Development and External Validation of a Multidimensional Deep Learning Model to Dynamically Predict Kidney Outcomes in IgA Nephropathy

Key Points A dynamic model predicts IgA nephropathy prognosis based on deep learning. Longitudinal clinical data and deep learning improve predictive accuracy and interpretability in GN. Background Accurately predicting kidney outcomes in IgA nephropathy is crucial for clinical decision making. Insufficient use of longitudinal data in previous studies has limited the accuracy and interpretability of prediction models for failing to reflect the chronic nature of IgA nephropathy. The aim of this study was to establish a multivariable dynamic deep learning model using comprehensive longitudinal data for the prediction of kidney outcomes in IgA nephropathy. Methods In this retrospective cohort study of 2056 patients with IgA nephropathy from 18 kidney centers, a total of 28,317 data points were collected by the sliding window method. Among them, 15,462 windows in a single center were randomly assigned to training (80%) and validation (20%) sets and 8797 windows in 18 kidney centers were assigned to an independent test set. Interpretable multivariable long short-term memory, a deep learning model, was implemented to predict kidney outcomes (kidney failure or 50% decline in kidney function) based on time-invariant variables measured at biopsy and time-variant variables measured during follow-up. Risk performance was evaluated using the Kaplan–Meier analysis and C-statistic. Trajectory analysis was performed to assess the various trends of clinical variables during follow-up. Results The model achieved a higher C-statistic (0.93; 95% confidence interval, 0.92 to 0.95) on the test set than the machine learning prediction model that we developed in a previous study using only baseline information (C-statistic, 0.84; 95% confidence interval, 0.80 to 0.88). The Kaplan–Meier analysis showed that groups with lower predicted risks from the full model survived longer than groups with higher risks. Time-variant variables demonstrated higher importance scores than time-invariant variables. Within time-variant variables, more recent measurements showed higher importance scores. Further interpretation showed that certain trajectory groups of time-variant variables such as serum creatinine and urine protein were associated with elevated risks of adverse outcomes. Conclusions In IgA nephropathy, a deep learning model can be used to accurately and dynamically predict kidney prognosis based on longitudinal data, and time-variant variables show strong ability to predict kidney outcomes.