Applying Prediction Models and AI to Optimize Growth Hormone Therapy in Children with Short Stature

Abstract

ABSTRACT Background: Growth hormone (GH) therapy has revolutionized the management of short stature in children, offering improved growth outcomes for those with GH deficiency and other related conditions. This thesis focuses on optimizing GH therapy through predictive tools and advanced machine learning models, with an emphasis on individualizing care to enhance patient outcomes and cost-effectiveness. Studies summaries: Study 1 evaluated the Gothenburg prediction model, which accurately identified poor responders to GH therapy, reducing the likelihood of unnecessary treatment. In contrast to a Nordic study, where 28% of children with GH deficiency failed to achieve a sufficient growth response on treatment, 98% of patients selected using the Gothenburg model achieved adequate growth outcomes with GH therapy. Study 2 compared the Gothenburg and KIGS (Pfizer International Growth Database) prediction models, both of which demonstrated exceptional accuracy in predicting first-year growth (r = 0.99 for both models). This comparison highlighted the flexibility of clinical choice, as either model could effectively guide GH therapy planning. Insulin-like growth factor 1 (IGF-1) is a potential biomarker that reflects the body´s response to GH. Study 3 explored the complexities of IGF-1 interpretation during GH therapy, particularly in early puberty, where discrepancies between clinical pubertal characteristics and biochemical hormone levels led to overestimated IGF-1 standard deviation scores (SDS). This study underscored the need for improved tools to enhance IGF-1 interpretation for better clinical decision-making. Study 4 introduced machine learning approaches to predict IGF-1 SDS and optimize GH dosing strategies. Symbolic regression emerged as a powerful tool, providing accurate predictions with minimal input variables, while Explainable Boosting Machine (EBM) excelled in identifying complex feature relationships. These models not only predicted IGF-1 SDS with high accuracy (R² = 0.47 - 0.51) but also demonstrated clinical applicability by balancing precision with interpretability. Conclusion: This thesis establishes a strong foundation for incorporating predictive models into routine clinical practice, facilitating more personalized and effective GH therapies. By utilizing advanced technologies and addressing key clinical challenges, this work contributes to optimizing treatment strategies, improving patient outcomes, and advancing precision healthcare.