Deep Learning for Predicting Post-Stroke Cognitive Decline Using Multimodal Data: A Synthetic Proof-of-Concept Study

Cognitive impairment is a frequent and debilitating outcome of stroke, profoundly affecting patient independence, recovery trajectories, and long-term quality of life. Despite its prevalence, accurate early prediction of post-stroke cognitive decline (PSCD) remains an unmet challenge due to the multifactorial interplay between structural brain lesions, neurophysiological changes, and diverse clinical comorbidities. In this study, we present a multimodal deep learning framework designed to classify stroke patients at risk of PSCD by jointly analysing synthetic magnetic resonance imaging (MRI), electroencephalography (EEG), and clinical features. The architecture comprises three modality-specific branches: a 3D convolutional neural network (CNN) to extract lesion-based structural patterns from MRI volumes, a 1D CNN for temporal feature extraction from EEG signals, and a multilayer perceptron for tabular clinical predictors. These representations are integrated through late fusion to enable comprehensive multimodal risk stratification. Experiments conducted on a balanced synthetic dataset (N = 800; MRI: 32 × 32 × 32, EEG: 8 × 256, Clinical: 12 features) demonstrated strong predictive performance, achieving a best validation area under the curve (AUC) of 0.7327, which outperformed all unimodal baselines. To enhance clinical interpretability, we employed permutation feature importance, gradient-based saliency mapping, and UMAP visualization of latent embeddings. Results highlighted clinically meaningful drivers of prediction, including baseline Montreal Cognitive Assessment (MoCA) scores, age, and lesion burden. Contribution analysis further confirmed that multimodal fusion consistently yielded superior predictive capacity compared to individual modalities. Although based on synthetic data, these findings establish a robust methodological foundation for multimodal AI-driven risk stratification in stroke rehabilitation. Future work will extend validation to real-world multi-center clinical datasets, investigate time-to-event modelling, and explore integration into clinical decision-support systems to guide personalized interventions.