Development and internal validation of an interpretable machine learning model using routine laboratory data for systemic sclerosis-associated pulmonary arterial hypertension.

This study aimed to develop and validate an interpretable machine learning (ML) model using routine laboratory data to support clinical decision-making and deliver personalized care-with a specific focus on optimizing the utilization of diagnostic resources for patients with systemic sclerosis-associated pulmonary arterial hypertension (SSc-PAH). Clinical and demographic data were retrieved from patients with SSc at a tertiary hospital in China. The diagnosis of SSc-PAH was confirmed by right heart catheterization in all relevant patients. The cohort was split into a training set (n = 367) and an independent test set (n = 158). A total of 39 routinely accessible laboratory test indicators and demographic variables were extracted from electronic health records (EHRs). Seven ML algorithms were developed, and their predictive performance was compared. The Least Absolute Shrinkage and Selection Operator (LASSO) technique was employed to screen for the most informative feature subset, thereby enhancing model performance and interpretability. The discriminative ability of the models was evaluated using the area under the receiver operating characteristic curve (AUC) and the area under the precision-recall curve (AP). Meanwhile, SHapley Additive exPlanations (SHAP) values were used to analyze feature importance and interpret individual prediction results. Potential as a preliminary screening tool was assessed via Decision Curve Analysis (DCA) were applied to further explore the potential associations between features and outcomes. Among the seven evaluated ML algorithms, the Adaptive Boosting (AdaBoost) classifier exhibited the optimal predictive performance. Following the screening of the top five most important features, an interpretable AdaBoost model was constructed. This final model demonstrated excellent discriminative ability in both the training set (AUC = 0.892, 95% confidence interval [CI]: 0.824-0.942) and the test set (AUC = 0.968, 95% CI: 0.930-0.996). DCA results indicated that, compared with the "treat all" or "treat none" strategies, this model delivered a higher net clinical benefit across the entire range of predictive probability thresholds. Further analysis using SHAP values revealed that N-terminal pro-brain natriuretic peptide (NT-proBNP), anti-triiodothyronine (rT