Answers are grounded in this paper and cite their sources.
Background To develop and prospectively validate an interpretable machine-learning model integrating lung ultrasound and routinely available clinical data for early recognition of severe Mycoplasma pneumoniae pneumonia (SMPP) in children. Methods This single-center cohort study included children aged 1 month to 14 years hospitalized with confirmed Mycoplasma pneumoniae pneumonia (MPP) at Beijing Friendship Hospital, Capital Medical University, from January 2018 to June 2025. A retrospective cohort of 1,696 children was randomly stratified 7:3 into a training set (n = 1,187) and an internal validation set (n = 509); an independent prospective cohort of 233 children was used for validation. Clinical, laboratory, and lung ultrasound variables available within 24 hours of admission were considered. Integrated feature ranking and recursive feature elimination selected predictors. Ten individual machine-learning models and stacking ensembles were evaluated using discrimination, calibration, decision-curve analysis, and SHapley Additive exPlanations. Results The retrospective cohort comprised 1,256 children with MPP and 440 with SMPP; the prospective cohort comprised 117 and 116 children, respectively. Twelve predictors were selected, including shred sign, lung ultrasound score, fever duration, age, inflammatory indices, and routine biochemical variables. The RF+XGBoost stacking model achieved the best internal-validation performance (AUC, 0.934; 95% CI, 0.920–0.949; accuracy, 0.848; sensitivity, 0.802; specificity, 0.884; F1 score, 0.821). In prospective validation, its AUC was 0.841 (95% CI, 0.786–0.887), exceeding that of RF (AUC, 0.743; 95% CI, 0.678–0.802; McNemar P = 0.0031). At the prespecified threshold of 0.497, accuracy, sensitivity, specificity, and F1 score were 0.755, 0.776, 0.768, and 0.759, respectively. The prospective Brier score was 0.167; the significant Hosmer-Lemeshow test indicated calibration deviation requiring recalibration before broad deployment. Conclusions An interpretable 12-feature RF+XGBoost stacking model combining lung ultrasound with routine clinical data showed strong discrimination for early SMPP recognition in prospective validation. Further multicenter validation and prospective recalibration are required before broad clinical deployment. Trial registration: The study was retrospectively registered at ClinicalTrials.gov (NCT07333833).
No takes yet. Share an insight, caveat, or question.
Gao et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: