Interpretable laboratory-data model for risk stratification of elevated NT-proBNP and its deployment in diagnostic...

Area of Science:

• Biomedical informatics
• Clinical diagnostics
• Machine learning in healthcare

Background:

• Heart failure (HF) presents a significant global health challenge.
• N-terminal pro-B-type natriuretic peptide (NT-proBNP) is crucial for HF diagnosis, but its high cost and limited analyzer availability hinder widespread use.
• Routine laboratory tests offer a potential low-cost alternative for HF risk stratification.

Purpose of the Study:

• To develop and validate an interpretable decision tree model for stratifying the risk of elevated NT-proBNP levels (>300 pg/mL).
• To assess the model's performance and clinical utility when deployed in a diagnostic support system (DSS).

Main Methods:

• An interpretable decision tree was developed using 20 candidate predictors from 19,889 patient encounters.
• Hyperparameters were optimized via 10-fold cross-validation, with a focus on high sensitivity (≥0.90) for triage.
• Performance was evaluated using AUROC, sensitivity, specificity, PPV, and NPV on internal and external validation cohorts.

Main Results:

• The decision tree identified key predictors including serum albumin, eGFR, and age.
• Internal validation showed an AUROC of 0.804 and sensitivity of 0.879.
• External validation in the DSS demonstrated robust performance with an AUROC of 0.806 and sensitivity of 0.882.

Conclusions:

• An interpretable decision tree derived from routine laboratory data effectively identifies elevated NT-proBNP levels with high sensitivity.
• The model demonstrates robust performance post-deployment, offering a scalable, low-cost solution for risk-directed triage in heart failure.
• This approach can improve resource allocation and diagnostic efficiency in managing heart failure.