Macrophage activation syndrome (MAS) secondary to Still’s disease is a potentially fatal complication, associated with mortality rates exceeding 10%. Early identification is critical for survival but remains challenging due to the lack of specific predictive biomarkers. To develop and test an explainable model to predict MAS in adult patients with Still’s disease using routine baseline clinical parameters, and to implement it as an interactive tool. We conducted a retrospective model development and testing study across four hospital sites from Aug 1, 2019 to Jul 31, 2025. Adults meeting the Yamaguchi criteria for Still’s disease were included. Demographics, imaging/physical findings, and routine laboratory tests within 48 hours of admission were analyzed. Predictors were selected using nested cross-validated LASSO, and five algorithms (logistic regression, random forest, SVM, XGBoost, and LightGBM) were compared. Model interpretability was assessed with SHAP, and a bedside score was derived using Firth’s penalized logistic regression. A total of 312 patients with Still’s disease was included, with model development in two centers (n=226) and testing in two independent centers (n=86). The final XGBoost model retained five key predictors: ferritin, splenomegaly, platelet count, total cholesterol, and erythrocyte sedimentation rate, achieving an AUC of 0.839 in the test set, with a sensitivity of 0.824, specificity of 0.710, acceptable calibration (Brier 0.136), and favorable net clinical benefit. The derived 0–10 bedside risk score stratified the training cohort into low- (1%), intermediate- (14.6%), and high-risk (75%) MAS groups. We present an interpretable machine learning model based on baseline data and simplified risk score for predicting in-hospital MAS in adult patients with Still’s disease. To our knowledge, this study represents one of the larger adult cohorts assembled for Still’s disease-associated MAS.
Zheng et al. (Sun,) studied this question.