Advancing precision in GH therapy: AI models for predicting IGF-1

Helena-Jamin Ly; Juulia Suvilehto; Beatrice Skyman; Carina Ankarberg-Lindgren; Hans Fors; Jovanna Dahlgren

doi:10.1530/endoabs.110.P611

P611

Prev Next

Section Contents Cite

Endocrine Abstracts (2025) 110 P611 | DOI: 10.1530/endoabs.110.P611

ECEESPE2025 Poster Presentations Growth Axis and Syndromes (91 abstracts)

Advancing precision in GH therapy: AI models for predicting IGF-1

Helena-Jamin Ly ¹ , Juulia Suvilehto ² , Beatrice Skyman ³ , Carina Ankarberg-Lindgren ¹ , Hans Fors ¹ & Jovanna Dahlgren ¹

Author affiliations

¹Institute of Clinical Sciences, Department of Pediatrics, Gothenburg, Sweden; ²AI Competence Center, Gothenburg, Sweden; ³Medical Physics and Biomedical Engineering, Gothenburg, Sweden

JOINT497

Background: International guidelines recommend dosing growth hormone (GH) based on body weight or body surface area. However, since GH regulates the production and release of IGF-I in the liver, using IGF-I as a biomarker presents a promising strategy for optimizing GH therapy dosing. Despite its potential, this approach is complicated by the inherent complexity and variability of IGF-I. AimTo optimize dosing strategies by showcasing the potential benefits of advanced machine learning techniques through a comparison of three different models and identifying key features for predicting IGF-1 SDS.

Methods: Data from 63 boys from previous clinical trial were included containing variables related to early growth, the start of GH treatment, and around the prediction time. 3-month predictions of IGF-1 SDS during maintenance and puberal growth phases on GH treatment were constructed using three different machine learning models; linear regression, symbolic regression and explainable boosting machine (EBM). Linear regression was used as a baseline representing classic statistical regression.

Results: Linear regression demonstrated poor performance with an R² of 0. 07, whereas the more advanced machine learning models—symbolic regression and EBM—achieved significantly better results, both with an R² of 0. 47. Mean absolute error for linear regression showed 0. 78 while both symbolic regression and EBM showed 0. 55. Key features identified by both models were baseline IGF-1 SDS, height velocity before GH treatment, weight and ΔIGF-1 at 1 year on treatment.


Machine Learning models	Predictors	R²	Mean Absolute Error	Mean Absolute Percentage Error
Linear Regression	33	0. 07	0. 78	0. 68
Symbolic Regression	7	0. 47	0. 55	0. 47
Explainable Boosting Machine	63	0. 47	0. 55	0. 32

Conclusion: Symbolic regression demonstrated clinical practicality by providing accurate predictions using only seven predictors, while EBM offered valuable interpretability and deeper insights into factors influencing IGF-1 SDS. Notably, EBM identified GH dose as an important feature, a relationship not detected by symbolic regression or linear regression, likely due to its complexity. Together, these models enhance individualized treatment strategies and advance precision medicine in GH therapy.