Real-time embedded detection of bradycardia, tachycardia, and arrhythmia using photoplethysmography and on-device convolutional neural networks

Cardiovascular diseases (CVDs) remain the leading cause of mortality worldwide, highlighting the need for accurate, accessible, and continuous cardiac monitoring solutions. This study presents the design and implementation of a photoplethysmography (PPG)-based, real-time heart health monitoring system that integrates specialized hardware, embedded computing, and artificial intelligence for fully autonomous operation. Unlike traditional approaches requiring external processing or cloud connectivity, the system uses a custom PPG acquisition module, an STM32F407 microcontroller, and a compact real-time convolutional neural network (RT-CNN) optimized for on-device execution. The hardware ensures robust signal acquisition under diverse conditions, while the RT-CNN processes one-dimensional PPG signals to detect normal rhythms and multiple cardiac anomalies with high accuracy. Experimental evaluation demonstrated 98.57% accuracy, a 500 ms reaction time, and consistently high recall, specificity, and F1-scores, outperforming comparable models. Open-source, modular architecture makes the platform scalable for telemedicine, home-based care, and resource-limited settings, while Grad-CAM visualizations enhance clinician trust in AI-assisted decisions. The system was evaluated on data collected from 640 subjects (320 normal and 320 patients with diagnosed cardiovascular conditions). This work offers a cost-effective, portable, and clinically relevant approach to real-time cardiac anomaly detection, addressing key limitations in existing PPG-based monitoring systems.