Key result
Six-hour XGBoost modeling improves prediction of poor diuretic response by ~0.08 AUROC in ICU heart failure.
Why the study?
Early response to loop diuretics may identify critically ill heart failure patients who will have low subsequent urine output, and it was unclear whether information recorded during the first six hours improves prediction across intensive care databases.
Does incorporating six-hour early-response clinical information into machine learning prediction models improve the prediction of poor subsequent urine output (hours 6-24) compared to baseline information alone in critically ill adults with heart failure receiving intravenous loop diuretics?
Cohort (n=1,995)
Yes
Does incorporating six-hour early-response clinical information into machine learning prediction models improve the prediction of poor subsequent urine output (hours 6-24) compared to baseline information alone in critically ill adults with heart failure receiving intravenous loop diuretics?
Incorporating six-hour early clinical response data significantly improves the prediction of subsequent poor urine output in critically ill heart failure patients compared to baseline data alone, though complex machine learning (XGBoost) offered no external advantage over parsimonious elastic-net regression.
Answers are grounded in this paper and cite their sources.
May improve early oliguria prediction in ICU HF; leaves open prospective validation before clinical adoption.
Background Early response to loop diuretics may help identify critically ill patients with heart failure who will have low subsequent urine output. We assessed whether information recorded during the first six hours improved prediction across intensive care databases. Methods We developed models in MIMIC-IV and externally evaluated them in eICU-CRD. The prediction landmark was six hours after the first qualifying intravenous loop-diuretic record. Poor subsequent response was defined as mean urine output during hours 6–24 at or below the development-cohort lower quartile (approximately 0.718 mL/kg/h). Four candidate algorithms underwent subject-grouped, candidate-specific nested cross-validation. A development-selected XGBoost model was applied unchanged externally. Evaluation included area under the receiver operating characteristic curve (AUROC), calibration, threshold performance, decision curves and inverse-probability-of-inclusion weighting. Results The development cohort included 1,450 stays and 363 events. The external cohort included 545 stays and 147 events across 69 hospitals. Development out-of-fold AUROC was 0.760 (95% confidence interval 0.732–0.787), conditional on algorithm-family selection and potentially optimistic. External AUROC was 0.729 (0.690–0.789). Compared with baseline elastic-net regression, external AUROC increased by 0.080 (0.028–0.126). However, early-response elastic-net regression achieved a higher external AUROC of 0.756; the paired XGBoost-minus-elastic-net difference was − 0.027 (− 0.057 to − 0.004). External calibration slope was 0.682. At the locked threshold, sensitivity was 75.5%, specificity 56.0% and positive predictive value 38.8%; incremental net benefit over reviewing all patients was uncertain. Inclusion-weighted external AUROC was 0.690 (0.632–0.750). Conclusions Six-hour early-response information improved prediction of subsequent low urine output, but the more complex algorithm showed no advantage. Calibration limitations, selection for observable outcomes and measurement differences between databases limit immediate clinical use. Prospective evaluation should compare parsimonious models and establish whether risk-guided review improves care.
No takes yet. Share an insight, caveat, or question.
庄小环 et al. (2026) conducted a cohort in Critically ill adults with heart failure receiving intravenous loop diuretics (n=1,995). Six-hour early-response XGBoost prediction model vs. Baseline elastic-net regression model was evaluated on Poor subsequent urine-output response during hours 6-24 (mean urine output <=0.718 mL/kg/h) (AUROC difference 0.080, 95% CI 0.028-0.126). Six-hour early-response XGBoost modeling improved external prediction of low 6-24h urine output over baseline regression in ICU heart failure patients.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: