Separating biomarker association from incremental EHR value: a reproducible LAR and troponin benchmark in older high-acuity patients
Background Biomarkers may show strong outcome associations without adding information to detailed electronic health record (EHR) models. We developed a reproducible benchmark separating association from incremental value for the lactate-to-albumin ratio (LAR) and troponin in older high-acuity patients, with the aim of locating where marker-specific information remains distinct after an information-rich baseline. Methods From 12,000 records in the official CinC 2012 Sets A, B, and C, we included 1,640 patients aged at least 65 years with computable LAR. First-48-h measurements from 37 raw variables produced a transparent 290-feature baseline. Baseline, LAR-augmented, missingness-aware troponin, integrated, and score-enhanced models used L2-regularized logistic regression, random forest, and gradient boosting. Set A supported development, Set B validation and recalibration, and Set C locked internal testing. Evaluation combined discrimination, calibration, paired bootstrap ablation, optimism correction, decision curves, permutation importance, and risk concentration. Results Mortality rose from 18.2% in the lowest LAR max quartile to 36.7% in the highest, with 2.60-fold higher odds of death. The association pointed in the same direction across four ICU categories (odds ratios per stratum-specific LAR interquartile range, 1.30–1.76). The score-enhanced random forest achieved the highest Set C AUROC (0.769; 95% CI: 0.729–0.809). Recalibration improved its Brier score from 0.200 to 0.170 and moved calibration-in-the-large from − 0.652 to 0.078, thereby reducing average overprediction. Paired-bootstrap intervals for LAR or troponin augmentation crossed zero, locating an information-saturation boundary after the 290-feature EHR baseline; decision-curve mean net benefits for baseline and LAR-augmented gradient boosting were 0.120 and 0.122. Troponin-tested patients had higher mortality than untested patients (31.6% vs. 23.6%). Conclusion The benchmark locates where marker value resides: LAR retained a cross-ICU mortality signal while the 290-feature EHR baseline captured most overlapping information; missingness-aware troponin modeling quantified assay values and selective testing. The reusable design distinguishes association, incremental information, probability reliability, and threshold utility. CinC 2012 provides a fixed reproducibility anchor, with contemporary assays and care pathways as the next validation setting.
Authors
- Mengdi Zhang (ORCID: https://orcid.org/0009-0007-8207-7093)
- Hong Zhang (ORCID: https://orcid.org/0000-0002-1770-6582)
- Zheng Ni
Institutions
- Capital Medical University (CN)
- Beijing Shijingshan Hospital (CN)
Publication Details
- Journal
- Frontiers in Medicine
- Published
- 2026-11-17
- DOI
- https://doi.org/10.3389/fmed.2026.1909486
- Primary Topic
- Machine Learning in Healthcare
- Type
- article
- Field-Weighted Citation Impact
- 0.00