This work offers a fresh hybrid method for heart disease prediction using unsupervised, semi-supervised, and supervised learning methods to improve accuracy, especially in situations with minimal labelled data. To efficiently use both labelled and unlabelled data, the method integrates data preparation, feature engineering, K-means clustering, label propagation, and ensemble learning. The semi-supervised learning phase uses label propagation to infer labels for unlabelled data points; the unsupervised learning phase uses K-means clustering to identify patient groupings. Modern ensemble approaches include Random Forest and stacking techniques, which help capture complicated interactions between features, and are included in the component of supervised learning. Extensive studies employing real-world clinical datasets have shown that the suggested hybrid approach regularly outperforms the conventional techniques. The paper also covers ethical issues in the creation and implementation of AI-based medical prediction models and investigates model interpretability. Better clinical decision-making resulting from the increased accuracy, resilience, and interpretability of this method should improve patient outcomes and more effective allocation of healthcare resources. Moreover, the flexibility of the framework to different medical diagnostic activities creates fresh research opportunities in several spheres of healthcare, thus supporting the disciplines of medical artificial intelligence and machine learning.