Machine Learning Identification of Multiple Predictively Equivalent Risk Models for Severe Hypoglycemia in Patients with Type 2 Diabetes

Abstract

Severe hypoglycemia (SH), hypoglycemia requiring medical assistance and either blood glucose ≤ 50 mg/dL or glucose administration, instills fear in patients treated for diabetes. While risk models for SH may categorize an individual’s risk, incomplete or unavailable (“missing”) data limit clinical application. To circumvent the “missing” data issue, we used machine learning to identify multiple predictively equivalent risk models for SH in patients with type 2 diabetes (T2DM). Data from the Action to Control Cardiovascular Risk in Diabetes (ACCORD) study (n=10,251) were analyzed. Over ACCORD follow-up (mean 4.7± SD 1.4 years), 721 incident SH events were observed. We examined 95 candidate risk factors for model construction. Multiple predictively equivalent risk models (n=194) for SH were induced using the SurvTIE* algorithm (an adaptation of TIE* algorithm to survival outcome) and constructed with Cox regression and time-varying covariates. For each risk model, the number of risk factors ranged from 18-23 (median = 19). Unbiased performance estimations of c-index 0.78±0.03 were obtained from repeated cross-validation. Table 1 shows several unique models with equivalent predictive ability. In patients with T2DM, multiple predictively equivalent risk models for SH can be identified, potentially personalizing model selection when data are “missing.” Disclosure S. Ma: None. P. Schreiner: None. R. Zmora: None. E.R. Seaquist: Advisory Panel; Self; Eli Lilly and Company. Consultant; Self; Eli Lilly and Company, Sanofi, Zucera, InfoMed, 360 consulting. Other Relationship; Self; Novo Nordisk Inc. L.S. Chow: Research Support; Self; Eli Lilly and Company, National Institute of Diabetes and Digestive and Kidney Diseases.