Current models for predicting 30-day readmission risk among people with diabetes vary in performance. We previously published the Diabetes Early Readmission Risk Indicator (DERRITM) , a logistic regression (LR) model with modest predictive performance (C-statistic 0.69) . The current study aims to develop a more accurate model using deep learning (DL) on electronic health record (EHR) data. We analyzed electronically abstracted data from 36,563 patients with diabetes and at least 1 hospitalization at an urban, academic medical center between 7/1/2010-12/31/2020. One hospitalization per patient was randomly selected for analysis. A DL, long short-term memory Recurrent Neural Network (RNN) was developed and compared to traditional models: LR, and Multi-layer Perceptron (MLP) . Models were developed using demographics, vitals, diagnostic and procedure codes, medications, laboratory tests, and administrative data as defined by the PCORnet Common Data Model. A look-back time of 1 year before the index hospitalization was used. Data dimensionality was reduced to 2,500 features by Singular Value Decomposition. The RNN C-statistic is significantly greater than that of the other models (Figure) . The RNN model outperforms the DERRI and is based on more generalizable EHR data. Varying the compression and time window might improve performance. RNN models may outperform traditional models at predicting readmission risk.
A.Abdel hai: None. M.Weiner: None. A.Paranjape: None. A.Livshits: None. J.Brown: None. Z.Obradovic: None. D.J.Rubin: Research Support; AstraZeneca.
National Institutes of Health (R01DK122073)