INTRODUCTION: Clinical guidelines stress early, effective glycemic control for obese patients with type 2 diabetes (T2D), but real-world treatment patterns lag behind guidelines, possibly contributing to adverse outcomes.

OBJECTIVES: To identify longitudinal A1c level trends among obese patients with T2D prescribed U-100 basal-only regimens over a 3-year period with an unsupervised machine learning algorithm

METHODS: Adult obese patients with T2D prescribed U-100 basal-only regimens (first claim = index date) were identified in the Veterans Health Administration database (OCT2013-SEP2018). Included patients had continuous enrollment for ≥6 months pre-index and ≥3 years post-index. Patient characteristics and treatment patterns were evaluated descriptively. A1c was measured at 6-month intervals over a 3-year period; patients with ≥4 A1c values were clustered. An unsupervised longitudinal trajectory clustering method was implemented to examine 24 features of A1c trajectory, followed by feature reduction using factor analysis. K-means clustering was used to cluster patients by A1c trajectory within dynamic (fluctuating) and stable subgroups.

RESULTS: Of 60,198 total patients, 51,327 had ≥4 post-index A1c values. Patients were predominantly white males aged ≥65 years. Identified A1c longitudinal trajectory patient clusters included dynamic-descending: low (N=6,616, 11.0%) and high (N=3,443, 5.7%); dynamic-ascending: low (N=5,853, 9.7%%) and ascending high (N=3,235, 5.4%); static-high (N=11,778, 19.6%, persistent mean A1c 8.9%); and static-low (N=20,402, 33.9%, persistent mean A1c 7.1%). Small proportions of patients across high A1c clusters had treatment intensification (0.6%-2.0%) and 3-year mean insulin total daily doses >200 units (1.2%-3.8%). Observable adherence was poorest among the dynamic-ascending high cluster (3-year mean proportion of days covered [PDC]: 0.72; adherent patients [PDC >0.80]: 46.6%.

CONCLUSIONS: This study used machine learning to identify over 54% of obese T2D patients prescribed basal-only regimens with poor A1c control over 3 years. Results warrant continued research exploring the clinical drivers of temporal A1c trends within patient clusters with poor glycemic control.