[Construction of a predictive model of death for sepsis-associated acute kidney injury].

OBJECTIVE

To establish a predictive model nomogram for 30-day death in patients with sepsis-associated acute kidney injury (SA-AKI) by using the data from the large international database, the Electronic Intensive Care Unit-Collaborative Research Database (eICU-CRD), and to validate its predictive performance.

METHODS

A retrospective cohort study was conducted using data from the eICU-CRD. Data of SA-AKI patients were screened from the eICU-CRD database, including demographic characteristics, medical history, SA-AKI type, Kidney Disease: Improving Global Outcomes (KDIGO)-AKI staging, severity of illness scores, vital signs, laboratory indicators, and treatment measures; with admission time as the observation start point, death as the outcome event, and a follow-up time of 30 days. Relevant variables of patients with different 30-day prognoses were compared. Univariate Logistic regression analysis and multivariate Logistic regression forward likelihood ratio analysis were used to screen for risk factors associated with 30-day death in SA-AKI patients, and a predictive model nomogram was constructed. Receiver operator characteristic curve (ROC curve), calibration curve, and Hosmer-Lemeshow test were used to validate the predictive performance of the model.

RESULTS

A total of 201 SA-AKI patients' data were finally enrolled, among which 51 survived for 30 days and 150 died, with a mortality of 74.63%. Compared with the survival group, patients in the death group were older [years old: 68 (60, 78) vs. 59 (52, 69), P < 0.01], had lower body weight, proportion of transient SA-AKI, platelet count (PLT) and blood glucose [body weight (kg): 79 (65, 95) vs. 91 (71, 127), proportion of transient SA-AKI: 61.33% (92/150) vs. 82.35% (42/51), PLT (×10/L): 207 (116, 313) vs. 260 (176, 338), blood glucose (mmol/L): 5.5 (4.4, 7.1) vs. 6.4 (5.1, 7.6), all P < 0.05] and higher proportion of persistent SA-AKI, sequential organ failure assessment (SOFA) score, lactic acid (Lac), and total bilirubin [TBil; proportion of persistent SA-AKI: 38.67% (58/150) vs. 17.65% (9/51), SOFA score: 7 (5, 22) vs. 5 (2, 7), Lac (mmol/L): 0.4 (0.2, 0.7) vs. 0.3 (0.2, 0.4), TBil (μmol/L): 41.0 (17.1, 51.3) vs. 18.8 (17.1, 34.2), all P < 0.05]. Univariate Logistic regression analysis showed that age [odds ratio (OR) = 1.035, 95% confidence interval (95%CI) was 1.013-1.058, P = 0.002], body weight (OR = 0.987, 95%CI was 0.977-0.996, P = 0.007), persistent SA-AKI (OR = 2.942, 95%CI was 1.333-6.491, P = 0.008), SOFA score (OR = 1.073, 95%CI was 1.020-1.129, P = 0.006), PLT (OR = 0.998, 95%CI was 0.996-1.000, P = 0.034), Lac (OR = 1.142, 95%CI was 1.009-1.292, P = 0.035), TBil (OR = 1.422, 95%CI was 1.070-1.890, P = 0.015) were associated with 30-day death risk in SA-AKI patients. Multivariate Logistic regression forward likelihood ratio analysis showed that age (OR = 1.051, 95%CI was 1.023-1.079, P = 0.000), body weight (OR = 0.985, 95%CI was 0.974-0.995, P = 0.005), cardiovascular disease (OR = 9.055, 95%CI was 1.037-79.084, P = 0.046), persistent SA-AKI (OR = 3.020, 95%CI was 1.258-7.249, P = 0.013), SOFA score (OR = 1.076, 95%CI was 1.013-1.143, P = 0.017), and PLT (OR = 0.997, 95%CI was 0.995-1.000, P = 0.030) were independent risk factors for 30-day death in SA-AKI patients. Based on the above risk factors, a predictive model nomogram for 30-day death in SA-AKI patients was constructed. ROC curve analysis showed that the area under the ROC curve (AUC) of the model was 0.798 (95%CI was 0.722-0.873), with a sensitivity of 86.7% and a specificity of 62.7%. Calibration curve showed that the fitted curve was close to the standard line, indicating that the predicted probability was close to the actual probability, suggesting good predictive performance of the model. Hosmer-Lemeshow test showed χ = 6.393, df = 8, P = 0.603 > 0.05, suggesting that the model could fit the observed data well. The quality of model fitting was judged by the accuracy of model prediction. The results showed that the prediction accuracy rate of the model was 95.3%, and the overall prediction accuracy rate of the model was 81.6%, indicating good model fitting.

CONCLUSIONS

A predictive model for 30-day death in SA-AKI patients based on risk factors can be successfully constructed, and the model has high accuracy, sensitivity, reliability, and certain specificity, which can help to early identify high-risk patients for death and adopt more proactive treatment strategies.