Objective: AI algorithms can potentially identify postprandial glucose responses (PPGR) directly from continuous glucose monitoring (CGM) data, reducing the burden of manual meal logging and PPGR interpretation. However, no standard AI approach for this task currently exists.
Methods: We developed a wavelet AI algorithm to identify PPGRs using a public CGM dataset (Hall'18) from individuals without a prior diagnosis of diabetes (Fig 1). The dataset also included clinic-recorded meal timings on a subset of days. Briefly, the algorithm works by computationally finding the best match for a template shape (the wavelet) in a time series. The algorithm was trained using a 5-fold cross-validation approach, and evaluated using two measures: (i) the time difference between true and predicted PPGR start times, and (ii) the ratio of predicted to true 2-hr area under the PPGR curve (AUC).
Results: We tested our algorithm on 19 participants with 2±1 days of both CGM data and clinic-recorded meal timings. The wavelet algorithm predicted start time of PPGR to within a median 15 [10,26] minutes of true meal time. The median ratio of predicted to true 2-hr AUC PPGR was 1.01 [1.00,1.02], indicating the wavelet algorithm’s ability to faithfully capture the PPGR.
Conclusion: A wavelet AI algorithm accurately identified PPGRs from CGM data, providing a novel automated approach to evaluate the effect of different foods on an individual’s glycemia.
W. Lim: None. S. Barua: None.
Abramson-Stires award from NYUGSOM Department of Medicine
