Answers are grounded in this paper and cite their sources.
Retrospective observational study finds routine laboratory tests outperform specialized immunology assays in predicting immune diagnoses among hospitalized adults, indicating diagnostic test overuse.
Overuse of immunology laboratory testing remains a challenge in acute care, increasing workload, downstream investigations, and diagnostic complexity. We applied explainable machine learning to evaluate real-world immunology test utilization among hospitalized patients who underwent immunology testing. We analyzed electronic health records from a tertiary-care hospital between 2018 and 2023, including 7,402 adults and 75 model predictors. Discharge diagnoses were clinically reviewed and mapped to a retrospective immune-related versus non-immune outcome. Multiple machine-learning models were evaluated, with XGBoost selected for detailed interpretation using SHapley Additive exPlanations (SHAP). Immunology assays were frequently ordered in both groups, with only modest differences in ordering frequency. XGBoost achieved an ROC-AUC of 0.749 and weighted F1 of 0.75. SHAP analyses showed that routine clinical and laboratory variables generally contributed more strongly to discrimination than most immunology assays, while complement-related markers showed higher contribution and assay effects varied across clinical contexts. Subgroup analyses demonstrated context-dependent variability, and outcome-independent sensitivity analyses showed largely stable performance when information from individual assays was randomly masked. These findings show how explainable machine learning can support evaluation of diagnostic test-utilization patterns and provide a practical framework for identifying potential misalignment between ordering frequency and model-derived predictive contribution within immunology-tested populations.
No takes yet. Share an insight, caveat, or question.
Shabtay et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: