(P20-016-25) Unsupervised Machine Learning to Identify Obesity Phenotypes in Brazilian Women

Objectives: This exploratory study aimed to develop a strategy based on artificial intelligence (AI) to identify distinct obesity phenotypes that wouldn’t be identified through BMI.

Methods: Data from 446 non-athlete Brazilian women (20-59y) with BMI from 18,5 to 40 kg/m² were analyzed. Data were extracted from the baseline datasets of 8 studies conducted by Immunometabolism Research Group. The study population included mainly individuals with no reported major diseases. Body composition data by electrical bioimpedance, weight, height, waist circumference (WC), and hip circumferences were used as variables and for index calculations. K-Means and Hierarchical Clustering Machine Learning algorithms were applied to identify obesity phenotypes, with different combinations of 22 variables as data input. Parametric or non-parametric tests were carried out to characterize the resulting clusters.

Results: Clustering models were set to obtain 4 clusters. The final model was achieved with 7 variables log transformed as data input for the k-means clustering algorithm. The resulting clusters were named: Normal Adiposity, Moderate Adiposity, Risk Adiposity, and Elevated Risk Adiposity. All 4 clusters presented anthropometric and body composition data statistically different in the post-hoc analysis. The Normal Adiposity cluster included all women with normal BMI and the lowest anthropometric measurements. Moderate Adiposity cluster included women with normal and overweight BMI. Women classified as lean in the Moderate Adiposity cluster presented higher anthropometric measurements and body fat values than those in the Normal Adiposity cluster while having similar fat-free mass. The Risk Adiposity and Elevated Risk Adiposity clusters included women with the highest BMI (overweight and obesity I classes in the Risk Adiposity cluster and obesity I and II classes in the Elevated Risk Adiposity cluster), body fat, WC, and WC indexes.

Conclusions: The K-Means method clustered 4 adiposity phenotypes using 7 current nutritional anthropometric measurements that distinguish obesity phenotypes regardless of BMI classification. Thus, using AI may be a promising tool to better identify people with higher metabolic risk related to adiposity accumulation.

Funding Sources: FAPEMIG, CAPES and CNPq.