Predicting renal function recovery and short-term reversibility among acute kidney injury patients in the ICU: comparison of machine learning methods and conventional regression

Abstract

Background: Acute kidney injury (AKI) is one of the most frequent complications of critical illness. We aimed to explore the predictors of renal function recovery and the short-term reversibility after AKI by comparing logistic regression with four machine learning models.

Methods: We reviewed patients who were diagnosed with AKI in the MIMIC-IV database between 2008 and 2019. Recovery from AKI within 72 h of the initiating event was typically recognized as the short-term reversal of AKI. Conventional logistic regression and four different machine algorithms (XGBoost algorithm model, Bayesian networks [BNs], random forest [RF] model, and support vector machine [SVM] model) were used to develop and validate prediction models. The performance measures were compared through the area under the receiver operating characteristic curve (AU-ROC), calibration curves, and 10-fold cross-validation.

Results: A total of 12,321 critically ill adult AKI patients were included in our analysis cohort. The renal function recovery rate after AKI was 67.9%. The maximum and minimum serum creatinine (SCr) within 24 h of AKI diagnosis, the minimum SCr within 24 and 12 h, and antibiotics usage duration were independently associated with renal function recovery after AKI. Among the 8364 recovered patients, the maximum SCr within 24 h of AKI diagnosis, the minimum Glasgow Coma Scale (GCS) score, the maximum blood urea nitrogen (BUN) within 24 h, vasopressin and vancomycin usage, and the maximum lactate within 24 h were the top six predictors for short-term reversibility of AKI. The RF model presented the best performance for predicting both renal functional recovery (AU-ROC [0.8295 ± 0.01]) and early recovery (AU-ROC [0.7683 ± 0.03]) compared with the conventional logistic regression model.

Conclusions: The maximum SCr within 24 h of AKI diagnosis was a common independent predictor of renal function recovery and the short-term reversibility of AKI. The RF machine learning algorithms showed a superior ability to predict the prognosis of AKI patients in the ICU compared with the traditional regression models. These models may prove to be clinically helpful and can assist clinicians in providing timely interventions, potentially leading to improved prognoses.

Keywords: Acute kidney injury; machine learning; renal function recovery; renal function recovery time.

Figure 1.

Flowchart depicting the number of critically ill patients included in the analysis after applying the exclusion criteria.

Figure 2.

Importance of the matrix plot of the AKI predictors in the random forest model among critically ill patients. AKI: acute kidney disease; HR: heart rate; SAPS II: Simplified Acute Physiology Score II; SCr: serum creatinine; SpO2: peripheral oxygen saturation; T: temperature.

Figure 3.

Receiver operating characteristic curve for estimating the discrimination of the logistic regression model, XGBoost model, random forest model (RF), and support vector machine model (SVM).

Figure 4.

Calibration curve of the renal function recovery prediction models (A) logistic regression and (B) random forest in the training set.

Figure 5.

Variable importance plot for the short-term reversibility in the random forest model. GCS: Glasgow Coma Scale; SCr: serum creatinine.