Inclusion of a multi-ancestry polygenic risk score into a machine learning weight gain prediction model: an...

Obesity is a public health crisis with few effective preventative treatments limited by challenges in identifying who is at high risk of weight gain. To address this gap, our group has developed machine learning weight gain prediction models with modest predictive power. In this study, we aimed to build upon our prior work by incorporating a multi-ancestry polygenic risk score (MAPRS) into a machine learning model and hypothesized that its inclusion would improve weight gain prediction performance.

Adults aged 18-70 years with weight measurements 2 years apart were identified within the NIH All of Us Research Program. Participants with a history of cancer, bariatric surgery, or pregnancy were excluded. XGBoost machine learning models were developed using demographics, vitals, laboratory data, obesity-related comorbidities, and MAPRS. The primary outcome was total body weight (TBW) gain of ≥10% over 2 years. Performance was compared using the area under the curves (AUCs).

The cohort consisted of 40 205 individuals (mean [SD] age 51.1[13.5] years; 56.1% female; 48.5% White, non-Hispanic), 3806 (9.5%) of whom gained ≥10% TBW over 2 years. Model performance with and without MAPRS was similar, with AUCs of 0.705 (95% confidence interval [CI] 0.692-0.718) and 0.697 (0.683-0.710), respectively. Baseline age and MAPRS had the highest variable importance. The best net reclassification index was 0.7%, indicating marginal improvements in risk classification with MAPRS.

Despite the known hereditary components of obesity and the strong association between the MAPRS and weight gain, incorporating a MAPRS did not substantially enhance model performance.

Further optimization of integrating polygenic risk scores into machine learning weight gain prediction models is needed before they can be applied in the clinical setting.