Background: This study aimed to investigate the ability of a deep learning model to predict the outcome of severe hypoglycemia (SH) using longitudinal electronic medical records data.
Methods: Structured data were extracted from the clinical data warehouse, including demographics, laboratory, medications, and diagnostic information. Between January 2010 and December 2012, 17,670 adult patients were enrolled; participants had type 2 diabetes, were aged over 30 years, and visited the diabetes centers of two hospitals. The primary outcome was SH, defined as diagnostic codes for hypoglycemia (E16.x, E11.63, E13.63, E14.63) in hospitalization or emergency room records. The performance of the longitudinal deep learning model (gated recurrent units) in predicting SH was quantified using the area under the receiver operating characteristic curve (AUROC) and evaluated in terms of net reclassification improvement (NRI) versus a traditional logistic regression model. Based on predicted risk of SH, we stratified patients into four risk groups: low (<1.0%), intermediate (1.0% to <2.0%), high (2.0% to <5.0%), and very high (≥5.0%).
Results: Of 17,670 patients, 528 patients developed SH events during the follow-up period (mean of 6.4 years). The performance of the longitudinal deep learning model was superior to the logistic model for predicting SH (AUROC of logistic model, 0.788 [95% CI 0.781-0.793] vs. longitudinal deep learning model, 0.858 [0.853-0.863]). The very high-risk group of SH categorized using the deep learning model had 73 times higher risk of SH than the low-risk group. The deep learning model improved case reclassification (NRI, 13.4%).
Conclusion: We found that the longitudinal deep learning model showed better performance and classification ability for predicting than the conventional logistic model.
G. Lee: None. S. Ko: None. Y. Ahn: None. E. Lee: None. S. Cha: None. D. Kim: None. J. Yun: None.
Korean Diabetes Association (2018-F to J-S.Y.)