A small-data machine learning framework with interpretable scorecards for disease activity assessment: A systemic...

BackgroundAssessment of Systemic Lupus Erythematosus (SLE) activity is complicated by clinical heterogeneity and data scarcity. Additionally, the "black-box" nature of many machine learning models creates significant barriers to their adoption in reliable clinical decision-making.ObjectiveThis study aims to develop an interpretable machine learning framework that transforms high-dimensional clinical data into a transparent scorecard for disease activity assessment, specifically addressing small-sample constraints.MethodsIn a cohort of 104 female SLE patients, LASSO regression selected 14 robust predictors from 149 variables spanning immunological and hematological domains. Eleven machine learning algorithms were systematically evaluated using stratified five-fold cross-validation. The optimal base model was transformed into a disease activity scorecard using a dynamic binning strategy to capture nonlinear effects.ResultsWhile Random Forest achieved the highest test-set AUC (0.883) among standalone algorithms, Logistic Regression was selected for its superior stability (AUC: 0.867) and interpretability. The derived scorecard demonstrated strong discrimination on the hold-out internal test set (AUC: 0.917), with the highest observed AUC among the evaluated models. At an optimized threshold, it achieved balanced sensitivity (0.917) and high specificity (0.840), with key predictors aligning with established pathophysiology.ConclusionsReliable prediction in data-limited settings can be achieved without sacrificing interpretability. The proposed framework provides an interpretable proof-of-concept decision-support tool for disease activity assessment in data-limited clinical settings. Although the proposed scorecard demonstrated promising performance on internal validation, prospective multi-center and external validation are required before routine clinical implementation.