Using explainable machine learning and fitbit data to investigate predictors of adolescent obesity

Abstract

Sociodemographic and lifestyle factors (sleep, physical activity, and sedentary behavior) may predict obesity risk in early adolescence; a critical period during the life course. Analyzing data from 2971 participants (M = 11.94, SD = 0.64 years) wearing Fitbit Charge HR 2 devices in the Adolescent Brain Cognitive Development (ABCD) Study, glass box machine learning models identified obesity predictors from Fitbit-derived measures of sleep, cardiovascular fitness, and sociodemographic status. Key predictors of obesity include identifying as Non-White race, low household income, later bedtime, short sleep duration, variable sleep timing, low daily step counts, and high heart rates (AUC_Mean = 0.726). Findings highlight the importance of inadequate sleep, physical inactivity, and socioeconomic disparities, for obesity risk. Results also show the clinical applicability of wearables for continuous monitoring of sleep and cardiovascular fitness in adolescents. Identifying the tipping points in the predictors of obesity risk can inform interventions and treatment strategies to reduce obesity rates in adolescents.

Figure 1

This figure illustrates the iterative process of building an Explainable Boosting Machine (EBM) for binary classification. Starting with a basic model, the EBM sequentially updates its predictions by cycling through each feature to learn its unique contribution via gradient boosting. This round-robin approach ensures gradual refinement, where each feature's effect is independently modeled and then aggregated to form the final predictive model. Interactions between pairs of features are also explored and integrated, enhancing the model's accuracy. The process emphasizes continuous learning from residuals, leading to a highly interpretable and precise predictive model that offers clear insights into the influence of individual features and their interactions on the outcome. P is the probability of the target variable belonging to the positive class (obesity), β0 is the intercept; fi (xi) represents the smooth function for the i-th feature; xi is the i-th feature; n is the total number of features. Figures are just for illustration, generated using the seaborn (Version 0.11, URL: https://seaborn.pydata.org/generated/seaborn.heatmap.html) Python (Version 3.8.5) package.

Figure 2

Body mass index distribution for the complete ABCD Study sample at the Year 2 assessment (N = 7552) and for the sub-sample of participants in the analytical sample with Fitbit and weight data (n = 2971) based on the CDC guidelines for BMI percentiles (Underweight: less than the 5th percentile; Healthy Weight: 5th percentile to less than the 85th percentile; Overweight: 85th percentile to less than the 95th percentile; Obesity: 95th percentile or greater).

Figure 3

Mean absolute contribution score of each term (feature or interaction) in the best performing EBM model trained to predict obesity in young adolescents (n = 2971). The top 20 features are sorted by their contribution (C_i), where C_i represents the contribution score for the i-th feature to the prediction. These scores are expressed in terms of log odds, derived from a combination of univariate models for each feature, considering their individual contributions as well as interactions with other features.

Figure 4

Interaction terms in the EBM model: two-way interaction between the Sleeping heart rate and White race (A), Sleeping heart rate and food access (B), White race and Resting heart rate (C), Black race and Sleeping heart rate (D) and the two-way interaction between the Sleeping heart rate and Household income (E) in the prediction of obesity. The color of the heatmap represents the direction of the effect (red: values associated with higher prediction score and increased risk, green: values associated with decreased risk). Figures were generated using the seaborn (Version 0.11, URL: https://seaborn.pydata.org/generated/seaborn.heatmap.html) Python (Version 3.8.5) package.

Figure 5

Contribution to the prediction (C) (blue line) for the predictors with the highest contribution in the best performing EBM: (A) Average Sleeping heart rate—weekdays (HR), (B) Average Resting heart rate—weekdays (HR). (C) Total step count—weekends, D. Total step count—weekdays, (E) Sleep duration in minutes—weekends, (F) Sleep duration in minutes—weekdays. (G) Bedtime in hours—weekends. The grey lines represent the standard error. These graphs showcase the nuanced effect of each variable on obesity prediction, highlighting zones where their influence shifts from being protective to risky, or vice versa, allowing us to identify tipping points.