Construction and validation of prognostic models for acute kidney disease and mortality in patients at risk of malnutrition: an interpretable machine learning approach.

BACKGROUND

Acute kidney injury (AKI) is a prevalent complication in patients at risk of malnutrition, elevating the risks of acute kidney disease (AKD) and mortality. AKD reflects the adverse events developing after AKI. This study aimed to develop and validate machine learning (ML) models for predicting the occurrence of AKD, AKI and mortality in patients at risk of malnutrition.

METHODS

We retrospectively reviewed the medical records of patients at risk of malnutrition. Eight ML algorithms were employed to predict AKD, AKI and mortality. The performance of the best model was evaluated using various metrics and interpreted using the SHapley Additive exPlanation (SHAP) method. An artificial intelligence (AI)-driven web application was also created based on the best model.

RESULTS

A total of 13 395 patients were included in our study. Among them, 1751 (13.07%) developed subacute AKD, 1253 (9.35%) were transient AKI, and 1455 (10.86%) met both AKI and AKD criteria. The incidence rate of mortality was 6.74%. The light gradient boosting machine (LGBM) outperformed other models in predicting AKD, AKI and mortality, with area under curve values of 0.763, 0.801 and 0.881, respectively. The SHAP method revealed that AKI stage, lactate dehydrogenase, albumin, aspirin usage and serum creatinine were the top five predictors of AKD. An online prediction website for AKI, AKD and mortality was developed based on the final models.

CONCLUSIONS

The LGBM models provide an effective method for predicting AKD, AKI and mortality at an early stage in patients at risk of malnutrition, enabling prompt interventions. Compared with the AKD model, the models for predicting AKI and mortality perform better. The AI-driven web application can significantly aid in creating personalized preventive measures. Future work will aim to expand the application to larger, more diverse populations, incorporate additional biomarkers and refine ML algorithms to improve predictive accuracy and clinical utility.