ESPE Abstracts

Background: Prediction of the effectiveness of GH therapy in children with short stature is an important issue. Artificial neural networks (ANN) seem to be promising tool for this purpose, not requiring any assumption on functions linking independent and dependent variables.

Objective and hypotheses: The aim of the study was to compare ANN models of GH therapy effectiveness, based on the data available at therapy onset with multiple linear regression (MLR) model.

Method: Retrospective analysis comprised the data of 150 short children (101 boys), diagnosed with isolated GH deficiency, treated with GH up to the attainment of final height (FH). The following parameters (input variables) were assessed before treatment for each patient: gender, chronological age, bone age, mothers’ and father’s height, pubertal status, height velocity, GH peak after falling asleep and in two stimulation tests, IGF1 and IGFBP-3 concentrations, birth weight and gestational age. The output variable was FH or FH SDS.

Results: The best MLR model included height SDS of the patient and of parents, pre-treatment height velocity and IGF1 SDS as significant variables and explained 44% of variability of FH SDS in learning group and 36% in testing group, with the mean error (RMSE) of predicted FH 3.5 and 3.8 cm respectively. The best ANN model for the same input variables explained 43% of variability of FH SDS for learning group and 40% for testing group with RMSE 3.6 and 3.7 cm respectively. The best ANN model eliminated GH peak after falling asleep and father’s height and explained 86% of variability of FH SDS for both learning and testing group with RMSE 3.2 and 3.4 cm respectively.

Conclusion: Neural networks are more accurate in FH prediction and explain more variability of FH in children with isolated GH deficiency than linear regression.