Identification of weight loss predictors using machine learning approaches in adolescents with obesity

Despite the effectiveness of lifestyle multidisciplinary (LMD) weight loss interventions in pediatric obesity, outcomes remain variable between individuals. Machine learning (ML), capable of capturing complex relationships between variables, offer a promising avenue to better understand and predict this variability.

This study aimed to identify baseline predictors of 9-month LMD success in adolescents with obesity. A pooled database of 471 adolescents (42.9% males, 13.6 ± 1.4 years, BMI z-score 3.06 ± 0.51) from 21 clinical trials was used. Five ML models (Random Forest (RFC), XGBoost, LightGBM, Logistic Regression, and Support Vector Machine) were developed and evaluated to predict treatment response using both binary and multiclass classification RESULTS: Mean BMI z-score reduction at 9 months was -0.53 ± 0.34. For binary classification, RFC achieved the highest performance (accuracy = 0.97; AUC-ROC = 0.97). For multiclass prediction, XGBoost performed best (accuracy = 0.80; AUC-ROC = 0.90).

The most influential predictors consistently belonged to two main domains: (1) functional aptitudes, (e.g. forced vital capacity, maximal aerobic power, vertical jump height, resting VO2) and (2) dietary profiles (e.g. total energy intake, protein intake, hunger/fullness scores). Metabolic markers (eg. Insulin, cholesterol) also contributed to multiclass prediction. ML shows promise for personalizing adolescent weight loss strategies and deserves further study.

While many fields of healthcare utilize artificial intelligence-based (AI) methods to achieve personalized and patient-centered care, these methods remain underexplored in the field of pediatric obesity. Identifying baseline predictors of lifestyle multidisciplinary weight loss interventions (LMD) would further improve the effectiveness of such interventions, currently hindered by inter-individual variability in response to treatment. Machine learning models were successful in predicting LMD's success in adolescents with obesity with high accuracy, and uncovered a novel, promising dietary factors. Machine learning is a promising tool for personalization of weight management strategies in the field of pediatric obesity.