An uncertainty-aware multimodal deep learning framework for Alzheimer’s disease classification using MRI and electronic health records

Despite significant advances in artificial intelligence, achieving early and accurate diagnosis of Alzheimer’s disease (AD) remains a challenging task due to the complexity, heterogeneity, noise, and uncertainty inherent in neurological and clinical data. Conventional machine learning and deep learning models generally rely on deterministic decision boundaries, which limits their ability to manage diagnostic uncertainty and ambiguous cases, particularly during the early Mild Cognitive Impairment (MCI) stage. This study presents a reinforcement learning–optimized fuzzy deep learning framework for uncertainty-aware prediction of AD. The proposed framework integrates convolutional neural networks (CNN) and bidirectional long short-term memory (BiLSTM) networks to learn comprehensive multimodal representations by extracting spatial features from MRI images and temporal patterns from electronic health record (EHR) data. The learned features are further processed through a fuzzy inference system, enabling soft classification and explicit modelling of uncertainty in diagnostic decisions. Additionally, a reinforcement learning (RL)-based optimization mechanism is introduced to adaptively optimize classification parameters and fuzzy membership functions, thereby enhancing model robustness, generalization capability, and predictive performance. Experimental results demonstrate that the proposed framework achieves superior performance compared with conventional machine learning and deep learning approaches, attaining an accuracy of 98.9%, precision of 98.7%, recall of 98.8%, F1-score of 98.8%, and an AUC of 0.99. In contrast, the baseline CNN–BiLSTM model achieved 96.4% accuracy and an AUC of 0.97. Statistical evaluation using the DeLong test confirmed that the performance improvement was significant ( p < 0.05). These findings indicate that the proposed uncertainty-aware framework provides a reliable and effective computational approach for early AD diagnosis and clinical decision support.