Beyond Time in Range: Hidden Statistical Challenges of Continuous Glucose Monitoring Data in Diabetes Drug Development

Integrating continuous glucose monitoring (CGM) systems as digital health technologies (DHT) has emerged as a transformative approach in anti-diabetic drug development. Data collected from CGM offer real-time insights into glycemic control that complement traditional measures like hemoglobin A1c. However, CGM use in the clinical trials presents complex statistical challenges that require comprehensive frameworks for data quality and traceability assessment and missing data management. While the Food and Drug Administration has provided guidance ¹ for the use of DHT in the clinical investigations, the unique statistical challenges posed by CGM data demand further attention.

Continuous glucose monitoring devices generate high-volume data—up to 1440 glucose measurements daily per participant. This data richness comes with analytical challenges in evaluating treatment effects in clinical trials. The transformation from raw sensor data (epoch-level) to regulatory endpoints such as Time in Range requires rigorous statistical strategies, which are often insufficiently addressed in current studies.

The multilayered structure of CGM data—from epoch-level to the summary level—presents data quality and traceability concerns. Epoch-level readings collected from the sensor every one or five minutes must undergo multiple derivation steps with additional patient and timing information before yielding summary-level endpoints. Each layer introduces potential sources of error that can propagate throughout the analysis. For example, five-minute CGM readings may contain irregular time intervals and duplicate data points due to factors such as time zone changes, daylight saving adjustments, or device malfunctions. If these irregularities are not similar across treatment arms in the clinical trials, bias can be introduced into treatment comparisons.

Continuous glucose monitoring systems generate continuous data streams where missingness can occur at various levels from individual readings to entire days without data.^2-4 Several imputation methods are explored for addressing the impact of missing data. ⁵ The selection of appropriate methods depends on factors such as the reason for missingness, the extent of missing observations at each data level, and the duration of the data gaps. The estimand framework provides a valuable foundation for addressing missing data for CGM-derived endpoints. ⁶

As CGM-derived endpoints gain prominence in drug approval decisions, we recommend several immediate actions to address challenges. First, clinical trial analysis plan should include minimum requirements for data completeness and prespecified rules for handling common data anomalies to ensure data quality. Second, CGM data quality metrics should be assessed to ensure similar patterns of missing data, sensor errors, and data anomalies across treatment arms, thereby supporting valid treatment comparisons. Third, comprehensive sensitivity analyses should be conducted to assess the robustness of conclusion under different missing data assumptions.

The promise of CGM in diabetes research is substantial, but realizing this potential requires immediate attention to statistical rigor and regulatory standards. Establishing robust statistical strategies for CGM data analysis is essential for scientific integrity while supporting continued innovations in diabetes treatment.