Background Acute kidney injury (AKI) is characterized by an abrupt reduction in the kidney’s functioning, and it has long-term repercussions. Predictive models are used widely in predicting mortality, identifying patients who are at risk, and making diagnoses. This study was conducted to compare the predictive accuracy of machine learning models with logistic regression (LR) for mortality among patients with AKI. Materials and Methods This study consists of data from 994 patients who underwent treatment for AKI in a tertiary health care center in South India between 2013 to 2021. Univariate Analysis was used to identify potential AKI predictors. The predictive performance of Multiple Binary logistic regression (MBLR) and machine learning models was compared using accuracy rate, sensitivity, specificity, and area under the curve (AUC). The split sample method was used for internal validation. Results In the training dataset, the Decision Tree (DT) and Random Forest (RF) achieved high AUCs of 0.87 and 0.86, respectively. However, in the testing dataset, their performance declined, suggesting potential overfitting. In contrast, LR and artificial neural network (ANN) demonstrated stable accuracy in both training and testing, with an AUC of 0.80, indicating better generalizability for clinical application. Conclusion While DT and RF showed strong predictive capabilities in training, their reduced performance in testing limits their clinical applicability. LR and ANN demonstrated consistent accuracy across datasets, making them more reliable for real-world mortality prediction in patients with AKI. These findings highlight the importance of carefully validating machine learning models before clinical implementation.
Renukadevi et al. (Fri,) studied this question.