Development and Validation of Explainable Artificial Intelligence Models for Early Prediction of Neurodegenerative Disorders in Middle Eastern Populations: A Multimodal Approach Integrating Clinical, Genetic and Neuroimaging Data

Background: Early detection of neurodegenerative disorders remains challenging, particularly in Middle Eastern populations where diagnostic delays and limited biomarker availability impact treatment outcomes. This study aims to build and validate explainable artificial intelligence (AI) models for early prediction of Alzheimer’s disease, Parkinson’s disease (PD) and multiple sclerosis using multimodal data from Saudi Arabian and regional populations. Methods: We constructed deep learning models using data from 12,458 participants (4872 with neurodegenerative disorders and 7586 controls) across 8 centers in Saudi Arabia (KSA) (2018–2024). Models integrated clinical variables, genetic markers, neuroimaging features and digital biomarkers. Model performance was evaluated using AUROC, sensitivity, specificity and prospective validation. Explainability was assessed using SHapley Additive exPlanations values and attention maps. Results: The multimodal ensemble model achieved superior performance (AUROC: 0.92, 95% confidence interval: 0.90–0.94) compared to single-modality models for predicting neurodegenerative disorders 3–5 years before clinical diagnosis. Performance remained robust across different demographic subgroups (age, gender and education levels). Key predictive features included hippocampal volume, specific genetic variants (APOE, LRRK2 and MAPT), subtle gait parameters and digital voice biomarkers. Explainability analyses revealed distinct predictive patterns for each disorder, with neuroimaging features dominating Alzheimer’s predictions and digital biomarkers showing the highest importance for PD. Conclusion: Our explainable AI approach demonstrates high accuracy for early prediction of neurodegenerative disorders in Middle Eastern populations, with potential to reduce diagnostic delays by 2–4 years. The integration of multimodal data and emphasis on model explainability addresses critical barriers to clinical implementation. Our findings highlight the importance of population-specific screening approaches and targeted early interventions for neurodegenerative disorders in Saudi Arabia and similar healthcare settings.