Data-driven explainable machine learning for personalized risk classification of myasthenic crisis.

Area of Science:

• Neurology
• Artificial Intelligence
• Clinical Informatics

Background:

• Myasthenic crisis (MC) is a severe complication of Myasthenia gravis (MG) requiring intensive care.
• Accurate risk stratification for MC is crucial for timely treatment adjustments and preventing disease exacerbation.

Purpose of the Study:

• To evaluate the feasibility of using explainable machine learning (ML) models to classify MG patients into low or high risk for MC.
• To identify distinct clinical features associated with MC risk using routine medical data.

Main Methods:

• A single-center, pseudo-prospective cohort study involving 51 MG patients (13 with MC).
• Machine learning models (Lasso regression, random forest) were trained on real-world clinical data from a hospital management system.
• Patients were classified as high or low risk for MC based on model predictions.

Main Results:

• Random forest models demonstrated a predictive accuracy of 76.5% for MC risk, outperforming Lasso regression (68.8% AUC).
• Explainable feature importance analysis identified multimorbidity as a significant factor in distinguishing high-risk from low-risk patients.
• Model performance was evaluated using cross-validated AUC across 5100 runs.

Conclusions:

• This study provides proof-of-concept for ML-based MC risk classification using routine clinical data and explainable AI.
• Future research should focus on multi-center data, larger patient cohorts, and advanced ML models incorporating free-text data for improved prediction.