Early multibiomarker machine-learning prediction of in-hospital mortality in critically ill patients with heart failure and atrial fibrillation

Q: What are the key findings of this study?

Lactate-to-hemoglobin ratio was associated with in-hospital mortality, while locked XGBoost did not outperform raw-component logistic regression.

Q: What is the clinical evidence from this study?

Study design: Cohort. Population: Critically ill patients with heart failure and atrial fibrillation (n=4328). Intervention: Early lactate-to-hemoglobin ratio (LHR), BUN-to-hemoglobin ratio (BHR), and locked XGBoost machine-learning model vs. Clinical base models and logistic regression with raw lactate, BUN, and hemoglobin components. Primary outcome: All-cause post-landmark in-hospital mortality (OR 1.26, 95% CI 1.13-1.41, p=<0.001).

Q: What does this research mean for the field?

In critically ill patients with heart failure and atrial fibrillation, machine-learning models and composite biomarker ratios offer no meaningful discrimination advantage over standard logistic regression using clinical features and raw laboratory values for predicting in-hospital mortality. Novelty: ClaimNovelty.INCREMENTAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

Q: What question did this study set out to answer?

To evaluate whether early multibiomarker composite ratios and machine-learning algorithms improve in-hospital mortality prediction compared to simpler models in critically ill patients with heart failure and atrial fibrillation.

Cevahir Alioglu; Nuray Kahraman Ay; Hasan Burak İşleyen

doi:10.1038/s41598-026-79598-5

November 14, 2026Scientific ReportsOpen Access

Early multibiomarker machine-learning prediction of in-hospital mortality in critically ill patients with heart failure and atrial fibrillation

View Full Paper

Ask AI

Bookmark

Share

Key result

Higher lactate-to-hemoglobin ratio is linked to ~26% higher in-hospital mortality.

OR 1.26
95% CI 1.13-1.41
P<0.001
n=4,328

Why the study?

Early prognostic assessment in critically ill patients with HF and AF may benefit from multidomain biomarkers, but composite ratios and machine-learning models require direct benchmarking against simpler alternatives.

Does a locked machine-learning (XGBoost) model or composite biomarker ratios (lactate-to-hemoglobin and BUN-to-hemoglobin) improve prediction of in-hospital mortality compared to logistic regression with raw biomarker components in critically ill patients with heart failure and atrial fibrillation?

Key Points

To evaluate whether early multibiomarker composite ratios and machine-learning algorithms improve in-hospital mortality prediction compared to simpler models in critically ill patients with heart failure and atrial fibrillation.
Retrospective observational study utilizing MIMIC-IV v3.1 for model development and internal cross-validation, and eICU-CRD v2.0 for locked independent external validation.
Included adult ICU patients with heart failure, atrial fibrillation, and measurements of lactate, blood urea nitrogen (BUN), and hemoglobin within 6 hours of admission, applying a fixed 6-hour landmark window (N=3,925 in MIMIC; N=403 in eICU).
Compared locked XGBoost machine learning against logistic regression models incorporating clinical variables, biomarker ratios (lactate-to-hemoglobin ratio [LHR] and BUN-to-hemoglobin ratio [BHR]), or individual raw laboratory values for post-landmark in-hospital mortality.
In the development cohort, in-hospital mortality occurred in 760 of 3,925 patients (19.4%), compared to 125 of 403 patients (31.0%) in the external validation cohort.
The locked XGBoost model achieved an external AUC of 0.668 (95% CI 0.610–0.723), demonstrating an AUC 0.053 higher than clinical features plus LHR (95% CI 0.002–0.101).
XGBoost showed no discrimination advantage over a standard logistic regression using clinical features plus raw lactate, BUN, and hemoglobin (AUC 0.672; delta AUC −0.004, 95% CI −0.042 to 0.034), and removing biomarker ratios changed XGBoost AUC by only 0.0005.

Study Design

Type

Cohort (n=4,328)

Multicenter

Yes

Structured PICO

Does a locked machine-learning (XGBoost) model or composite biomarker ratios (lactate-to-hemoglobin and BUN-to-hemoglobin) improve prediction of in-hospital mortality compared to logistic regression with raw biomarker components in critically ill patients with heart failure and atrial fibrillation?

P

Population

4,328 critically ill adults (3,925 in MIMIC-IV development and 403 in eICU external validation; median age 75 years, ~40% female) with concomitant heart failure and atrial fibrillation and complete 6-hour lactate, BUN, and hemoglobin measurements, followed for post-6-hour in-hospital mortality.

E

Exposure

Early multibiomarker risk prediction using composite biomarker ratios—lactate-to-hemoglobin ratio (LHR) and blood urea nitrogen-to-hemoglobin ratio (BHR)—and a locked machine-learning model (XGBoost) incorporating clinical and early laboratory variables within 0 to 6 hours of ICU admission.

C

Comparator

Simpler logistic regression prediction models including clinical base variables alone, clinical variables plus LHR, clinical variables plus BHR, clinical variables plus both ratios, and clinical variables plus raw biomarker components (lactate, BUN, and hemoglobin).

O

Outcome

All-cause post-landmark in-hospital mortality (occurring after the 6-hour ICU admission landmark).hard clinical

Main Result

Odds Ratio: 1.26 (95% CI 1.13–1.41)

p-value: p=<0.001

In critically ill patients with heart failure and atrial fibrillation, a complex XGBoost machine-learning model and composite biomarker ratios (LHR and BHR) offered no external discrimination advantage over a simpler logistic regression model using raw lactate, BUN, and hemoglobin components for predicting in-hospital mortality.

Limitations

No a priori sample-size calculation was performed, and precision in the external validation cohort was limited by only 125 deaths.
Diagnosis fields were admission-level rather than hour-stamped, so HF and AF were treated as documented comorbid diagnoses rather than conditions proven to be newly established within the first 6 hours.
Requiring all three biomarkers within 6 hours selected a higher-risk population with higher mortality than excluded landmark patients.
Substantial case-mix differences existed between MIMIC-IV and eICU-CRD (SMDs > 0.5 for anion gap, cardiac ICU, and medical ICU), contributing to reduced transportability and imperfect external calibration.
Admission-level rather than hour-stamped diagnosis fields meant HF and AF were treated as documented comorbid diagnoses rather than conditions proven to be newly established within the first 6 hours.
Precision limits in external validation imposed by only 125 external deaths in the eICU cohort.
Selection bias toward a higher-risk population due to complete-biomarker eligibility within 6 hours (mortality was higher among included than excluded landmark patients: MIMIC 19.4% vs 12.0%; eICU 31.0% vs 14.1%).
Substantial case-mix differences between databases (SMDs > 0.5 for anion gap, cardiac ICU, and medical ICU) that may contribute to reduced transportability and decline in external performance.
Systematic underprediction and imperfect external calibration in the locked XGBoost model (mean predicted risk 25.3% vs observed mortality 31.0%; calibration-in-the-large 0.324, calibration slope 0.726).

Ask about this paper

Answers are grounded in this paper and cite their sources.

Implication

Simple logistic regression with raw biomarkers matches XGBoost discrimination; leaves open whether machine learning adds value in this observational cohort.

Abstract

Early prognostic assessment in critically ill patients with heart failure (HF) and atrial fibrillation (AF) may benefit from multidomain biomarkers, but composite ratios and machine-learning models require direct benchmarking against simpler alternatives. We used MIMIC-IV v3.1 for development and eICU-CRD v2.0 for external validation. Adults with HF and AF and valid lactate, blood urea nitrogen (BUN), and hemoglobin within 6 h of ICU admission entered a fixed 6-hour landmark cohort; deaths or discharges at or before 6 h were excluded. The primary outcome was all-cause post-landmark in-hospital mortality. MIMIC included 3,925 patients (760 deaths; 19.4%) and eICU 403 (125 deaths; 31.0%). Adjusted lactate-to-hemoglobin ratio was associated with mortality in both cohorts, whereas BUN-to-hemoglobin ratio was imprecise externally. Locked XGBoost achieved external AUC 0.668 (95% CI 0.610–0.723); its AUC was 0.053 higher than clinical plus LHR (95% CI 0.002–0.101) but showed no clear discrimination advantage over clinical plus raw lactate/BUN/hemoglobin logistic regression (AUC 0.672; delta AUC − 0.004, 95% CI -0.042 to 0.034). Removing both ratios changed XGBoost AUC by 0.0005.

Discussion

No takes yet. Share an insight, caveat, or question.

Cite This Study

Alioglu et al. (2026) conducted a cohort in Critically ill patients with heart failure and atrial fibrillation (n=4,328). Early lactate-to-hemoglobin ratio (LHR), BUN-to-hemoglobin ratio (BHR), and locked XGBoost machine-learning model vs. Clinical base models and logistic regression with raw lactate, BUN, and hemoglobin components was evaluated on All-cause post-landmark in-hospital mortality (OR 1.26, 95% CI 1.13-1.41, p=<0.001). Lactate-to-hemoglobin ratio was associated with in-hospital mortality, while locked XGBoost did not outperform raw-component logistic regression.

synapsesocial.com/papers/6af9ebcdd393cbc380a0505d https://doi.org/10.1038/s41598-026-79598-5

View Full Paper

Ask AI

Bookmark

Share

Early multibiomarker machine-learning prediction of in-hospital mortality in critically ill patients with heart failure and atrial fibrillation

Key Points

Study Design

Structured PICO

Main Result

Limitations

Ask about this paper

Implication

Abstract

Discussion

Cite This Study

Early multibiomarker machine-learning prediction of in-hospital mortality in critically ill patients with heart failure and atrial fibrillation

Key Points

Study Design

Structured PICO

Main Result

Limitations

Ask about this paper

Implication

Abstract

Discussion

Cite This Study

Keep reading

Keep reading