AKI Risk Model in Older Heart Failure Inpatients

Key Takeaways

• Among older adults hospitalized with heart failure at a single Chinese center, 11.2% developed in-hospital AKI.
• Admission cystatin C, lower LVEF, chronic preadmission diuretic use, and several comorbidities were associated with AKI risk.
• Logistic regression had the highest held-out validation AUC among the compared approaches.
• Calibration was acceptable, but modeled decision-curve net benefit was confined to lower risk thresholds.

For an older adult admitted with heart failure, kidney function and cardiac status are assessed while acute kidney injury (AKI) remains a potential complication of the hospital stay. Admission information collected before AKI develops may help distinguish who will experience it, provided its predictive performance is tested against subsequent outcomes.

In Liu and colleagues’ AKI prediction-model study in Clinical Interventions in Aging, investigators retrospectively analyzed 7,076 heart failure inpatients aged at least 65 years at the General Hospital of the Northern Theater Command in Shenyang, China. Patients with AKI at admission or incomplete predictor or outcome data were excluded; missing values were not imputed. In-hospital AKI was defined by changes in serum creatinine or urine output. The prediction clock started at admission: predictors were measured within the first 24 hours, before AKI, while chronic medication use was recorded from before admission. Investigators screened candidate variables with univariate analysis, cross-validated least absolute shrinkage and selection operator (LASSO), and multivariable logistic regression. They trained multiple approaches on a random internal training split; the held-out validation subset was not used to select or fit variables.

During hospitalization, 796 patients developed AKI. The selected logistic predictors were chronic preadmission diuretic use, hypertension, diabetes, valvular heart disease, chronic obstructive pulmonary disease (COPD), chronic kidney disease (CKD), admission cystatin C, and left ventricular ejection fraction (LVEF); higher LVEF tracked lower estimated risk. Logistic regression ranked first among the compared approaches for validation area under the receiver operating characteristic curve (AUC), at 0.804 (95% confidence interval 0.777–0.830). Random forest and Bayes approaches had lower validation AUCs, without a reported formal test of between-model superiority.

The logistic model’s training AUC was 1.000, substantially above its held-out validation performance. Its validation calibration slope was 0.835; the reported Hosmer–Lemeshow tests, slopes, and Brier scores across training and validation were collectively consistent with acceptable rather than excellent calibration. The validation curve showed mild overestimation at the lower end of predicted risk. Validation decision-curve net benefit was positive below an approximately 0.28 risk threshold and negative above it, a modeled finding rather than demonstrated improvement in patient outcomes. Investigators also derived a nomogram and an integer-based risk score from the selected logistic model.

The retrospective cohort came from one Chinese hospital, excluded patients with incomplete data, and was evaluated with a random internal split rather than an independent external cohort. Investigators raised possible overfitting or complete separation as explanations for the near-perfect training discrimination, limiting claims about performance elsewhere. Modeled utility depended on the decision threshold, while neither the nomogram nor decision-curve analysis tested implementation or patient outcomes. Chronic preadmission diuretic use was an associated predictor, not evidence that diuretics caused AKI.

Liu and colleagues described the admission-based logistic model as a candidate for risk stratification among older heart failure inpatients, with acceptable internal validation performance. They called for independent external testing, including broader prospective evaluation, before routine implementation.

Clinician Questions

Were dialysis patients or kidney transplant recipients included in the heart failure AKI prediction model?

Liu and colleagues excluded patients receiving dialysis, patients with end-stage renal disease, and kidney transplant recipients from their cohort of older adults hospitalized with heart failure in China. The AKI model was therefore not evaluated in those groups.

Did NYHA functional class or NT-proBNP independently predict AKI in older heart failure inpatients?

New York Heart Association functional class and N-terminal pro–B-type natriuretic peptide were not independent predictors of in-hospital acute kidney injury among the older hospitalized heart failure patients analyzed by Liu and colleagues.

What serum creatinine value was used as baseline to identify AKI in older heart failure inpatients?

For older heart failure inpatients, Liu and colleagues used the serum creatinine recorded at hospital admission as baseline when assessing subsequent AKI during hospitalization; urine output was also part of the AKI definition.