Explainable predictive models of short stature and exploration of related environmental growth factors: a case-control study

Abstract

Background: Short stature is a prevalent pediatric endocrine disorder for which early detection and prediction are pivotal for improving treatment outcomes. However, existing diagnostic criteria often lack the necessary sensitivity and specificity because of the complex etiology of the disorder. Hence, this study aims to employ machine learning techniques to develop an interpretable predictive model for normal-variant short stature and to explore how growth environments influence its development.

Methods: We conducted a case‒control study including 100 patients with normal-variant short stature who were age-matched with 200 normal controls from the Endocrinology Department of Nanjing Children's Hospital from April to September 2021. Parental surveys were conducted to gather information on the children involved. We assessed 33 readily accessible medical characteristics and utilized conditional logistic regression to explore how growth environments influence the onset of normal-variant short stature. Additionally, we evaluated the performance of the nine machine learning algorithms to determine the optimal model. The Shapley additive explanation (SHAP) method was subsequently employed to prioritize factor importance and refine the final model.

Results: In the multivariate logistic regression analysis, children's weight (OR = 0.92, 95% CI: 0.86, 0.99), maternal height (OR = 0.79, 95% CI: 0.72, 0.87), paternal height (OR = 0.83, 95% CI: 0.75, 0.91), sufficient nighttime sleep duration (OR = 0.48, 95% CI: 0.26, 0.89), and outdoor activity time exceeding three hours (OR = 0.02, 95% CI: 0.00, 0.66) were identified as protective factors for normal-variant short stature. This study revealed that parental height, caregiver education, and children's weight significantly influenced the prediction of normal-variant short stature risk, and both the random forest model and gradient boosting machine model exhibited the best discriminatory ability among the 9 machine learning models.

Conclusions: This study revealed a close correlation between environmental growth factors and the occurrence of normal-variant short stature, particularly anthropometric characteristics. The random forest model and gradient boosting machine model performed exceptionally well, demonstrating their potential for clinical applications. These findings provide theoretical support for clinical identification and preventive measures for short stature.

Keywords: Growth environment; Machine learning; Predictive model; SHAP; Short stature.

Fig. 1

Performance of machine learning models to predict normal-variant short stature. This figure shows the Receiver Operating Characteristic (ROC) curves for various machine learning models used to predict normal-variant short stature. The ROC curve plots the true positive rate (sensitivity) against the false positive rate (1-specificity) at various threshold settings. The area under the curve (AUC) for each model is also provided, indicating the model’s overall performance. Higher AUC values represent better model performance

Fig. 2

SHAP Plot on the importance of factors. This figure displays the SHAP (SHapley Additive exPlanations) plot, which illustrates the importance of various factors in the machine learning model’s predictions. Each dot represents a SHAP value for a particular feature for a specific observation. The color of each dot indicates the feature’s value (e.g., high or low). Features are ranked by their importance, with the most important features at the top. This plot helps in understanding how each feature contributes to the model’s output