An Explainable Multimodal AI Software Framework for 36-Month MCI-to-Alzheimer’s Disease Progression Prediction: A Methodological Evaluation

Abstract

The complexity involved in predicting the progression of Alzheimer’s Disease (AD) occurs because of differences in neuroimaging data and the many inconsistencies in biomarker data and cognitive data alike. The current research examined an AI model that uses deep learning, organized data preprocessing, and post-experiment interpretations. A tracing examination of the provided computing material established that the experiments were not conducted using actual ADNI or OASIS-3 data but rather fabricated ADNI/OASIS-type data; therefore, the partition of OASIS was considered an independent synthetic test in the context of external data testing and not clinical validation. The complete dataset used in the research involved 603 records of MCI participants and used only good-quality MRI recordings. In the research, MRI data were combined with demographic, cognitive, APOE4-related, PET, and biomarker data. The output of the experiment was based on post-experiment explanation tools identified as SHAP and Grad-CAM. The internal AUROC of the gated model reached 0.676; sensitivity and specificity were equal to 0.524 and 0.690; the F1 score was equal to 0.489. The external domain testing results were AUROC 0.675; AUPRC 0.571; sensitivity 0.593; specificity 0.656; F1 score 0.551. The logistic regression model slightly surpassed the gated model in terms of internal AUROC; the simple multimodal concatenated model slightly surpassed the gated model in terms of external AUROC and AUPRC; therefore, the predictive edge cannot be claimed. The preprocessing benchmark yielded a throughput of approximately 708–738 records/s and a batch-8 inference latency of approximately 1.25 ms/record. These results support technical feasibility within the tested workload but do not establish large-scale big-data scalability or clinical applicability.