Algorithm beat human CGM readouts
A recent T2D study reported an algorithm that outperformed human glucose monitoring interpretation—fueling debate over AI’s role in real‑time glucose decision support. (x.com)
The validation poster lists Tong Sheng, PhD; Linda Parks, MS, RN; David Kerr, MD; and Michael S. Greenfield, MD, and was produced by Glooko with involvement from Sansum Diabetes Research Institute. (glooko.com) Researchers applied Glooko’s described “best day” algorithm to continuous glucose monitor (CGM) records from 10 people with diabetes, each providing seven continuous days of data. (glooko.com) The algorithm’s day-by-day rankings (from “best” to “worst” glycemic control) were compared against assessments from 57 clinicians: 29 endocrinologists and 28 diabetes educators. (glooko.com) Clinicians’ self-reported CGM experience in the sample showed a median of 8 years for endocrinologists (IQR 3–10) and a median of 25 years for diabetes educators (IQR 10–40). (glooko.com) Validation used identification of the correct “best day” plus three similarity metrics — R2, weighted Cohen’s Kappa, and mean absolute error (MAE) — to quantify agreement between the algorithm and clinician rankings. (glooko.com) The poster shows dataset-level results with R2 values roughly between 0.46 and 0.98, weighted Kappa between about 0.68 and 0.98, and MAE spanning approximately 0.14 to 1.43 across sampled days. (glooko.com) Example dataset entries printed in the poster include Dataset 1 (R2 0.98, weighted Kappa 0.98, MAE 0.14), Dataset 4 (R2 0.67, Kappa 0.81, MAE 1.00), and Dataset 5 (R2 0.46, Kappa 0.68, MAE 1.43). (glooko.com) Glooko published the poster PDF on its website in June 2025 under the title “Validation of a novel algorithm for interpreting glycemic control from CGM data.” (glooko.com)
