An AI-driven model for risk stratification and prototype intervention mapping in motoric cognitive risk syndrome and subjective cognitive decline

As Alzheimer's disease and other dementias continue their global rise, there is a compelling need for early identification of pre-clinical cognitive decline. Motoric Cognitive Risk Syndrome and Subjective Cognitive Decline represent crucial transition phases to facilitate targeted intervention, and existing unimodal approaches are neither sensitive nor sufficiently multimodal for appropriate risk classification and individualized intervention mapping (prototype). This study introduces a hybrid multimodal CNN fusion framework based on an improved version of light spectrum optimizer as metaheuristic optimizer to address this challenge. The framework fuses multiple modality-specific information through parallel three-dimensional convolutional neural network branches (MRI and PET) and encoders of motor, cognitive, and biomarker information followed by a combination of intermediate and late fusion layers with parameters tuned by the optimized framework. The framework is meticulously assessed on the BIOCARD longitudinally dataset and achieved an accuracy of 93.7% and an area under the ROC curve (AUC) of 0.965. It significantly surpassed seven state-of-the-art methods and revealed clinical interpretability through longitudinal risk trajectories and SHAP attribution of features.