A Wavelet AI Algorithm to Automatically Identify Postprandial Glucose Responses From Continuous Glucose Monitoring Profiles in People Without Diabetes

Background: Automated identification of postprandial glucose responses (PPGR) from continuous glucose monitoring (CGM) profiles may detect early dysglycemia in people without diabetes. However, no standard approach for this task currently exists. Methods: We developed a wavelet transform-based AI algorithm to identify PPGRs using only CGM data. The algorithm was evaluated on a public CGM dataset of 25 normoglycemic adults and three independent validation cohorts (n = 65 total) with a mix of normoglycemia and prediabetes. Performance metrics included mealtime prediction error and total and incremental areas under the PPGR curve (tAUC and iAUC, respectively). Associations between AI-derived PPGR parameters and clinical markers such as HbA1c and fasting glucose were also examined. Results: In the public dataset, 25 participants (age 40 ± 14 years, BMI 26 ± 6 kg/m 2 , HbA1c 5.4 ± 0.4%) provided 3 ± 1 days of paired CGM data and ground-truth mealtimes. The algorithm predicted PPGR start time with a median error of 10 [IQR: 4, 19] minutes relative to ground-truth mealtimes. Postprandial glucose response parameters including tAUC and iAUC derived using ground-truth mealtimes versus AI-predicted mealtime were similar (all P > .1), indicating the algorithm faithfully captured PPGR characteristics. In adjusted analysis, AI-derived PPGR iAUC was independently associated with laboratory markers including HbA1c (β = 0.57 [95% CI: 0.19, 0.95], P = .006) and fasting glucose (β = 0.52 [95% CI: 0.12, 0.92], P = .013). Algorithm performance remained consistent across the three validation cohorts. Conclusions: The wavelet AI algorithm accurately identified PPGRs from CGM data in people without diabetes, offering a novel automated approach to monitor early signs of postprandial dysglycemia in this population.