Classifying the risk for myasthenic crisis using data-driven explainable machine learning with informative feature design and variance control – a pilot study

Abstract

AbstractImportanceMyasthenic crisis (MC) is a critical progression of Myasthenia gravis (MG), requiring intensive care treatment and invasive therapies. Classifying patients at high-risk for MC facilitates treatment decisions and helps prevent disease progression.ObjectiveTo test whether machine learning models trained with real-world routine clinical data can aid precisely identifying MG patients at risk for MC.DesignThis is a pseudo-prospective cohort study of MG patients presenting since January 2010.SettingSingle center.ParticipantsA cohort of 51 MG patients was used for model training based on a defined set of real-world clinical data. The cohort was created from a convenience sample of 13 MC patients matched based on sex, five-year age band, antibody status, thymus pathology with MG patients who had not suffered an MC. Data analyses and model refinements were performed from June 2022 to May 2023.ExposureClassification of MG patients to high or low risk for MC using Lasso regression or random forest machine learning models.Main Outcomes and MeasuresThe accuracy of the risk classification was assessed by patient.ResultsThis study included 51 MG patients (13 MC, 38 non-MC; median age MC group 70.5, non-MC group 65.5). The mean cross-validated AUC classifying MG patients as high or low risk for MC based on simple or compound features derived from real-world routine clinical data showed a predictive accuracy of 68.8% for the regularized Lasso regression and of 76.5% for the random forest model. Feature importance scores suggest that multimorbidity may play a role in risk classification. Different thresholds were applied to tune model performance to optimal parameters. Studying result stability across 100 runs further indicated that the random forest model was better suited to cope with feature variance. Studying feature importance across 5100 model runs identified explainable features to distinguish MG patients at high or low risk for MC.Conclusions and RelevanceIn this study, feasibility of classifying risk for MC based on real-world routine clinical data using machine learning was shown. The models showed accurate and consistent performance indicating the utility of personalized risk assessment in MG patients using machine learning models.Key PointsQuestionCan machine learning models be used to classify Myasthenia gravis patients into groups at high or low risk for myasthenic crisis with high precision based on explainable data-driven features derived from real-world clinical data?FindingsIn this pseudo-prospective study of 51 Myasthenia gravis patients, the risk of myasthenic crisis using real-world clinical data was accurately classified employing two machine learning models with explainable features.MeaningThese findings suggest that it is possible to classify the risk for myasthenic crisis in patients based on real-world clinical data with high precision.