Abstract: TH-PO892
A Novel Deep Learning Model Outperforms Cox Regression Model to Predict Renal and Cardiovascular Risk in Patients with Diabetic Kidney Disease
Session Information
- Diabetic Kidney Disease: Basic - I
November 07, 2019 | Location: Exhibit Hall, Walter E. Washington Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Diabetic Kidney Disease
- 601 Diabetic Kidney Disease: Basic
Authors
- Belur nagaraj, Sunil, University of Groningen, Groningen, Netherlands
- Pena, Michelle, University of Groningen, Groningen, Netherlands
- Hamidi, Habib, University of Michigan, Ann Arbor, Michigan, United States
- Ju, Wenjun, University of Michigan, Ann Arbor, Michigan, United States
- L Heerspink, Hiddo Jan, University of Groningen, Groningen, Netherlands
Group or Team Name
- BEAt-DKD consortium
Background
Predicting long-term risk in patients with type 2 diabetes and chronic kidney disease is important in clinical practice. We hypothesize that by using short-term dynamic changes in clinical characteristics, deep learning algorithm can accurately predict long-term renal and cardiovascular risk.
Methods
In total 3228 patients with type 2 diabetes and chronic kidney disease from two randomized controlled trials were used in this study: RENAAL ( = 1513), IDNT ( = 1715). We used a 2D convolutional neural network (CNN) to predict renal (doubling of serum creatinine and/or end-stage renal disease) and cardiovascular (CV; myocardial infarction, stroke and cardiovascular death) outcomes. We compared the prediction performance with a traditional Cox proportional hazard regression (Cox) model. Eighteen clinical characteristics from baseline until 6 months follow-up were used as predictors to train the model on RENAAL data. The model was then externally validated on the IDNT trial. The area under the receiver operator curve (AUC) was used to assess the performance of the CNN and Cox model in the IDNT trial.
Results
A total of 462 (27%) and 518 (30%) of patients in IDNT experienced a renal or CV outcome respectively during a median follow-up of 2.6 years. The AUC of the CNN model, including UACR, HbA1c, SBP, albumin and uric acid as important predictors, was significantly higher compared to the Cox regression model (figure) and obtained the state-of-the-art performance to predict the long-term renal and CV outcomes.
Conclusion
Using 6-month short-term dynamic changes in clinical characteristics, a deep learning algorithm identifies patterns to accurately predict long-term renal and CV risk. The proposed method offers the potential to create accurate and automated risk predictions models to identify high-risk patients who could benefit from intensified therapy.
The performance comparison (mean AUC ± standard deviation) of CNN and Cox model for the prediction of renal and cardiovascular risks.