Abstract
Introduction: Timely identification of the transition from relapsing-remitting multiple sclerosis (RRMS) to secondary- progressive MS (SPMS) is of great weight for effective treatment planning. However, this transition is typically diagnosed with an average delay of three years, leading to missed opportunities for early intervention. Building on our previous artificial intelligence (AI) work with Swedish electronic health record data, we extend our approach to the global MSBase cohort. Objectives/Aims: To develop AI models that predict disease progression from RRMS to SPMS proactively in a globally heterogeneous MS population while enabling user-defined confidence levels and interpretable predictions. Methods: We utilized two large-scale MS registries: MSBase (110,000 patients; 1.3 million visits across 45 countries) and the Swedish MS registry (22,000 patients; 200,000 visits). We trained random forest classifiers to predict disease states at each clinical encounter, integrating conformal prediction to quantify predictive uncertainty and explainable AI to enhance interpretability and transparency. Results: The global model achieved an F1 score of 0.83 and outperformed country-specific models in several regions. However, in certain countries, local models were better fitted. Calibration curves revealed marked differences in RRMS and SPMS diagnoses across countries. We identified groups with aligned predictions by clustering countries based on calibration similarity. While the global model generalized well, clustered models improved local accuracy. Conclusion: We developed AI models that provide accurate and interpretable predictions of MS progression trained on registry data from 18 countries. The global model offers scalability, while localized approaches better capture regional diagnostic practices. This framework supports harmonizing diagnostic standards and can enhance clinical trial design and international data interpretation.