An Explainable Machine Learning Model for Early Prediction of Incident Myocardial Injury in Patients With Severe Fever With Thrombocytopenia Syndrome

ABSTRACT

This study aimed to develop an explainable machine learning–based model to enable early prediction of incident myocardial injury during hospitalization among patients with severe fever with thrombocytopenia syndrome (SFTS). This multicenter retrospective cohort study included and analyzed clinical data from 1,088 patients with SFTS who were hospitalized for the first time across four medical institutions between May 2011 and October 2024. The dataset was randomly split into an internal training set and a test set at a 7:3 ratio. After candidate predictors were selected from baseline clinical characteristics and laboratory indices, nine machine learning models were developed and compared, and external validation was conducted using retrospective cohorts from two independent medical centers. SHapley Additive Explanations (SHAP) were used to interpret model predictions and quantify feature importance. The ten key predictors identified comprised five baseline clinical characteristics and five laboratory indices. The CatBoost model demonstrated robust predictive performance in the internal test set and across two external validation cohorts, with AUCs of 0.750 (95% CI 0.698–0.803), 0.704 (95% CI 0.555–0.853), and 0.729 (95% CI 0.573–0.886), respectively. SHAP analysis further identified clinically relevant risk thresholds for the key laboratory predictors. In conclusion, this study developed a CatBoost‐based machine learning model that enables early prediction, at the time of hospital admission, of the risk of incident myocardial injury during hospitalization among patients with SFTS.