Endocrine Abstracts

JOINT1955

There is increasing evidence that the incidence of false positive results of growth hormone (GH) stimulation tests (GHST) in children with short stature may be relatively high. The challenge is to identify the patients who may be overdiagnosed with GH deficiency (GHD) for this reason. Interpretation of GHST in children does not take into account their nutritional status. The aim of the study was an attempt to use machine learning methods for prediction of isolated GHD and idiopathic short stature (ISS) in children with excluded other causes of short stature (other endocrinopathies, chronic diseases, genetic defects, malnutrition, cancers, brain injuries), based on auxological indices and IGF-1 concentration. Records of 1592 children with short stature (height SDS <-2. 0), aged 10. 3±3. 4 years (mean±SD), were used for creating classification models for predicting GHD or ISS with different techniques of machine learning. In all of the patients height and weight were measured, two GHST (after clonidine and glucagon) were performed with the cut-off for GH peak 10 µg/l, IGF-1 concentrations were determined, bone age (BA) was assessed, and body mass index (BMI) was calculated. Models were created on raw data (age, sex), age-related variables expressed as SDS for age and sex (hSDS, IGF-1 SDS, BMI SDS), while BA as its ratio to chronological age (BA/CA). Based on GH peak in both GHST <10 µg/l, GHD was diagnosed in 604 patients (37. 9%), including 378 out of 985 boys (38. 4%) and 226 out of 607 girls (37. 2%); the remaining children were diagnosed with ISS. In the decision tree, GHD was predicted in 156 patients, in 149 ones based on BMI SDS >0. 91. Naive Bayes classifier (the method that should include real size of groups) predicted GHD in only 118 cases. The best multilayer perceptron (MLP) neural network predicted GHD in 352 patients. Logistic regression with forward stepwise variable selection classified 269 patients to GHD group, keeping only BMI SDS and IGF-1 SDS as significant variables. All the obtained models classified much less patients to GHD group than the results of GHST <10 µg/l. The significance of higher BMI SDS for prediction of GHD shown in AI models seems a possible cause of inaccuracies of prediction. It seems that these results speak for the need to interpret the results of GHST depending on the nutritional status of children rather than for inability of machine learning techniques to create prediction models.