Digital phenotyping of CGM engagement reveals distinct glycemic outcomes.

Benefits from continuous glucose monitors (CGMs) may depend on how devices are used over time-not only how often they are used. We linked one year of device-generated CGM wear data from 2,351 U.S. Veterans to electronic health records (EHRs) to characterize real-world usage during the first year after CGM initiation. To compare longitudinal use patterns in the presence of unsynchronized sensor replacement-related gaps and intermittent interruptions, we aligned daily wear streams using a complexity-adjusted, time-adaptive optimal transport (TAOT) distance and applied spectral clustering to identify data-driven usage phenotypes. To estimate the adjusted association between CGM usage phenotypes and clinical outcomes, we used a double/debiased machine learning framework to quantify 12-month changes in time in range (ΔTIR) and mean glucose (ΔMG). We identified three reproducible patterns of CGM usage: Consistent, Fluctuating, and Low engagement. Relative to Consistent wear, Fluctuating usage was associated with worse glycemic change (ΔTIR = -3.56%, 95% CI: -4.76 to -2.36; ΔMG = +7.12 mg/dL, 95% CI: 4.91 to 9.32), and Low engagement showed larger deterioration (ΔTIR = -7.00%, 95% CI: -10.80 to -3.14; ΔMG = +14.09 mg/dL, 95% CI: 6.52 to 21.67). Notably, 73.2% of Fluctuating users still met a common "adherence" threshold (≥80% days worn), indicating that simple coverage metrics can miss clinically relevant instability. The differences in glycemic change (ΔTIR and ΔMG) between the Consistent and Fluctuating groups were most pronounced among subgroups with more intensive diabetes management needs, such as insulin pump or glucagon users, suggesting that sustained CGM usage may be particularly important when clinical management is more complex. Beyond CGM and diabetes, this work provides a generalizable framework for characterizing longitudinal usage patterns of intermittently used digital health technologies and linking derived usage phenotypes to clinical outcomes. The approach can support more precise evaluation, monitoring, and intervention design for a wide range of real-world digital health tools.