Machine Learning Prediction of Growth Hormone Response in Children Non-Growth Hormone-Deficient Short Stature

Height gain under recombinant human growth hormone (rhGH) varies widely in children with short stature, making early, reliable response prediction essential for individualized care. Using routinely collected data from Bambino Gesù Children's Hospital (Rome, Italy) with up to five years of follow-up, we built and evaluated machine learning (ML) models for children with Idiopathic Short Stature (ISS) and Small for Gestational Age (SGA). Predictions were updated at baseline, end of the first treatment year, and end of the second treatment year, and categorized as poor, mild, or good responders. Results indicated that response classification was most accurate at the end of the first treatment year, consistent with clinical evidence that year one is the most informative checkpoint for long-term response. The most influential predictors were perinatal size, age at treatment initiation, mid-parental target height, baseline stature and growth trajectory, and initial rhGH dose. These results support identifying responders who may benefit from dose optimization, adherence reinforcement, or additional evaluation rather than simple continuation or discontinuation.