Explainable time-to-progression predictions in multiple sclerosis.

Area of Science:

• Machine learning applications in neurology
• Survival analysis in clinical research
• Explainable artificial intelligence in healthcare

Background:

• Current multiple sclerosis (MS) prognostic models use binary predictions, failing to account for individual disease severity.
• There is a need for more nuanced prognostic tools in MS research.

Purpose of the Study:

• To model the time to disability progression in MS patients.
• To enhance model interpretability using explainable machine learning techniques.

Main Methods:

• Utilized a subset of 29,201 patients from the MSBase registry.
• Employed random survival forests for time-to-event prediction of disability progression (Expanded Disability Status Scale).
• Applied SHAP and Bellatrex for model explainability, providing global and local insights.

Main Results:

• Random survival forest achieved state-of-the-art performance, comparable to prior random forest models.
• The model accurately predicted progression over a 10-year horizon (AUROC >60%).
• Explainability techniques revealed clinically relevant insights, such as the impact of recent MS therapies on progression rates.

Conclusions:

• Transitioning from binary classification to time-to-event modeling in MS prognosis does not compromise performance.
• Explainable AI is crucial for understanding and validating prognostic models in MS care.
• This approach allows for more comprehensive and personalized patient prognosis.