A Deep Learning Model for Dynamic Prediction of Acute Kidney Injury in Heart Failure Patients.

BACKGROUND

In this we study aimed to develop and validate a dynamic prediction model for acute kidney injury (AKI) in heart failure (HF) patients.

METHODS

Using data from 7636 HF patients in the Medical Information Mart for Intensive Care IV (MIMIC-IV) version 3.1 database, we constructed a long short-term memory model with dynamic focal loss to handle class imbalance. We designed 2 prediction perspectives: for short-term prediction w used different data collection windows (6-72 hours) to dynamically predict the risk of AKI occurrence within subsequent specific time windows (12-72 hours); for long-term prediction we used data from specific time points after admission (12-72 hours) to predict the occurrence of AKI during the entire hospitalization.

RESULTS

The model showed robust performance across all prediction tasks (area under the receiver operating characteristic curve range, 0.80-0.94). Analysis of prediction lead time showed that the model could provide early warnings: the median lead times for predicting AKI occurrence within 12, 24, 48, and 72 hours were 9.57, 12.89, 18.77, and 27.25 hours, respectively. Feature importance analysis revealed that urine output, Sequential Organ Failure Assessment (SOFA) score, and systolic blood pressure played dominant roles in short-term prediction, whereas troponin T and history of cardiovascular surgery were more important in long-term prediction.

CONCLUSIONS

The long short-term memory-based model proposed in this study captures dynamic physiological changes in HF patients and provides dynamic risk assessments for AKI with sufficient lead time.