ESPE Abstracts (2016) 86 P-P1-602

Prediction of First Year Response to Growth Hormone Treatment in Neural Network Models

Urszula Smyczynskaa, Joanna Smyczynskab & Maciej Hilczerb,c


aDepartment of Automatics and Biomedical Engineering, AGH University of Science and Technology, Cracow, Poland; bDepartment of Endocrinology and Metabolic Diseases, Polish Mother’s Memorial Hospital – Research Institute, Lodz, Poland; cDepartment of Pediatric Endocrinology, Medical University of Lodz, Lodz, Poland


Background: Accurate prediction of responsiveness to growth hormone (GH) therapy is an important issue. The 1st year response to treatment is regarded as significant predictor of the attained final height. Neural networks are techniques of machine learning which do not require any assumptions and preprocessing of data, contrary to linear regression models.

Objective and hypotheses: The aim of the study was to predict height velocity (HV) during 1st year of therapy (HV-1) in GH treated children with isolated GH deficiency.

Method: Our retrospective analysis comprised data of 253 patients, age 11.5±2.8 (2.5–15) years, for whom we tried to predict HV-1 using multilayer perceptron (MLP) and radial basis function (RBF) neural networks. Mean HV before treatment (HV-0) in those patients was equal to 4.2±1.3 cm/year, while during 1st year of treatment 9.6±1.9 cm/year (4.9–17.0 cm/year). As potential predictors of HV-1 we included height and HV-0, parental heights, IGF-I and IGFBP-3 concentrations, chronological and bone age (BA), results of GH stimulation tests with clonidine and glucagon, and patient’s gender.

Results: Best MLP network predicted HV-1 with mean error 1.77 cm/year in learning data and 1.70 cm/year in testing set. It included all predictors with exception of gender. Best RBF network was characterized by averaged error equal to 1.76 cm/year in original data and 1.77 cm/year in testing dataset, but it included only BA, father’s height, result of test with glucagon and concentrations of IGF-I and IGFBP-3.

Conclusion: Models tend to reproduce general, averaged tendencies rather than extreme values for particular patients, thus the range of answers they produced was narrower (e.g. for MLP network 7.4–12.4 cm/year) than in the case of real values. Together with obtained relatively low error this feature may allow us to use such models for identifying patients with poor response to treatment.

Article tools

My recent searches

No recent searches.