Prediction of Antipsychotic Associated Weight Gain in Children and Adolescents Taking Second Generation Antipsychotics: A Machine Learning Approach

OBJECTIVES: Second Generation Antipsychotics (SGA) are associated with serious cardiometabolic side effects in children and adolescents, especially antipsychotic-associated weight gain (AAWG). However, there is no study that has focused on real-time AAWG prediction among children and adolescents. The objective of this study is to develop a prognostic-based machine learning algorithm that would dynamically predict the real-time risk of AAWG.

METHODS: This study encompassed children and adolescents who were naïve SGA recipients included in the IQVIA Ambulatory EMR- US database between 2016 and 2021. The outcome was the relative change between the baseline BMI z-score and the last BMI z-score categorized as severe, moderate, and minor AAWG. These models included 6 supervised machine-learning algorithms, which were the multiclass Logistic Regression (MLR) model, the multiclass Classification and Regression Trees (CART) model, the multiclass Random Forest (MRF) model, the multiclass Vector Generalized Additive Model (VGAM) model, and the multiclass and the 2-stage binomial Extreme Gradient Boosting (Xgboost) models.

RESULTS: A total of 10,997 patients who met the eligibility criteria were identified. The proportions of patients who experienced minor, moderate, and severe weight gain were 64%, 10%, and 26% respectively. Of the 6 models developed, the multiclass Xgboost model exhibited the highest AUC of 0.921, and the highest sub-categorical sensitivity for the identification of patients at risk of minor (0.911) and severe (0.803) AAWG. The top 5 important features were BMI z-score slope, Baseline BMI z-score, SGA duration, duration between the last BMI z-score measure, and the predicted outcome and counts of BMI z-score at the follow-up period.

CONCLUSIONS: In conclusion, the dynamic and individual-level change of AAWG in children and adolescents with SGA treatment could be accurately predicted using the machine learning algorithm. The algorithm developed in the study can be applied in practice to guide personalized monitoring and timely interventions of AAWG.