546-P: Deep Learning Predictions of Diabetic Retinopathy Associated with Progression of Renal Disease in Type 1 Diabetes

Abstract

Background and Aims: We considered if deep learning (DL) models trained to predict diabetic retinopathy (DR) are also useful for prediction of renal disease in people with type 1 diabetes. Materials and Methods: We used 35000 retinal images and associated ground truth DR grades from the online Kaggle dataset. These were used to train a Resnet convolutional neural network (CNN) to predict the ground truth DR grades. The CNN was then applied to retinal images for each patient in the SDRNT1BIO available from the national screening program generating 10 DL outputs corresponding to DR gradings for both eyes. Longitudinal eGFR readings and risk factor data were available for the SDRNT1BIO through linkage to clinical records. Univariate associations of DL outputs with final eGFR at follow-up was tested via linear regression models. A penalised Bayesian approach using 10-fold cross-validation was used to select the most predictive DL features and assess prediction performance. Results: We found 7 DL outputs had a significant univariate association with follow-up eGFR adjusting for age, sex, duration of diabetes, baseline eGFR, and length of follow-up at p<10-4. 4 DL outputs remained significant on adjusting for the screening program manual DR grading of the image (p ranging from 1.62x10-8 to 9.9x10-5). For predicting achieved eGFR we found that DL outputs improved test log-likelihood by 22.2 natural log units (equivalent to a p-value of < 10-9) increasing the R2, from 0.64 to 0.65. Conclusion: This small study demonstrates that retinal images contain information for predicting eGFR decline and that deep learning can extract this information. Future work on a larger dataset using DL to predict renal disease progression directly is therefore warranted. Disclosure J.C. Mellor: None. A.J. Storkey: None. H. Colhoun: Advisory Panel; Self; Eli Lilly and Company, Novartis Pharmaceuticals Corporation, Sanofi-Aventis. Research Support; Self; AstraZeneca, Boehringer Ingelheim International GmbH, Eli Lilly and Company, Novo Nordisk A/S, Pfizer Inc., Regeneron Pharmaceuticals, Roche Pharma, Sanofi-Aventis. Speaker's Bureau; Self; Eli Lilly and Company, Regeneron Pharmaceuticals, Sanofi. Stock/Shareholder; Self; Bayer AG, Roche Pharma. P.M. McKeigue: Advisory Panel; Spouse/Partner; Eli Lilly and Company, Novartis Pharmaceuticals Corporation, Sanofi. Research Support; Spouse/Partner; AstraZeneca, Boehringer Ingelheim International GmbH, Eli Lilly and Company, Novo Nordisk A/S, Pfizer Inc., Regeneron Pharmaceuticals, Roche Pharma, Sanofi-Aventis. Stock/Shareholder; Spouse/Partner; Bayer AG, Roche Pharma.