Growth Hormone Treatment Response and Machine Learning-Based Prediction in Idiopathic GHD and ISS: Analysis of the Korean LG Growth Study.
Machine learning approaches provided limited improvements in predicting growth response to rhGH in iGHD, with no clear benefits in ISS patients.
Where it sits
this study against the rest of the hgh (somatropin) corpusSummary and findings
This study evaluated the prediction of growth response to recombinant human growth hormone (rhGH) therapy in children with idiopathic growth hormone deficiency (iGHD) and idiopathic short stature (ISS). A total of 2215 children were analyzed, with primary outcomes being changes in height SDS (ΔHSDS) at 1 and 2 years and mid-parental height (MPH) attainment. Machine learning models were compared to conventional statistical methods for predictive accuracy.
Abstract
<h4>Objective</h4>Individual responses to recombinant human growth hormone (rhGH) therapy vary widely among children with idiopathic growth hormone deficiency (iGHD) and idiopathic short stature (ISS), making accurate prediction of treatment outcomes clinically important. This study aimed to develop and compare machine learning (ML)-based and conventional statistical models to predict short-term growth response and mid-parental height (MPH) attainment following rhGH therapy in iGHD and ISS patients.<h4>Design</h4>Retrospective observational cohort study using a nationwide, real-world registry.<h4>Patients</h4>A total of 2215 children (1877 with iGHD and 338 with ISS) treated with rhGH were identified from the Korean LG Growth Study database. All included patients had at least 1 year of follow-up with available clinical data.<h4>Measurements</h4>Primary outcomes were 1- and 2-year changes in height SDS (ΔHSDS) and achievement of MPH SDS. Predictive models included multiple linear regression, logistic regression, Random Forest, eXtreme Gradient Boosting, and Elastic Net. Model performance was evaluated using R², error metrics and area under the receiver operating characteristic curve. Model interpretability was assessed using SHAP values.<h4>Results</h4>In the iGHD group, ensemble ML models modestly outperformed linear regression for predicting 1-year ΔHSDS (R² ≈ 0.19 vs. 0.16), but predictive performance declined at 2 years across all models. Prediction of MPH attainment showed high specificity but very low sensitivity at 1 year, with no clear advantage of ML over logistic regression. In ISS patients, all models demonstrated poor predictive performance for both ΔHSDS and MPH attainment, reflecting substantial clinical heterogeneity.<h4>Conclusions</h4>ML approaches provided limited but clinically meaningful improvements in predicting short-term growth response in iGHD, while offering no clear benefit in ISS. These findings highlight both the potential and limitations of ML models based solely on routine clinical variables and underscore the need for integrating multimodal data to improve growth prediction, particularly in heterogeneous ISS populations.
Background
The variability in responses to rhGH therapy among children with iGHD and ISS necessitates effective prediction models for treatment outcomes. Prior studies have indicated that individual growth responses can differ significantly, emphasizing the importance of accurate forecasting. This research aims to assess the efficacy of machine learning models compared to traditional statistical methods in predicting growth outcomes.
Methods
This retrospective observational cohort study utilized data from the Korean LG Growth Study, comprising 2215 children (1877 with iGHD and 338 with ISS) who received rhGH therapy. All patients had at least 1 year of follow-up data. Primary outcomes included changes in height SDS (ΔHSDS) at 1 and 2 years and attainment of MPH SDS, with predictive models evaluated using various statistical methods.
Results
In the iGHD group, ensemble machine learning models achieved an R² of approximately 0.19 for predicting 1-year ΔHSDS, compared to 0.16 for linear regression. However, predictive performance declined at 2 years across all models. For MPH attainment, high specificity was noted, but sensitivity was very low at 1 year, indicating limited predictive capability.
Interpretation
The findings suggest that while machine learning models showed modest improvements in predicting short-term growth response in iGHD, the effect sizes are not clinically significant. The poor predictive performance in ISS highlights the complexity of this population and the limitations of using routine clinical variables alone for growth prediction.
Key findings
- R² ≈ 0.19 for ensemble ML models predicting 1-year ΔHSDS in iGHD vs. 0.16 for linear regression.
- High specificity but very low sensitivity for predicting MPH attainment at 1 year in iGHD.
- All models showed poor predictive performance for ΔHSDS and MPH attainment in ISS patients.
Limitations
- Retrospective observational design may introduce confounding.
- Limited predictive performance in ISS patients.
- Small effect sizes not clinically meaningful.
- Single-site study may limit generalizability.