Effect of Spontaneous GH Secretion and the GH Sampling Period on the Accuracy of Models for Predicting Growth Responses to GH Treatment

WE HAVE recently shown and validated (1) that the initial 1- and 2-yr growth responses to GH treatment can be accurately predicted with the use of mathematical algorithms, so-called “prediction models.” The patient group consisted of prepubertal short (<−2 sd score) children aged 3–15 yr and diagnosed as having GH deficiency (GHD) or idiopathic short stature (ISS). The treatment GH dose was 0.1 IU/kg·d (0.033 mg/kg·d). Five prediction models were constructed, using increasing amounts of pretreatment data and a nonlinear multivariate approach. The first model included exclusively auxology from the start of GH treatment (Basic model), and the second was constructed with the addition of growth data from the first 2 yr of life (Basic+Early growth model). The third and fourth models included the maximum GH level (GHmax) from provocation tests or the IGF-I sd score from the start of treatment, respectively. The best prediction model (i.e. the model with the most narrow prediction interval) included all the auxological data and data from the spontaneous 24-h GH secretion profile. The best predictor from the profile was the single maximum GH value. However, it is well known that the spontaneous maximum GH level usually occurs during nighttime and the 24-h GH profile is laborious to perform. The aim of this study was, therefore, to determine whether the accuracy of prediction could be maintained if the GH sampling period was shortened, but still using all the auxological data (i.e. using the best prediction model). To be able to answer this question data from a large group of children were required with a wide range of GH levels. Data from a total of 279 prepubertal short children (94 with GHD and 185 with ISS, diagnosed according to results from two provocation tests, and aged 3–15 yr) were therefore studied. Data from children used in the construction and validation of the models were included as well as data from children later diagnosed as having GHD or ISS. The studies were approved by the Ethics Committees of the Medical Faculties of the Universities of Göteborg, Lund, Linköping, Uppsala, and Umeå and of the Karolinska Institute. Informed consent was obtained from the parents of each child and from the child, where appropriate. The GHmax values were distinctly clustered around two time points, 2400 h and 0500 h, irrespective of the actual serum GHmax level. The period 2000 h to 0800 h contained 89% of the GHmax values, and the period 2000 h to 0200 h contained 53% (Fig. 1). Guided by the timing of the GHmax values over 24 h, with the aim of including as many GHmax values as possible, the periods 2000 h to 0800 h and 2000 h to 0200 h were selected for comparison with the prediction result from the full 24-h GH sampling period. The GHmax values were distinctly clustered around two time points, 2400 h and 0500 h, irrespective of the actual serum GHmax level. The cutoff level for the diagnosis of GHD was 32 mU/liter with the monoclonal assay and WHO IRP 80/505, corresponding to 10 μg/liter with the polyclonal assay and WHO IRP 66/127. The GHmax values were distinctly clustered around two time points, 2400 h and 0500 h, irrespective of the actual serum GHmax level. The cutoff level for the diagnosis of GHD was 32 mU/liter with the monoclonal assay and WHO IRP 80/505, corresponding to 10 μg/liter with the polyclonal assay and WHO IRP 66/127. The GHmax from the selected time periods was used for predicting the growth response. For those children whose GHmax over 24 h did not occur during the selected time periods, this approach resulted in overprediction of the growth response. The effect of using different GHmax levels on the positions of the prediction intervals for an individual child is illustrated in Fig. 2. Depiction in one child of the positions of the prediction intervals resulting from the use of GH sampling periods shorter than 24 h, when the GHmax24 h occurred during daytime hours. Left, 1, using the peak from the 1600 h clock-time (24-h GH sampling); middle, 2, using the GH peak between 2000 h to 0200 h; right, 3, using the peak at 0300 h (GH sampling 2000 h to 0800 h). The inset shows the actual growth of the child before the start of GH therapy (broken line) and the first-year growth during GH therapy (solid line). The bar in the prediction interval (symbol) represents ±1 sd, and the whiskers represent ±1.96 sd. Depiction in one child of the positions of the prediction intervals resulting from the use of GH sampling periods shorter than 24 h, when the GHmax24 h occurred during daytime hours. Left, 1, using the peak from the 1600 h clock-time (24-h GH sampling); middle, 2, using the GH peak between 2000 h to 0200 h; right, 3, using the peak at 0300 h (GH sampling 2000 h to 0800 h). The inset shows the actual growth of the child before the start of GH therapy (broken line) and the first-year growth during GH therapy (solid line). The bar in the prediction interval (symbol) represents ±1 sd, and the whiskers represent ±1.96 sd. When the 12-h period (2000 h to 0800 h) for GH sampling was used, overprediction of the growth response to GH therapy was observed in 11% (range, 0–0.32 sd scores; median, 0.07), but it was substantial (>0.1 sd score) in only 3% of the total group. The corresponding values for the 6-h period (2000 h to 0200 h) were 50% (range, 0–0.56 sd scores; median, 0.08) and 12%. The performance of a prediction model is estimated by the differences (residuals) between the predicted and the observed outcomes, hereΔ height sd score. Results are presented as the sd for the residuals, sdres, a measurement of the model accuracy. The lower the sdres, the better. For the total group of 279 children, the prediction accuracy for the first year Δheight was estimated by the sdres, using the prediction based on the GHmax from the 12-h and 6-h sampling periods. When data on growth in early life were included in the estimates, the sdres were 0.194 and 0.207 for 12 h and 6 h, respectively, and 0.202 and 0.214 for 12 h and 6 h, respectively, if early growth data were not included. As expected, the sdres for these shorter GH sampling periods were higher than those obtained using the true GHmax over 24 h (0.191 with data on growth in early life included and 0.199 if early growth data were not included; Ref. 1), but still lower (i.e. better) than the corresponding values achieved using models using the GHmax from an arginine-insulin tolerance test (GHmaxAITT) or levels of IGF-I before the start of treatment (0.268 and 0.239) (1), respectively. Overnight blood sampling for estimation of spontaneous GH secretion may be regarded as laborious and stressful for the child, but compared with a provocation test it has many advantages: it is physiological, induces no discomfort, and, with the use of a withdrawal pump, is convenient and almost risk free. In addition, the resulting GH estimate used in the prediction model is a good predictor of the growth response to GH treatment and can be used to aid in the selection of children for an expensive treatment that may continue for many years. Thus, the expense of the investigation in terms of costs, time, discomfort, and risk should be compared with the reliability of the results achieved, together with the expense of treatment. The GHmax obtained during a sampling period limited to 12 h at night results in an acceptable prediction level for 97% of the children and is still a better predictor of the growth response than the GHmax AITT or IGF-I levels when used in our prediction models. When considering GH treatment in a short child, we suggest that this work-up procedure, with the use of a computerized prediction model, may ultimately replace the hazardous provocation test that is such a poor predictor of the growth response to treatment. GH deficiency; idiopathic short stature.