Cardiovascular

Integrating 72-h BUN/Cr trajectories with interpretable machine learning for 28-day mortality after acute stroke: a dual-center landmark study.

Cheng Z, Lu Z, et al. • Frontiers in neurology • 2026

PubMed 42903085 DOI 10.3389/fneur.2026.1865384

TL;DR

BUN/Cr trajectories over the first 72 hours after acute stroke were associated with 28-day post-landmark mortality and summarized the three-day renal course, but trajectory class added no clear discrimination gain beyond routine clinical and static renal variables, and the LightGBM model's performance fell in external validation.

Key Findings

Results

A quadratic three-class trajectory model identified low/stable, high U-shaped, and rising-to-plateau 72-h BUN/Cr patterns with graded differences in mortality.

The study included 2,805 ICU patients with acute ischemic stroke, nontraumatic intracerebral hemorrhage, or nontraumatic subarachnoid hemorrhage from MIMIC-IV version 3.1.
Mortality rates were 21.6%, 35.9%, and 39.7% for Class 1 (low/stable), Class 2 (high U-shaped), and Class 3 (rising-to-plateau).
The trajectory models were developed in the training set (n = 1,964) and applied without modification to the internal test set (n = 841) and the Qinghai external validation cohort (n = 387).

Results

Both higher-risk trajectory classes had significantly higher adjusted hazards of post-landmark mortality than the low/stable class.

Adjusted hazard ratio versus Class 1 was 1.87 (95% CI 1.25-2.79) for Class 2.
Adjusted hazard ratio versus Class 1 was 1.73 (1.31-2.29) for Class 3.
Analysis used a landmark design with Cox regression.

Results

The training-selected LightGBM model showed moderate discrimination and good calibration in the internal test set.

AUROC was 0.734 (95% CI 0.695-0.773).
Calibration slope was 0.894 and the observed-to-expected (O/E) ratio was 1.015.
Eight algorithms were evaluated, and LightGBM was selected in the training set.

Results

Performance of the frozen LightGBM model degraded substantially in the Qinghai external validation cohort.

AUROC was 0.624 (95% CI 0.559-0.689).
Calibration slope was 0.459, indicating poor calibration.
O/E ratio was 1.920, indicating underestimation of observed mortality relative to predictions.
The model was applied without modification (frozen external evaluation) to the cohort of 387 patients.

Results

Adding trajectory class to a model already containing routine clinical and static renal variables did not clearly improve discrimination.

AUROC change was 0.003 in the internal test set (p = 0.515).
AUROC change was -0.014 in the Qinghai external validation cohort (p = 0.130).
Prespecified nested models and paired comparisons were used.

Conclusions

The authors conclude that BUN/Cr trajectories summarize the three-day renal course and provide an interpretable framework for 72-h reassessment, while adding only modest information beyond routine variables.

Interpretability was provided by TreeSHAP.
Evaluation also included calibration, decision curves, and paired comparisons.
Trajectory class was associated with post-landmark mortality despite lacking a clear incremental discrimination gain.

What This Means

This study looked at how the ratio of two routine blood tests, blood urea nitrogen and creatinine (BUN/Cr), changed over the first three days in intensive care patients with stroke. Using records from 2,805 patients, researchers found three typical patterns: a low and stable ratio, a high ratio that dipped then rose (U-shaped), and a ratio that rose and then leveled off. Death within 28 days occurred in about 22% of the low/stable group, versus about 36% and 40% in the other two groups. Even after accounting for other factors, the two higher-ratio patterns were linked to higher risk of death. The team also built machine-learning models to predict 28-day death. The best one, LightGBM, performed moderately well on a held-out portion of the same database (AUROC 0.734). On a separate hospital cohort from Qinghai (387 patients), it did noticeably worse (AUROC 0.624) and its risk estimates were poorly calibrated, underestimating the actual number of deaths. Importantly, adding the trajectory pattern to a model that already included standard clinical information and a single kidney measurement did not clearly improve predictions. This research suggests that BUN/Cr trajectories describe a meaningful pattern linked to outcomes and may help organize a 72-hour reassessment, but they are not a stand-alone improvement in predicting death, and models like this need local checking before wider use.

Have a question about this study?

Citation

Cheng Z, Lu Z, Wang S, He M, Lv Y, Duoji Z, et al.. (2026). Integrating 72-h BUN/Cr trajectories with interpretable machine learning for 28-day mortality after acute stroke: a dual-center landmark study.. Frontiers in neurology. https://doi.org/10.3389/fneur.2026.1865384

Key Findings

What This Means

Have a question about this study?

Related Research

Citation