A four-parameter laboratory risk stratification model at admission for 90-day mortality in hepatorenal syndrome-acute kidney injury: a single-center derivation and internal validation study.

Abstract

Background: Hepatorenal syndrome-acute kidney injury (HRS-AKI) is a high-mortality complication of advanced cirrhosis. Early risk stratification after hospital admission may support clinical assessment and research on escalation pathways, but simple admission-based models that avoid in-hospital data leakage remain limited.

Methods: We performed a retrospective single-center derivation and internal-validation study of hospitalized adults with chart-review-confirmed HRS-AKI according to contemporary ADQI/ICA criteria. Candidate predictors were restricted a priori to admission laboratory values to minimize data leakage. Admission variables were prioritized using random-forest importance rankings under prespecified parsimony constraints, followed by endpoint-specific closed-form logistic regression models using an identical four-predictor set across outcomes. The primary endpoint was 90-day all-cause mortality. In-hospital hemodialysis and ICU admission were prespecified secondary endpoints and are reported as exploratory. Internal validation used bootstrap resampling with assessment of apparent and optimism-corrected discrimination and overall accuracy; calibration slope and calibration-in-the-large/intercept were reported descriptively. For contextual benchmarking, admission MELD, MELD-Na, and MELD 3.0 were evaluated as single-predictor logistic benchmark models for 90-day mortality within the same cohort.

Results: The final analysis included 77 patients. The final predictor set comprised INR, albumin, ln(total bilirubin), and ln(C-reactive protein). For 90-day mortality, the final four-laboratory model showed an apparent AUC of 0.781 (95% CI, 0.715-0.926) and an optimism-corrected AUC of 0.758 (95% CI, 0.648-0.859), with an apparent Brier score of 0.144 (95% CI, 0.081-0.176), an optimism-corrected Brier score of 0.168 (95% CI, 0.117-0.212), an optimism-corrected calibration slope of 0.770, and an optimism-corrected calibration-in-the-large/intercept of 0.168. Performance for in-hospital hemodialysis and ICU admission was limited, with optimism-corrected AUCs of 0.580 (95% CI, 0.451-0.733) and 0.631 (95% CI, 0.514-0.737), respectively. MELD-family benchmarks showed numerically lower within-cohort discrimination than the four-laboratory model; these comparisons were descriptive and hypothesis-generating.

Conclusions: In this single-center derivation and internal-validation study, an admission-based four-laboratory model showed moderate internally validated discrimination for 90-day mortality in hospitalized patients with HRS-AKI. The model should be interpreted as a preliminary risk-stratification approach requiring external validation and, if necessary, recalibration before clinical use. Secondary endpoint models for hemodialysis and ICU admission are exploratory and should not be used for threshold-based decision-making.