Abstract
BACKGROUND: Heart disease remains a leading cause of global morbidity and mortality, emphasizing the need for early and accurate prediction. Traditional models, however, often struggle with high class imbalance, limited accuracy, inadequate hyperparameter tuning, reliance on sparsely labeled data, and poor interpretability. METHODS: To address these limitations, we introduce a novel 3-tier information fusioned framework integrating Oversampling Using the Propensity Score (OUPS), ConvRecurrentNet (CRNet), an optimized CRNet, and CRNet enhanced with Uncertainty-Based Sampling (UBS). Initially, the highly imbalanced heart disease health indicator dataset is balanced using OUPS while preserving the underlying data distribution. In the first tier, we develop a new deep model; CRNet which combines gated units to capture temporal dependencies and convolutional layers for efficient spatial feature extraction. The second tier leverages the grasshopper optimization algorithm to optimize CRNet configuration and classification performance through hyperparameter tuning. At the third tier, we use an active learning based UBS to address the problem of sparse labeled data selecting the most informative samples for effective CRNet model learning. Model performance and generalizability are assessed through 10-fold cross validation and one-way analysis of variance. Interpretability is ensured using explainable artificial intelligence (AI) methods, local interpretable model-agnostic explanations and shapley additive explanations. RESULTS: Improvements are observed across all performance metrics: 3.45%, 5.75%, and 8.05% in accuracy, 3.45%, 6.90%, and 12.64% in precision, 4.65%, 5.81%, and 4.65% in recall, 4.65%, 6.98%, and 15.12% in F1-score, 1.04%, 2.08%, and 3.13% in receiver operating characteristic-area under the curve, 8.00%, 9.33%, and 17.33% in Matthews correlation coefficient, 12.16%, 13.51%, and 18.92% in Cohen’s Kappa, and log loss reduction of 23.64%, 29.45%, and 45.82% across CRNet, optimized CRNet, and CRNet with UBS, respectively. CONCLUSION: The proposed 3-tier fusion framework enhances heart disease prediction by balancing data, optimizing parameters, and applying active learning to address labeled data scarcity. Explainable AI further improves transparency and supports real-world clinical adoption.