A Multimodal Deep Learning Framework for Neuroimaging Analysis, Neurological Prediction and Intelligent Clinical Decision Support

Neuroimaging and clinical assessments are complementary measures of Alzheimer's disease and neurological classification, and prediction of cognitive decline remain challenging. This work introduces a multimodal deep learning framework based on attention mechanism, combining three-dimensional convolutional neural networks (CNN) for magnetic resonance imaging (MRI) and positron emission tomography (PET), and intermediate feature fusion, where the clinical multilayer perceptron (MLP) is trained using the features extracted from the multimodal networks. OASIS-3 was used to develop and internally evaluate the models; ADNI was used for independent external evaluation. In an illustrative, non-executed experimental scenario, the proposed model reached 0.850 internal macro-F1 and ROC-AUC, while the external ones were 0.800 and 0.900, respectively. The mean absolute error (MAE) of the hypothetical performance was 1.03, the root mean squared error (RMSE) was 1.42 and the R² was 0.73 for the prediction of 24-month cognitive decline using the Clinical Dementia Rating-Sum of Boxes (CDR-SB). The framework will include explainability and probabilistic calibration, as well as uncertainty aware risk stratification, that will link predictive outputs to clinician review. In the proposed contribution, the aim is to provide a joint assessment of discrimination, progression and decision support. These illustrative comparisons raise the possibility of superiority over unimodal approaches, but cannot be substantiated by reproducible testing, subgroup analysis or clinical utility testing.