Machine learning models outperform glucose management indicator in estimating HbA1c
Click Here to Manage Email Alerts
Machine learning models using continuous glucose monitor and demographic data were more accurate at estimating HbA1c for people with type 1 diabetes compared with the glucose management indicator, according to study data.
Using data from five cohorts of people with type 1 diabetes, researchers developed machine learning models with as much as 26% fewer HbA1c estimation errors compared with the glucose management indicator formula.
“The COVID-19 pandemic interrupted HbA1c measurements in many ongoing studies,” David Scheinker, PhD, clinical associate professor of pediatrics – endocrinology and diabetes at Stanford University School of Medicine and the founder and director of SURF Stanford Medicine, told Healio. “This work uses machine learning and data from many different types of patients collected by CGM to better estimate HbA1c. This method may allow studies in which HbA1c measurement was interrupted to use this approach to estimate the effects on their population. Data collected by CGM and analyzed with machine learning should significantly reduce the need for HbA1c, a lab value measured approximately every 3 months, with a much richer, closer to real-time picture of glucose management.”
Scheinker and colleagues collected data from four cohorts described in studies listed by the Type 1 Diabetes Exchange and a fifth cohort described in a study on lifestyle intervention for teenagers with type 1 diabetes. People with HbA1c values between 5.5% and 11.5% and whose self-identified race was white or Black were included. CGM, HbA1c and demographic data were collected. All HbA1c values were accompanied by at least 5 days of CGM recordings.
Using HbA1c as the response variable and CGM glucose statistics and demographics as the features, researcher created an L1-regularized regression (LASSO) and a random forest (RF) regression with two-way interactions between all features. A simple ordinary least squares regression of HbA1c on mean glucose and race (OLSmgr) was also developed. The findings were published in Journal of Diabetes and Its Complications.
The analysis included 4,212 HbA1c measurements from 1,182 participants. The RF model was the highest performing model that did not use prior HbA1c, with a mean error of 0.67 percentage points, 19% lower than the glucose management indicator’s average error of 0.83 percentage points (P < .001). The OLSmgr model had an 8% higher average error than the RF model. HbA1c estimates in the RF model were within 1 percentage point of the true HbA1c value for 87% of participants compared with 81% for the glucose management indicator, with the strongest performance for HbA1c values of 9% and 10%. In cohorts with a median HbA1c higher than 8%, both the RF and OLSmgr models estimated median HbA1c more accurately than the glucose management indicator.
Researchers paired 2,352 HbA1c measurements from 872 people with an additional HbA1c measurement from 70 days prior. LASSO was the best performing machine learning model, with an average error of 0.49 percentage points, which was 26% lower than the glucose management indicator’s average of 0.67 percentage points (P < .001). LASSO estimated HbA1c within 1 percentage point for 95% of participants compared with 89% for the glucose management indicator. LASSO performed strongest in those with HbA1c of 9% and 10% when measuring error within 1 percentage point.
Researchers noted the simpler OLSmgr model performed nearly as well as LASSO, with an average error 0.4% higher despite factoring in only mean glucose, race and prior HbA1c.
“The performance of glucose management indicator could be dramatically improved by accounting for only two additional features and without sacrificing interpretability,” the researchers wrote.
For more information:
David Scheinker, PhD, can be reached at dscheink@stanford.edu.
Collapse
Click Here to Manage Email Alerts