RL-optimized CNN–BiLSTM: A multi-layer adaptive deep learning framework for heart disease diagnosis using IoT sensor data and reinforcement learning

Cardiovascular disease remains a leading cause of premature mortality worldwide, highlighting the need for accurate, continuous, and adaptive cardiac monitoring. This study proposes a three-layer IoT–Fog–Cloud framework integrated with a reinforcement learning (RL)-optimized CNN–BiLSTM model for early heart disease detection. The IoT layer continuously acquires physiological signals, including ECG, heart rate, and blood pressure; the fog layer performs preprocessing, feature extraction, dimensionality reduction, and classification; and the cloud layer supports long-term storage and large-scale analytics. The key novelty of this study is the joint integration of multimodal physiological and clinical data with dual RL-based optimization within the IoT–Fog–Cloud architecture. Specifically, a Deep Q-Network (DQN) adaptively selects discriminative features, while a policy-gradient mechanism dynamically optimizes classification decision thresholds, enabling the CNN–BiLSTM model to adapt beyond conventional static feature-selection and classification approaches. The framework was evaluated using the Cleveland, Hungarian, and Framingham cardiovascular datasets. Data preprocessing included normalization, anomaly detection, feature extraction, and PCA-based dimensionality reduction. The proposed RL-optimized CNN–BiLSTM achieved accuracies of 96.4%, 95.7%, and 94.1%, with AUC values of 0.98, 0.97, and 0.96, on the Cleveland, Hungarian, and Framingham datasets, respectively. Robustness testing under Gaussian noise further demonstrated stable predictive performance across heterogeneous datasets. These findings demonstrate the potential of the proposed adaptive framework for real-time cardiac risk assessment, remote patient monitoring, and intelligent clinical decision support.