinician-centred explainability (saliency maps and attention-based explanations). Results demonstrate that the integrated IoT+DCNN approach attains high sensitivity for early arrest detection while providing robust predictive performance for longer-term cardiovascular risk. The contributions of this work are: (i) a reproducible system architecture combining wearable IoT acquisition with a hybrid DCNN temporal model for dual tasking (acute detection + chronic prediction); (ii) rigorous evaluation on heterogeneous datasets showing clinically relevant performance gains; and (iii) practical design guidelines for real-world deployment, including latency, energy, and explainability constraints.