Interpretable Machine Learning for Predicting Suboptimal 12-Month Growth Response to Recombinant Human Growth Hormone in Children with Idiopathic Short Stature: A Dual-Center External Validation Study.
An interpretable machine learning model can predict which children with idiopathic short stature may not respond optimally to growth hormone therapy, potentially aiding in personalized treatment planning.
Where it sits
this study against the rest of the hgh (somatropin) corpusSummary and findings
This study aimed to develop and validate a machine learning model to predict suboptimal growth response to recombinant human growth hormone (rhGH) therapy in children with idiopathic short stature (ISS). The model was developed using data from 901 children and validated in an independent cohort of 51 children. The primary outcome was defined as a height gain of less than 0.5 standard deviation score after 12 months of treatment.
Abstract
Background/Objectives: Individual responses to recombinant human growth hormone (rhGH) therapy in children with idiopathic short stature (ISS) vary substantially, limiting pretreatment decision-making. This study aimed to develop and externally validate an interpretable machine learning model for predicting suboptimal 12-month growth response to rhGH therapy. Methods: In this retrospective dual-center study, 901 children from Center 1 were used for model development and internal testing, and 51 children from Center 2 formed an independent external validation cohort. Routinely collected baseline demographic, laboratory, hormonal, radiographic, and family-history variables were used to develop multiple machine learning models. A soft-voting ensemble classifier was constructed and interpreted using SHapley Additive exPlanations (SHAP). The primary outcome was suboptimal growth response, defined as failure to achieve a height gain of at least 0.5 standard deviation score after 12 months of treatment. Results: The optimized ensemble model showed strong discrimination in the internal test set, with an area under the receiver operating characteristic curve of 0.927, and maintained robust performance in the external validation cohort, with an AUC of 0.897. SHAP analysis identified luteinizing hormone, body mass index, TW3 RUS bone age, and insulin-like growth factor 1 as the leading contributors to predicted suboptimal-response risk. Conclusions: An interpretable ensemble machine learning model based on routinely available pretreatment data can predict suboptimal short-term rhGH response in children with ISS and may support individualized risk stratification in pediatric endocrine practice. Clinical trial registration was not required because this was a retrospective analysis.
Background
This paper addresses the variability in individual responses to rhGH therapy in children with ISS, which complicates treatment decisions. Prior research has indicated that predicting growth response can be challenging due to the multifactorial nature of growth disorders. The development of an interpretable machine learning model could enhance decision-making by providing personalized risk assessments based on baseline characteristics.
Methods
This retrospective dual-center study included 901 children for model development and 51 children for external validation. The study utilized routinely collected demographic, laboratory, hormonal, and radiographic data to develop multiple machine learning models. The primary outcome was defined as a height gain of less than 0.5 standard deviation score after 12 months of rhGH treatment.
Results
The optimized ensemble model demonstrated strong discrimination in the internal test set, with an area under the receiver operating characteristic curve (AUC) of 0.927. In the external validation cohort, the model maintained robust performance with an AUC of 0.897. SHAP analysis identified key predictors of suboptimal growth response, including luteinizing hormone and body mass index.
Interpretation
The findings suggest that the machine learning model can effectively predict suboptimal growth responses to rhGH therapy, which aligns with previous studies indicating the importance of individualized treatment approaches. However, the clinical significance of the model's predictions should be carefully evaluated, especially given the small size of the validation cohort and the retrospective nature of the study. These limitations may affect the generalizability of the results to broader clinical practice.
Key findings
- AUC of 0.927 in internal test set.
- AUC of 0.897 in external validation cohort.
- Leading contributors to predicted suboptimal-response risk included luteinizing hormone, body mass index, TW3 RUS bone age, and insulin-like growth factor 1.
Limitations
- retrospective study design
- small external validation cohort (n=51)
- relies on routinely collected data, which may introduce biases
- no clinical trial registration