📄 medicine

A three-axis burden score and in-hospital mortality in ICU patients with heart failure and atrial fibrillation: limited incremental discrimination and transportability in a retrospective cohort study

This retrospective cohort study concludes that while a three-axis burden score based on nutritional status, inflammation, and lactate clearance shows a descriptive association with in-hospital mortality in heart failure patients with atrial fibrillation within the derivation dataset, it fails to demonstrate transportable predictive value or significant incremental discrimination when validated in an external ICU cohort.

Original authors: Jiaodan Wang¹, Rui Yu

Published 2026-10-22

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jiaodan Wang¹, Rui Yu

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the intense, high-stakes environment of an intensive care unit, doctors constantly look for clues that might predict whether a patient will survive their stay. When a patient arrives with two serious conditions at once—heart failure, where the heart struggles to pump blood effectively, and atrial fibrillation, a chaotic rhythm that disrupts that pumping—the risk of death rises significantly. To navigate this danger, clinicians often rely on standard severity scores that weigh many factors, from blood pressure to organ function. However, researchers have long wondered if a simpler combination of three specific blood test results could offer a clearer picture. These three markers represent different aspects of the body's struggle: the nutritional reserve, measured by a score that combines protein levels and white blood cell counts; the level of systemic inflammation, indicated by the ratio of two types of infection-fighting cells; and the body's ability to clear a waste product called lactate, which builds up when tissues are starved of oxygen. The question was whether counting how many of these three markers were "bad" could serve as a quick, reliable warning system for doctors.

A team of researchers set out to test this idea using data from two massive, anonymized databases of intensive care patients across the United States. They focused specifically on adults admitted to the ICU who had both heart failure and atrial fibrillation. The team defined a simple "burden score" ranging from zero to three. A patient received one point for each of the three markers that fell into a high-risk category: a low nutritional score, a high inflammation ratio, or a poor rate of lactate clearance. The researchers first established the thresholds for what counted as "high risk" using data from one database, and then applied those exact same rules to a second, completely separate database to see if the pattern held up. Their goal was to determine if this simple count could predict death better than the complex, standard models doctors already use, and if it would work consistently across different groups of patients.

The results showed that the score did tell a story, but not the one the researchers had hoped for as a standalone tool. In the first database, which included over 7,500 patients, the risk of dying in the hospital clearly rose as the score increased. Patients with a score of zero had a mortality rate of about 11 percent, while those with a score of three faced a risk of nearly 25 percent. This trend was visible in the second database as well, suggesting that having more of these negative markers is indeed linked to a worse outcome. However, when the researchers tried to use this score to improve upon the existing, detailed medical models, the results were much less impressive. The standard models, which already included information about albumin levels, white blood cell counts, and lactate, were already quite good at predicting risk. Adding the simple three-point count to these models provided only a tiny, almost invisible improvement in accuracy.

The situation became even more complicated when the researchers looked at the second database more closely. While the overall trend of higher scores correlating with higher death rates remained, the specific ranking of the risk levels became messy. In this second group, patients with a score of one actually appeared to have a higher risk of death than those with a score of two, breaking the smooth, predictable pattern seen in the first group. Furthermore, when the researchers adjusted for the many other factors already included in the standard models, the score lost its statistical power in the second database. It no longer showed a clear, independent link to mortality. This suggests that the information provided by these three blood tests is largely already captured by the other data points the doctors are already collecting. The score did not reveal a hidden danger that the standard models missed; rather, it seemed to be a rough summary of risks that were already being measured in more precise ways.

The study also highlighted a practical hurdle for using this score in real-world hospitals. To calculate the third part of the score, the body's ability to clear lactate, doctors need two separate blood tests taken at different times. In the second database, this second test was missing for the vast majority of patients, making it impossible to calculate the score for most of them. Even in the first database, the test was missing for a significant portion of the patients. This means that while the score works in theory for those who have the data, it cannot be applied universally in a busy hospital setting where not every patient gets a second lactate test. The researchers concluded that while the count of these three markers is a useful way to describe the severity of a patient's condition in a research setting, it is not a validated or reliable tool for predicting individual outcomes in a new hospital. It does not offer a shortcut that improves upon the complex, established methods doctors currently use to care for critically ill patients.

Technical Summary: A Three-Axis Burden Score in ICU Patients with Heart Failure and Atrial Fibrillation

Problem Statement
Patients in the Intensive Care Unit (ICU) with concurrent heart failure (HF) and atrial fibrillation (AF) represent a high-risk population with substantial in-hospital mortality. While generic severity scores (e.g., APACHE II, SOFA) exist, they are not specifically designed for this cardio-metabolic profile. Previous research has linked individual biomarkers—Prognostic Nutritional Index (PNI), Neutrophil-to-Lymphocyte Ratio (NLR), and dynamic lactate clearance rate (dLCR)—to short-term risk. However, composite models are often complex for bedside use, and single-biomarker studies fail to capture the interplay of nutritional, inflammatory, and perfusion deficits. The study investigates whether a simple 0–3 "burden score," counting the presence of low PNI, high NLR, and low dLCR, can serve as a validated, transportable prediction tool that improves mortality discrimination beyond a robust reference model.

Methodology
This retrospective cohort study utilized two large, de-identified ICU databases:

Derivation Cohort: MIMIC-IV v3.1 (n = 7,562; 16.2% mortality).
External Validation Cohort: eICU-CRD 2.0 (n = 2,588; 16.0% mortality).

Data Processing and Score Definition:

Cohort Selection: Adults (≥18 years) with a first ICU stay ≥24 hours and concurrent HF and AF diagnoses.
The Burden Score: A 0–3 count based on three axes. Cut-points were defined as the medians of observed values in the MIMIC-IV cohort and "frozen" for application in eICU-CRD without re-estimation:
1. Nutritional: PNI < 38.0 (calculated from albumin and absolute lymphocyte count).
2. Inflammatory: NLR ≥ 9.24.
3. Perfusion: dLCR < 12.5% (relative fall in lactate between earliest and latest samples within a -24h to +48h window).
Handling Missing Data: Multiple imputation (m=10) using chained equations with Bayesian ridge regression was employed to handle missing laboratory values.
Reference Model: A 43-variable logistic model including demographics, comorbidities, vital signs, organ support, and a reconstructed SOFA score. Crucially, this model already contained albumin, white-cell count, and lactate level, meaning the score's incremental value was tested against these components.
Statistical Analysis: Associations were estimated via odds ratios (OR) from logistic models. Discrimination was assessed via Area Under the Curve (AUC) changes ( $\Delta$ AUC) using out-of-fold cross-validation and bootstrap resampling. External transportability was evaluated by applying MIMIC-IV coefficients to eICU-CRD, followed by two-parameter recalibration.

Key Results

Mortality Association: In imputed data, in-hospital mortality increased with the burden score in both cohorts (MIMIC-IV: 10.8% to 24.9%; eICU-CRD: 10.2% to 24.0%).
- In MIMIC-IV, each additional point was associated with an adjusted OR of 1.20 (95% CI 1.08–1.33).
- In eICU-CRD, the adjusted OR was 1.14 (95% CI 0.97–1.35), which was not statistically significant.
- When restricted to patients with all three axes measured (a sicker subset), the gradient in eICU-CRD was non-monotonic (score 1 mortality > score 2).
Incremental Discrimination:
- In MIMIC-IV, adding the 0–3 score to the reference model increased the AUC by only 0.0028 (95% CI 0.0011–0.0045). This gain was smaller than adding the three axes as continuous terms (+0.0046).
- In eICU-CRD, when models were refitted, the score provided no detectable incremental discrimination ( $\Delta$ AUC = +0.0009, CI included zero).
Transportability: The reference model transported to eICU-CRD with an AUC of 0.711 but was miscalibrated (slope 0.70, intercept -0.95). Two-parameter recalibration improved the Brier score from 0.140 to 0.124. However, the burden score itself did not add detectable value upon transport or refitting.
Individual Axes: After full adjustment in MIMIC-IV, only dLCR remained significantly associated with mortality (OR 0.79). PNI and NLR largely overlapped with information already present in the reference model (albumin and white-cell count).

Significance and Claims
The paper concludes that the three-axis burden score is primarily a descriptive stratification tool tied to the specific cut-points derived from the MIMIC-IV cohort, rather than a validated, transportable prediction instrument.

Limited Incremental Value: The score adds a negligible amount of discrimination (0.0028 AUC) to a comprehensive reference model in the derivation data, and this gain disappears in the external cohort.
Lack of Validation: The adjusted association is supported in MIMIC-IV but not in eICU-CRD. The score does not consistently improve discrimination across databases.
Clinical Utility Constraints: The absolute risks presented in the study (e.g., Table 3) are "standardized risks" derived from the full 43-variable model and cannot be calculated from the three laboratory values alone.
Conclusion: The count is not a standalone predictor. While it tracks mortality trends in imputed data, it does not offer a clinically meaningful improvement over existing routine laboratory data and reference models, nor does it function as a validated tool for external application without recalibration of the underlying reference model.

Technical Summary: A Three-Axis Burden Score in ICU Patients with Heart Failure and Atrial Fibrillation

More like this