138-LB: A Retrospective Algorithm to Identify Meal Intakes from CGM Time Series
The accurate assessment of dietary behavior in clinical trials, including number and size of meal intakes, is essential to evaluate participants’ metabolic health which, in turn, might reflect an increased risk for developing diabetes. Traditional dietary assessments leverage handwritten/digital foo...
Saved in:
Published in | Diabetes (New York, N.Y.) Vol. 72; no. Supplement_1; p. 1 |
---|---|
Main Authors | , , , , , , , , , , , , , |
Format | Journal Article |
Language | English |
Published |
New York
American Diabetes Association
20.06.2023
|
Subjects | |
Online Access | Get full text |
Cover
Loading…
Summary: | The accurate assessment of dietary behavior in clinical trials, including number and size of meal intakes, is essential to evaluate participants’ metabolic health which, in turn, might reflect an increased risk for developing diabetes. Traditional dietary assessments leverage handwritten/digital food logs, which are error-prone and time-consuming for both users and investigators. Instead, glucose timeseries unobtrusively collected by CGM could provide objective measures of meals’ consumption. Here we present a new algorithm to retrospectively identify meal intakes using CGM data only.
In the Geriatric Anorexia Study (NCT04858932), 50 healthy individuals (26 females, mean±SD age: 72.26±5.02 years, BMI: 24.96±3.33 kg/m2) were monitored in free-living conditions for 2 weeks with the Abbott FreeStyle Libre Pro, and recorded meal intakes using the Renpho Smart Food Scale. After removal of days with <2 meals recorded, a total of 1132 meals over 365 days were available for analysis. To identify meal intakes, the proposed algorithm scans CGM data, identifies local maxima, and measures their prominence, i.e., the relative height with respect to a neighborhood window of length L. Peaks with prominence higher than Pmin are labeled as meals. Subject data was divided into a ~70% training set (n=31, 775 meals), where L and Pmin are tuned, and ~30% test set (n=12, 357 meals), where the algorithm is evaluated in terms of absolute percent error (APE) between number of recorded and detected meals.
Across subjects, mean±sd daily APE is 25.1±9.70%. Similar values hold when considering only subjects with <3.5 meals/day (n=6, 144 meals, APE=24.5±13.2%) and ≥3.5 meals/day (n=6, 213 meals, APE=25.5±6.93%), demonstrating the reliability of the algorithm to identify the number of meals in subjects with different dietary behaviors.
These results evidence the suitability of CGM for the automatic quantification of meal intakes. Further assessment of the algorithm should be performed using more controlled meal intake recordings.
Disclosure
N. Camerlingo: Employee; Pfizer Inc. A. Messere: None. M. Santamaria: None. C. Demanuele: None. D. Caouette: Employee; Pfizer Inc. K. C. Thomas: None. N. Shaafi kabiri: None. D. Psaltos: Employee; Pfizer Inc. F. Karahanoglu: None. S. Khan: None. I. Messina: Research Support; Pfizer Inc. M. Wicker: None. M. Kelly: Research Support; Pfizer Inc. H. Zhang: Employee; Pfizer Inc. |
---|---|
ISSN: | 0012-1797 1939-327X |
DOI: | 10.2337/db23-138-LB |