914-P: Predictors of ED and Inpatient Admissions after Hypo- and Hyperglycemic Events: Leveraging Remote Monitoring Data of People with Diabetes

Abstract

Background: Predictive models to identify people with diabetes mellitus (DM) at high-risk for future ED visits and inpatient admissions (IA) are an area of clinical interest. The models do not include self-monitoring blood glucose (SMBG) levels. Livongo, the leader in Applied Health Signals, includes cellular-enabled BG meter that allows instantaneous uploading of SMBG values into the cloud with millions of values in people with DM. Methods: Leveraging data collected via the Livongo program with medical and pharmacy claims, machine learning techniques were used to identify the 25 most predictive variables of severe hypo- and hyperglycemia (BG ≤54 mg/dL and ≥400 mg/dL, respectively) resulting in an ED or IA encounter within three days. Four models were constructed for DM (type 1 and type 2) and encounter (ED and IA). Participants had to be enrolled for at least 12 months with continuous medical benefits eligibility for 24 months. Variable categories included in modeling were demographics, comorbidities based on ICD-10 codes, prior HCU, new and current medications, 30-days SMBG patterns with mean BG levels, and months on program. Area under the curve (AUC) was used to assess model performance. Results: There were 7,633 people selected. They had a mean age of 54 years, 48% were female, and 11% had type 1 DM. In this group, 924 and 1,518 severe hypo- and hyperglycemic with ED or IA encounters occurred. Random forest models had the highest AUC with values greater than 98% and sensitivity and specificity above 93% and 99%, respectively. HCU variables were the most predictive variables in all 4 models. Mean 7-day BG level, 30-day count of BG checks, and before-breakfast checking were also highly predictive. Conclusions: SMBG variables are independent predictors of hypo- and hyperglycemia with ED and IA encounters. Real-time BG remote monitoring programs have the capability to identify people at high-risk of costly HCU and develop interventions to improve care. Disclosure W. Lu: Employee; Self; Livongo Health. R. James: Advisory Panel; Self; AstraZeneca. Employee; Self; Livongo Health. S.L. Painter: Employee; Self; Livongo Health. B. Shah: Employee; Self; Livongo Health.