Interpretable machine learning identifies predictors of IGF-1 during growth hormone therapy in boys

Insulin-like growth factor-1 (IGF-1) is widely used to guide growth hormone (GH) therapy in children, yet the factors that determine IGF-1 levels during later stages of treatment remain incompletely understood. A new study published in Frontiers in Endocrinology applied interpretable machine learning approaches to predict IGF-1 standard deviation scores (SDS) at 3 and 12 months in GH-treated boys and to identify key predictors of IGF-1 dynamics during the maintenance phase and puberty. The research team, led by investigators at the Göteborg Pediatric Growth Research Center at the University of Gothenburg in Sweden, analyzed data from 72 boys with growth hormone deficiency or idiopathic short stature who participated in a previous Swedish clinical trial. The findings indicate that prediction of future IGF-1 levels is informed primarily by current IGF-1 SDS together with additional clinical variables related to treatment and pubertal development.

Correct dosing of growth hormone in children with or without classic GH deficiency can be challenging. Even after receiving treatment for several years, a significant percentage of patients fail to achieve target height based on parental heights, indicating that the effectiveness of treatment varies among individuals. GH dosing therefore requires individualized assessment and ongoing monitoring to optimize growth outcomes while minimizing risks. IGF-1-based dosing, where the GH dose is titrated to achieve a desired IGF-1 standard deviation score, has been proposed as a strategy for personalizing treatment. However, despite its widespread use in clinical practice, the predictors of IGF-1 levels during later stages of GH therapy remain incompletely understood, particularly during the maintenance phase and puberty where IGF-1 dynamics become more complex.

IGF-1 levels are influenced by multiple physiological and treatment-related factors. In GH-naïve individuals, age and sex are known determinants, with IGF-1 levels increasing throughout childhood and reaching a peak during puberty. In prepubertal children, IGF-1 is associated with height velocity, whereas during puberty, IGF-1 levels may remain elevated despite declining growth velocity. Following the initiation of GH therapy, an initial catch-up phase characterized by rapid growth is typically followed by a maintenance phase and subsequently a pubertal growth spurt. During the early phase of treatment, IGF-1 appears to be positively correlated with GH dose. In one of the few trials employing IGF-1-based dosing, higher GH doses were needed to maintain IGF-1 levels at the higher end of the normal range. Conversely, some studies did not show any significant association between IGF-1 z-scores and GH dose in their cohorts of prepubertal children on GH when adjusted for age, gender, race and insurance type. During maintenance and pubertal growth, the dose dependency is not as obvious, implying that other factors become more impactful at later treatment stages.

Most existing prediction models in pediatric GH therapy have focused on growth outcomes and have relied on multiple linear regression or empirical curve fitting approaches. More recently, machine learning methods have been applied to predict growth outcomes or treatment adherence, while prediction of IGF-1 levels has received less attention. Furthermore, while a model designed to predict growth response may maximize growth outcomes, it does not account for safety. This aspect of treatment can be addressed by focusing on IGF-1 as the predictive target. In non-GH-treated populations, both low and high IGF-1 levels are linked to mortality risk, underscoring the need for IGF-1 regulation in GH treatment. Machine learning approaches are well suited to identifying complex and potentially non-linear relationships between clinical variables, but in healthcare settings it is important that clinicians are able to follow the model’s reasoning and evaluate the resulting recommendation, which is possible in so-called white-box models.

In this study, the researchers applied three interpretable modeling approaches with good performance in small datasets: linear regression, Elastic Net regression, and explainable boosting machine (EBM). Linear regression was used as a baseline method representing standard statistical analysis. Elastic Net regression combines L1 and L2 regularization to improve model stability and variable selection, particularly when multiple predictors are correlated. EBM is a tree-based gradient-boosting method designed to balance flexibility and interpretability. These approaches allow detailed examination of how individual predictors contribute to IGF-1 levels in complex datasets. The study population originated from a larger cohort of 153 short-statured children who participated in the Catch-up growth study, of whom 98 children who achieved catch-up growth were included in a subsequent randomized, prospective, open-label clinical trial conducted in Sweden. Of the 98 participants, the present analysis was restricted to the 72 boys; the 26 girls were excluded because sex-related differences in hormonal regulation and growth patterns precluded combined modeling. Twenty-two boys had GH deficiency, defined as a peak GH concentration below 10 µg/L on both the arginine-insulin tolerance test and a 24-hour spontaneous GH profile, whereas the remaining 50 had idiopathic short stature.

The selected variables were based on previously identified predictors of growth and were categorized into three groups: early-life variables, variables related to the initiation of GH treatment, and variables corresponding to the time of prediction. Early growth data were obtained from child welfare centers, and additional clinical data including height, weight, testicular volume, IGF-1 levels, IGF-binding protein-3, and GH dose were collected at least annually and every three months during active participation in the Maintenance study. After applying the prespecified eligibility criteria and requiring complete data for the outcome and the common set of 33 candidate predictors, 62 boys contributed 165 observation pairs to the 3-month analysis and 65 boys contributed 306 observation pairs to the 12-month analysis. Hormone analyses were conducted at the Gothenburg Pediatric Growth Research Centre Laboratory, with estradiol and testosterone analyzed using validated high-sensitivity gas chromatography-tandem mass spectrometry and serum IGF-1 and IGFBP3 concentrations analyzed using a specific radioimmunoassay.

Model performance was evaluated using nested cross-validation with patient-level grouping. Participants were assigned to five reproducible outer folds, with all observations from a participant assigned to the same outer fold. In each outer iteration, models were developed using four folds and evaluated on the remaining fold, so that every eligible observation received one out-of-fold prediction from a model that had not been trained or tuned using that participant. Hyperparameter selection was performed only within the corresponding outer training data using three-fold patient-grouped inner cross-validation. Patient-balanced performance estimates were calculated from the pooled out-of-fold predictions, with each participant contributing equal total weight irrespective of the number of observations. Uncertainty was quantified using 2,000 patient-cluster bootstrap resamples, and performance estimates were reported with 95% confidence intervals.

The results showed that Elastic Net had the highest point estimates at 3 months, with an R-squared of 0.57 compared with 0.50 for full linear regression and 0.52 for EBM, but confidence intervals did not establish a difference between the primary models. No clear differences between models were found at 12 months, where R-squared values ranged from 0.41 to 0.45. As a reference, a persistence model using current IGF-1 SDS as the prediction of future IGF-1 SDS showed lower predictive performance than the multivariable models at both prediction intervals. Elastic Net identified five consistently selected predictive features at both prediction horizons, defined as predictors with non-zero coefficients across all five outer cross-validation models. For the 3-month prediction, these were IGF-1 SDS at prediction, IGF-1 increase, GH dose, testicular volume, and age at GH treatment start. For the 12-month prediction, the consistently selected features were IGF-1 SDS at prediction, testosterone decrease, testicular volume, age at prediction, and IGF-1/IGFBP-3 increase. IGF-1 SDS at prediction showed the largest median coefficient at both prediction horizons. In the EBM analysis, IGF-1 SDS at prediction showed the highest median main-effect importance at both prediction horizons.

The comparatively lower performance of ordinary linear regression may partly reflect the limitations of applying an unregularized linear model in a dataset with many potentially correlated predictors. Elastic Net may improve predictive performance in this setting by applying regularization and variable selection, thereby reducing the risk of overfitting. EBM provided a complementary assessment of predictive feature importance, although its greater flexibility did not result in clearly improved predictive performance compared with Elastic Net. GH dose received a non-zero Elastic Net coefficient in all five outer-fold models for the 3-month prediction and was also among the features with the highest median importance in the EBM analysis, supporting its relevance for short-term prediction of IGF-1 SDS. This is consistent with previous studies showing an association between GH dose and IGF-1 levels, although the strength of this relationship may vary across treatment phases and patient populations. Puberty-related variables contributed particularly to the 12-month prediction, with testicular volume consistently selected by Elastic Net at both prediction horizons, while age at prediction and testosterone decrease also received non-zero coefficients in all five outer-fold models at 12 months. These findings support the relevance of pubertal development for IGF-1 dynamics during longer prediction intervals, though interpretation of IGF-1 SDS during puberty is challenging, as pubertal development and sex steroid levels are not fully accounted for by age-based reference values.

Despite the relatively small cohort, the prospective longitudinal design, repeated measurements, and limited missing data allowed investigation of IGF-1 dynamics over several years of treatment. However, the models were developed and evaluated within a single cohort, and the limited sample size and absence of external validation restrict the generalizability of these findings. Although the best-performing models explained approximately half of the observed variance in IGF-1 SDS, a substantial proportion remained unexplained, reflecting the complexity of IGF-1 dynamics during the maintenance phase and puberty. In current clinical practice, GH dose adjustments are not based solely on IGF-1 levels but are interpreted together with growth response, pubertal development, treatment adherence, and other clinical factors. A prediction error of approximately 0.6 to 0.7 SDS would therefore currently limit the use of these models as stand-alone tools for dose adjustment. The primary aim of the study was not to develop a dosing algorithm for clinical use but to evaluate whether IGF-1 levels during the maintenance phase and puberty can be predicted and to identify predictors of IGF-1 levels over time. The present findings demonstrate that interpretable modeling approaches can explain a substantial proportion of the observed variation in IGF-1 levels during a treatment phase where regulation is influenced by multiple interacting factors. Further validation in larger cohorts will be required before clinical implementation can be considered, but these results provide new insight into the prediction of IGF-1 levels during later phases of GH treatment and may support the future development of more individualized treatment strategies.

Subject of Research: Prediction of IGF-1 levels during growth hormone therapy in boys using interpretable machine learning

Article Title: Interpretable machine learning identifies predictors of IGF-1 during growth hormone therapy in boys

Article References: Ly, H.-J., Suvilehto, J. T., Skyman, B., Ankarberg-Lindgren, C., Fors, H., Havsol, J., & Dahlgren, J. (2026). Interpretable machine learning identifies predictors of IGF-1 during growth hormone therapy in boys. Frontiers in Endocrinology, 17, Article 1873245. https://doi.org/10.3389/fendo.2026.1873245

Image Credits: AI Generated

DOI: 10.3389/fendo.2026.1873245

Keywords: growth hormone therapy, IGF-1, machine learning, pediatric endocrinology, prediction model, personalized medicine, Elastic Net, explainable boosting machine, puberty, maintenance phase, cross-validation, boys