A machine-learning algorithm correctly identified 98.7% of meal events from continuous glucose monitoring data, with predicted 2-hour glucose AUCs highly correlated with actual responses (p<0.0001).
Does a machine-learning algorithm accurately identify meal events and postprandial glucose responses from continuous glucose monitoring data in inpatient participants?
A novel machine-learning algorithm can accurately identify meal events and estimate postprandial glucose responses directly from continuous glucose monitoring data.
Introduction and Objective: Poor long-term glycemic control increases the risk of complications such as cognitive decline and cardiovascular disease. Continuous glucose monitoring (CGM) enables the assessment of glycemic control by providing time-stamped glucose data. However, significant, rapid rises or dips in blood sugar can occur independent of eating because of metabolic, hormonal, and environmental factors, decreasing the accuracy of deciphering meal times over extended periods under free-living conditions. Methods: We developed a machine-learning algorithm—a Bidirectional Long Short-Term Memory network with an Autocorrelation Function—that accurately identifies meal events directly from CGM data (Dexcom G4 Platinum). The algorithm was trained and cross-validated on CGM data collected from 39 inpatient participants during inpatient ad libitum feeding trials (range 45%-75% kcal CHO). In total, the meal times of 1,548 meals were recorded by research nurses at the NIH Clinical Center. Results: Overall, the algorithm correctly identified 98.7% of meal events. Among the identified meals, 83% were detected within 10 minutes of the recorded meal time, 90% within 15 minutes, and 95% within 25 minutes. The 2-hour glucose AUCs of the predicted and actual postprandial glucose responses were highly correlated (r = 0.986 and 0.861 for absolute and incremental AUCs, respectively; p 0.0001). For predicted absolute AUCs, Bland-Altman analysis revealed no significant linear bias (−0.01 ± 0.009 mg/dL, p = 0.37) and a trivial mean bias (−0.3 ± 0.14 mg/dL, p = 0.037), with an RMSE of 2.78 mg/dL. Similarly, for predicted incremental AUCs, no significant linear bias (−0.02 ± 0.028 mg/dL, p = 0.41) or mean bias (−0.49 ± 0.42 mg/dL, p = 0.24) was observed, with an RMSE of 8.21 mg/dL. Conclusion: This algorithm can be used to assess dietary compliance in outpatient studies and to evaluate long-term glycemic control in patients consuming diets of 45-75% energy from carbohydrates under free-living conditions. Disclosure J. Guo: None. V. Darcey: None.
Guo et al. (Fri,) conducted a other in Glycemic control (n=39). Machine-learning algorithm for CGM data vs. Recorded meal times was evaluated on Correct identification of meal events. A machine-learning algorithm correctly identified 98.7% of meal events from continuous glucose monitoring data, with predicted 2-hour glucose AUCs highly correlated with actual responses (p<0.0001).
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: