Development and Validation of Multivariable Machine-Learning Models for the Prediction of Multisystemic Inflammatory Syndrome Outcomes in Latin American Children

Aim: We aimed to develop and test machine learning algorithms for the prediction of severe outcomes associated with MIS-C. Method: An observational ambispective cohort study was conducted including children aged from 1 month to 18 years old in 84 hospitals from the REKAMLATINA (Red de la Enfermedad de Kawasaki en America Latina) network diagnosed with MIS-C from 1st January 2020 to 31st June 2022. Multiple models were developed to predict four main outcomes: paediatric intensive care unit (PICU) admission, need for inotropes, need for mechanical ventilation, and death. Performance measures were accuracy for PICU admission, inotropes use and mechanical ventilation, and the area under the receiver operating characteristic curve (AUROC) for death. Variable contribution was analysed using Shapley Additive Explanations (SHAP) values. Results: We included 1303 children with a diagnosis of MIS-C. The model for the prediction of PICU admission (random forest [RF]) reached an accuracy of 0.80 (95% CI: 0.76–0.84), the model for inotrope use (RF) an accuracy of 0.86 (95% CI: 0.82–0.90), the model for mechanical ventilation (histogram-based gradient boosting [HBGB]) an accuracy of 0.84 (95% CI 0.80–0.88), and the model for death (RF) reached an AUROC of 0.85 (95% CI 0.77–0.93). Conclusions: We developed and validated machine learning models for the prediction of MIS-C related outcomes that can help clinicians risk stratify patients to identify those most likely to have a severe outcome from MIS-C.