Explainable Machine Learning Models for Alzheimer’s Diagnosis: Integrating Neuroimaging, Biomarkers, and Clinical Features for Improved Trustworthiness

Many types of dementia exist, but Alzheimer’s Disease (AD) is biggest contributor to dementia worldwide. As such, it is essential that we accurately identify AD as soon as possible as possible in order to properly treat and plan for care. At present, traditional methods of diagnosing patients for AD primarily include neuroimaging and clinical assessments, both of which often use subjective judgement to reach a diagnosis and are likely to be low in sensitivity during early AD. DL methods may be an accurate way to predict early onset AD however, due to their lack of interpretability, they have not received much clinical success. This work addresses this issue by providing an explainable multimodal DL framework for predicting the early onset of AD using MRIs, PET scans, and clinical biomarkers. The authors use CNN with an attention fusion mechanism to extract and combine complementary information from multiple modalities. SHAP and Grad CAM techniques were used to increase the interpretability of the multiple DL networks, which allow the transparency of feature and brain region identification. Experimental results using benchmark datasets demonstrate that the proposed framework produces comparable classification accuracy to existing methods and provides meaningful clinical insights. Furthermore, the paradigm that has been suggested strikes a compromise between generating predictive performance and interpretability; thus, the proposed framework supports trustworthy and evidence-based decisions, indicating its potential for reaching deployment in clinical settings.