An interpretable decision tree using routine laboratory data detected elevated NT-proBNP (>300 pg/mL) with an AUROC of 0.806 (95% CI 0.799-0.813) and sensitivity of 0.882 in an external cohort.
Observational (n=34,792)
No
Does an interpretable decision tree model using routine laboratory data accurately predict elevated NT-proBNP >300 pg/mL in clinical encounters?
A low-cost, interpretable decision tree model based on routine laboratory data can effectively triage patients for elevated NT-proBNP, potentially improving resource allocation in heart failure diagnosis.
Effect estimate: AUROC 0.806 (95% CI 0.799-0.813)
Background: Heart failure (HF) is a growing global burden. Although N-terminal pro–B-type natriuretic peptide (NT-proBNP) guides diagnosis, assay cost and analyzer availability limit routine use. Routine laboratory data may offer a low-cost triage alternative. Methods: We developed and validated an interpretable decision tree to stratify the risk of elevated NT-proBNP >300 pg/mL and assessed deployment in a diagnostic support system (DSS). We analyzed 19,889 encounters at Gifu University Hospital (Aug 2022–May 2024). All 20 candidate predictors were included without prior feature selection to capture non-linear associations. Hyperparameters were tuned by 10-fold cross-validation. Final classification used a fixed decision rule optimized for high sensitivity (≥0.90 in training) to support effective triage. Performance comprised AUROC (DeLong 95% CIs) and sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy (Wilson 95% CIs) on an internal hold-out set (n=3,978) and a temporal external cohort (n=14,903; Jun 2024–Jun 2025). Analyses were complete-case with no imputation. Results: The decision tree inherently utilized clinically relevant predictors including serum albumin, eGFR, and age. Internal test performance: AUROC 0.804 (0.791–0.818); sensitivity 0.879 (0.863–0.893); specificity 0.505 (0.484–0.526); accuracy 0.674 (0.660–0.689). External performance within the DSS: AUROC 0.806 (0.799–0.813); sensitivity 0.882 (0.874–0.890); specificity 0.518 (0.508–0.529); NPV 0.852 (0.842–0.861). Calibration and decision-curve analysis supported clinical utility. Conclusions: An interpretable tree built from routine laboratories detects clinically relevant NT-proBNP elevation with high sensitivity and performs robustly after deployment. This scalable, low-cost approach could enable risk-directed triage and more efficient resource allocation.
Hidekazu et al. (2026) conducted an observational in Heart failure (elevated NT-proBNP) (n=34,792). Interpretable decision tree model was evaluated on Elevated NT-proBNP >300 pg/mL (AUROC 0.806, 95% CI 0.799-0.813). An interpretable decision tree using routine laboratory data detected elevated NT-proBNP (>300 pg/mL) with an AUROC of 0.806 (95% CI 0.799-0.813) and sensitivity of 0.882 in an external cohort.