Proposal of a clinical response score and predictors of clinical response to 2 years of GH replacement therapy in adult GH deficiency

in European Journal of Endocrinology
View More View Less
  • 1 Medizinische Klinik und Poliklinik IV, Klinik für Neurochirurgie, Institut für Klinische Chemie und Labormedizin, Universitätsklinik für Innere Medizin III, Department of Endocrinology, Klinik für Innere Medizin, Pfizer Endocrine Care; Sollentuna, Klinikum der Universität München, Ziemssenstr. 1, 80336 Munich, Germany

Free access

Objective

There is no single clinical marker to reliably assess the clinical response to growth hormone replacement therapy (GHRT) in adults with growth hormone deficiency (GHD). The objective of this study was to propose a clinical response score to GHRT in adult GHD and to establish clinical factors that predict clinical response.

Design

This was a prospective observational cohort study from the international KIMS database (Pfizer International Metabolic Database).

Methods

We included 3612 adult patients with GHD for proposing the response score and 844 patients for assessing predictors of response. We propose a clinical response score based on changes in total cholesterol, waist circumference and QoL-AGHDA quality of life measurements after 2 years of GHRT. A score point was added for each quintile of change in each variable, resulting in a sum score ranging from 3 to 15. For clinical response at 2 years, we analysed predictors at baseline and after 6 months using logistic regression analyses.

Results

In a baseline prediction model, IGF1, QoL-AGHDA, total cholesterol and waist circumference predicted response, with worse baseline parameters being associated with a favourable response (AUC 0.736). In a combined baseline and 6-month prediction model, baseline QoL-AGHDA, total cholesterol and waist circumference, and 6-month change in waist circumference were significant predictors of response (AUC 0.815).

Conclusions

A simple clinical response score might be helpful in evaluating the success of GHRT. The baseline prediction model may aid in the decision to initiate GHRT and the combined prediction model may be helpful in the decision to continue GHRT.

Abstract

Objective

There is no single clinical marker to reliably assess the clinical response to growth hormone replacement therapy (GHRT) in adults with growth hormone deficiency (GHD). The objective of this study was to propose a clinical response score to GHRT in adult GHD and to establish clinical factors that predict clinical response.

Design

This was a prospective observational cohort study from the international KIMS database (Pfizer International Metabolic Database).

Methods

We included 3612 adult patients with GHD for proposing the response score and 844 patients for assessing predictors of response. We propose a clinical response score based on changes in total cholesterol, waist circumference and QoL-AGHDA quality of life measurements after 2 years of GHRT. A score point was added for each quintile of change in each variable, resulting in a sum score ranging from 3 to 15. For clinical response at 2 years, we analysed predictors at baseline and after 6 months using logistic regression analyses.

Results

In a baseline prediction model, IGF1, QoL-AGHDA, total cholesterol and waist circumference predicted response, with worse baseline parameters being associated with a favourable response (AUC 0.736). In a combined baseline and 6-month prediction model, baseline QoL-AGHDA, total cholesterol and waist circumference, and 6-month change in waist circumference were significant predictors of response (AUC 0.815).

Conclusions

A simple clinical response score might be helpful in evaluating the success of GHRT. The baseline prediction model may aid in the decision to initiate GHRT and the combined prediction model may be helpful in the decision to continue GHRT.

Introduction

Growth hormone replacement therapy (GHRT) exerts several effects in adult patients with growth hormone deficiency (GHD), including improvement in lipids, bone density, quality of life and body composition. This has been shown both in randomized placebo-controlled trials (1) and in large-scale observational studies (2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12).

GHRT is an expensive therapy and has a potential risk of side effects. Therefore, it is important to identify those patients who will benefit most from GHRT. In children, growth velocity is a good and simple clinical marker that allows monitoring response to GHRT. Moreover, reliable prediction models for final height have been established in children, based on growth response to GHRT (13, 14, 15).

However, in adults, response to GHRT cannot be measured reliably with one single clinical marker. This makes it difficult to quantify clinical response and, thus, identify potential predictors of clinical response in adult GHD. Generally, the dose of GH is titrated by IGF1 levels. However, it is not known if improvement in IGF1 automatically translates into a relevant improvement of the clinical condition of GHD patients.

Therefore, we aimed to propose a generalized criterion of clinical response to GHRT in adult patients with GHD using values from simple measurements that are performed during standard follow-up. In addition, we aimed to identify factors at baseline and during the early course of treatment that predict such a clinical response to GHRT.

We analyzed data from the KIMS database for this study.

Subjects and methods

Study cohort

KIMS (Pfizer International Metabolic Database), a long-term safety and outcome study of GH replacement therapy in adults, represents a large pharmaco–epidemiological survey initiated in 1994 (16). All patients gave written informed consent for their participation in KIMS. We included all GH replaced subjects from KIMS worldwide with data at baseline and 2 years on QoL-AGHDA, waist circumference and total cholesterol, except patients who had been GH-replaced before KIMS start (only treatment-naïve patients were included), patients diagnosed with Cushing's syndrome, acromegaly and craniopharyngioma and patients from the UK. Patients with Cushing's syndrome, acromegaly and craniopharyngioma were excluded because of the specific metabolic profiles of these diseases (6, 8). Patients from the UK were not included because QoL-AGHDA score is a criterion for GH subscription in UK, which might have caused potential bias. The data freeze was from August 20, 2012.

Methods

In KIMS, a large number of clinical parameters are documented. Insulin-like growth factor 1 (IGF1) was measured centrally and lipids were measured centrally or locally using standard procedures.

An analysis of 1926 observations with parallel central and local measurements of cholesterol showed a median and mean difference of 0.00 and 0.03 mmol/l. We considered this difference clinically not relevant and continued using central or local laboratory measurements for cholesterol with central measurement as preference when parallel measurements were reported for a single patient.

Until November 2002, serum IGF1 was determined by RIA after acid–ethanol precipitation of IGF-binding proteins (Nichols Institute Diagnostics, San Juan Capistrano, CA, USA). Thereafter, a chemiluminescence immunoassay (Nichols Advantage System, Bad Vilbel, Germany) was introduced. Long-term reproducibility, measured during >1 year, showed a coefficient of variation <9% in the concentration range of 130–850 mg/l. The assay detection limit was 30 mg/l (16). For each assay, age- and gender-specific reference parameter estimates were used to calculate IGF1 SDS.

Statistical analyses

For the purpose of a simple and easy to-use clinical response tool it seemed most promising to choose variables that respond well to GH replacement, are routinely assessed during follow-up and are not, or only weakly, correlated (for example, when choosing between total cholesterol and LDL cholesterol, only one should be selected, since both their baseline values and their response to GH are expected to be similar).

Variables that fulfilled these criteria were quality of life (measured by QoL-AGHDA), waist circumference and total or LDL cholesterol. QoL-AGHDA is a score consisting of 25 problem items to be answered with yes or no. A score of 25 indicates problems on all items and 0 no problems on any item. Analysis of KIMS has shown that the maximum effects of GH treatment, particularly on lipids, are seen after 2 years of treatment (5,6, unpublished data). Therefore, we assessed the change (‘delta’) on the following variables before and after 2 years of GH treatment in KIMS: QoL-AGHDA, waist circumference and total cholesterol.

We chose total cholesterol rather than LDL cholesterol for reasons of simplicity: previous data have shown that cholesterol changes are mainly driven by changes in LDL cholesterol, whereas HDL cholesterol remains unchanged after GH therapy. Total cholesterol is the easiest to determine.

We tested linear correlations with Pearson's and Spearman's correlation analyses. We checked for normality by visual analysis.

Quintiles and their cut-off points were determined based on the 2-year delta for the respective variables. By adding one score point for each quintile (1 to 5) for each of the three delta variables, an overall response score of 3–15 points was achieved. A minimum score of three indicates the lowest level of response (quintile 1) for each delta variable and a maximum score 15 indicates the highest level of response (quintile 5) for each delta variable. We considered all three parameters equally important. Therefore, we gave equal weight to each parameter.

Based on these scores, a dichotomous response variable Y was constructed, where response (Y=1) was defined as having a score of 7 or more and non-response as having a score of 6 or less (Y=0). The reason for this cut off was that the lowest two quintiles for a particular delta variable were associated with a worsening, whereas the upper three quintiles were associated with an improvement of that outcome (see Table 1). In clinical practice, a deterioration of the three response variables is observed over time without treatment (2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12). Therefore, we already considered an unchanged condition represented with a score of 7 as a response. This classification does not imply that those patients with a score of <7 did not show any clinical response to GH treatment.

Table 1

Quintiles for 2-year change. Quintile 1 scores 1, quintile 2 scores 2, etc. Each quintile in each category can be summed up to a response score a minimum of 3 (no response) and a maximum of 15 (maximal response).

Waist circumferenceQoL-AGHDATotal cholesterol
Quintile 1Delta W >3 cmDelta A ≥2 itemsDelta C >0.4 mmol/l
Quintile 20.0 <Delta W ≤3 cm0 ≤Delta A ≤1 items−0.1 <Delta C ≤0.4 mmol/l
Quintile 3−2.0 ≤Delta W≤0.0 cm−3 ≤Delta A≤−1 items−0.5 ≤Delta C ≤−0.1 mmol/l
Quintile 4−5.8 <Delta W <−2.0 cm−7 ≤Delta A ≤−4 items−1.1 ≤Delta C <−0.5 mmol/l
Quintile 5Delta W ≤−5.8 cmDelta A ≤−8 itemsDelta C <−1.1 mmol/l

Delta W, 2-year change in waist circumference; Delta A, 2-year change in QoL-AGHDA score; Delta C, 2-year change in total cholesterol.

In the aim of modelling the 2-year response probabilities, two logistic regression models were assessed: Model 1 examined eight KIMS baseline variables as potential response predictors. These were QoL-AGHDA, waist circumference, total cholesterol, height, age, gender, IGF1 SDS and the number of additional pituitary deficiencies. We also incorporated quadratic terms for all numerical variables. Analyses were performed by backward variable elimination. (Proc logistic, SAS 9.2). To test for potential inclusion bias we compared clinical variables of patients included in model 1 and patients excluded from model 1.

Model 2 incorporated all significant baseline variables from model 1 and three 6-month numerical delta variables. Final models were reached by backward elimination procedures. (Proc logistic, SAS 9.2.). In further steps we incorporated only those baseline variables that remained significant in model 1 and in a step-by-step approach we consecutively tested each single 6-month numerical delta in QoL-AGHDA, cholesterol and waist circumference. Finally, we incorporated only the significant baseline predictors and the delta variable that contributed most to predictive power and had the lowest missing values.

All variables, except the categorical 2-year response variable, gender and the number of additional pituitary deficiencies, were numerical. A quadratic term was also assessed for numerical variables that were found as statistically significant.

A variable was eliminated from a model if the Wald-based significance value was >5%. Significance level was set to 5%. CI were Wald-based.

Areas under the curve (AUC) were calculated with ROC methods (Proc logistic, SAS 9.2). Generally, AUC values between 0.70 and 0.90 are considered ‘moderate accuracy’ and values between 0.90 and 1 as ‘high accuracy’ in classifying a diagnostic test (17). Goodness of fit of the models was performed with the Hosmer-Lemeshow test (18). For outlier influence assessment we used residual analyses and leave-one-out methods (19, 20, 21).

In sensitivity analyses, we tested models 1 and 2 using ≥5 and ≥9 instead of ≥7 as alternative cut-off values for the response score.

Results

Correlations

All variables except baseline AGHDA score were close to normally distributed. The Pearson correlation coefficient between baseline cholesterol and baseline waist circumference was −0.04; between baseline cholesterol and baseline QoL-AGHDA score 0.009; between baseline waist circumference and baseline QoL-AGHDA score 0.076. The correlation between 2-year delta cholesterol and 2-year delta waist circumference was 0.11; between 2-year delta cholesterol and 2-year delta QoL-AGHDA score 0.03; and between 2-year delta waist circumference and 2-year delta QoL-AGHDA score −0.01. Spearman correlation analysis yielded similar results. This confirmed our primary assumption that the three clinical response variables were not or only weakly inter-correlated.

Quintiles

Out of 16 266 patients in KIMS at data close, 3612 patients had data at baseline and 2 years on QoL-AGHDA, or waist circumference or total cholesterol, and fulfilled the other inclusion and exclusion criteria. Quintiles were based on all patients with full baseline and 2-year data on the respective variable (waist circumference n=2110, QoL-AGHDA n=1545 total cholesterol n=1417). Table 1 displays the quintile values.

Predictors of response

A total of 844 patients fulfilled all inclusion and exclusion criteria and had full data at baseline and 2 years on QoL-AGHDA, waist circumference and total cholesterol. Table 2 displays the baseline characteristics of these patients. Note, however, that due to a varying number of missing values on candidate predictor variables, the different analyses were based on a different number of observations.

Table 2

Baseline characteristics of all patients with complete baseline and 2-year delta variables.

MenWomen
n462382
Age (years)50.6±13.046.9±13.2
Waist circumference (cm)100.7±12.493.5±14.8
Height (cm)177.4±8.2164.2±7.4
QoL-AGHDA score items (median (range))9 (0–25)11 (0–25)
Total cholesterol (mmol/l)5.70±1.205.82±1.24
IGF1 SDS−1.55±1.56−1.97±1.77

Data are shown as means±s.d. unless otherwise indicated.

Model 1 was based on 791 patients due to 51 missing observations on IGF1 SDS and 2 on height at baseline. Comparison of clinical characteristics between the 791 patients included in model 1 and those not included showed no significant differences in 2-year changes of cholesterol, AGHDA score, or waist circumference and baseline age, cholesterol, IGF1 SDS, or the number of additional deficiencies. However there were clinically marginal but statistically significant differences in AGHDA score (excluded patients mean +1.2 scores points; P<0.0001), height (−2.5 cm; P<0.0001) and percentage of males (−4.8%; P=0.0175).

In the backward elimination procedure, baseline variables height, age, sex and the number of additional pituitary deficits were non-significant and eliminated. Baseline variables QoL-AGHDA, waist circumference, total cholesterol and IGF1 SDS remained in the model as predictors. The odds ratio (95% CI) per item QoL-AGHDA was 1.070 (CI 1.037–1.105); per centimetre waist circumference, it was 1.026 (CI 1.010–1.042); per mmol/l total cholesterol, 1.941 (CI 1.595–2.366); and per unit for IGF1 SDS, 0.874 (CI 0.765–0.992). Table 3 summarizes the model-building process.

Table 3

Summary of backward elimination and final Model 1 based on 791 observations.

StepDFNumber inWald χ2P valueAUCs.e.m.
Effects removed
1 Age21140.01730.89520.757
2 Cholesterol2 (mmol/l)1130.21280.6446
3 Gender1120.32260.5700
4 No. of additional deficiencies21110.47350.4914
5 No. of additional deficiencies1100.00330.9545
6 IGF1 SDS2190.99920.3175
7 Height2181.40950.2351
8 Height (cm)170.81380.3670
9 Age (years)160.99890.3176
10 Waist2153.60200.0577
11 AGHDA score2143.78380.0518
12Final model: effects remaineda0.736
 Intercept1−5.205127.2477<0.00010.9972
 AGHDA score10.068017.2919<0.00010.0163
 Waist10.025610.56340.00120.00787
 Cholesterol(mmol/l)10.663343.5398<0.00010.1005
 IGF1 SDS1−0.13424.07950.04340.0665

Values of number in for final model are an estimate.

The area under the ROC curve (AUC) for this model was 0.736 (Fig. 1A). For prediction of 2-year response, the following equation for the predictive response probability was established:

Figure 1
Figure 1

ROC curves for prediction of 2-year response by baseline QoL-AGHDA, waist circumference, total cholesterol, and IGF1 SDS (Model 1) and for prediction of 2-year response by baseline QoL-AGHDA, waist circumference and total cholesterol and 6-month delta waist circumference (Model 2).

Citation: European Journal of Endocrinology 173, 6; 10.1530/EJE-15-0305

article image

where bX was estimated to −5.2271+0.0681A+0.0257W+0.6640C+−0.1377I (final model 1) where A stands for total AGHDA score, W for waist circumference in centimetre, C for total cholesterol in mmol/l and I for IGF1 SDS in SDS units, all at baseline.) If a sum score of ≥5 and ≥9 instead of ≥7 was used as a response cut-off, the AUC for model 1 were 0.712 and 0.705 instead of 0.736 respectively. With the cut-off of ≥5 and ≥9 instead of ≥7, only two and three instead of four variables remained significant predictors respectively.

For model 2, we performed backward elimination on all significant baseline variables and 6-month delta waist circumference only (454 non-missing observation), which ended up with AGHDA score, cholesterol and waist circumference at baseline and 6-month delta waist circumference as significant predictors. The AUC for this model was 0.815 (Fig. 1B). The Pearson's correlation coefficients between baseline values and delta changes for total cholesterol, QoL-AGHDA score and waist circumference were −0.43, −0.41 and −0.10 respectively. Thus, waist circumference showed the weakest correlation between baseline and delta change. This supports the use of waist circumference as a 6-month variable.

For prediction probabilities of 2-year response, the following equation was formulated based on baseline and 6 months delta information:

article image

where bX was estimated to −6.1032+0.0706A+0.0285W+0.8391C−0.1800 ΔW (where A stands for total AGHDA score, W for waist circumference in centimetre, C for total cholesterol in mmol/l, all at baseline, and ΔW for 6-month change in waist circumference.)

Performing the corresponding exercise for 6-month delta AGHDA score (295 non-missing observations) resulted in an equation with waist circumference and cholesterol at baseline and 6-month delta AGHDA score as significant predictors (parameter estimates not shown) with AUC of 0.784. Correspondingly for 6-month delta cholesterol (243 non-missing observations) the model resulted in an equation with AGHDA score and cholesterol at baseline and 6-month delta cholesterol (parameter estimates not shown) with an AUC of 0.778. In none of these models was baseline IGF1 SDS of importance. Therefore, as final model 2 we selected the model with baseline QoL AGHDA score, baseline waist circumference, baseline total cholesterol and delta waist circumference, as it had a high AUC and the highest number of non-missing observations.

If a sum score of ≥5 and ≥9 instead of ≥7 was used as a response cut-off, the AUC for model 2 were 0.726 and 0.732 instead of 0.815 respectively. With the cut-off of ≥5 and ≥9 instead of instead of ≥7, only one and three instead of four variables remained significant predictors respectively.

Assessment of the models' predictive probabilities

Using the equations provided for final models 1 and 2 respectively we could calculate predicted response probabilities ranging between 0 and 1.

We used the whole data set for model building and did an internal cross-validation using the leave-one-out approach. With this approach, we examined how well we could predict the observation that was left out for each fit, repeating this step until all observation were left out once. With this approach the percentage of outliers were 4.7 and 3.7% in model 1 and model 2 respectively. This was below the cut-off of 5% considered acceptable and showed a good validation.

By classifying patients in the datasets for model 1 and 2 respectively, we could define a patient as a predicted responder if the calculated predicted probability was equal to or greater than a chosen predicted probability. We compared predicted responder status with actual responder status after 2 years of treatment. Thereby we could calculate sensitivity, specificity, positive predictive value and negative predictive value respectively for each selected cut-off of the model predicted probabilities, which are shown in Fig. 2. Positive predictive value corresponds to the probability that the patient actually responded to the treatment (by having a score of at least 7 at the 2-year visit), given the patient was predicted as a responder. The negative predictive value corresponds to the probability that the patient actually did not respond (by having a score of 6 or less at the 2-year visit), given the patient was predicted as a non-responder. Sensitivity measures the probability that the patient was predicted as a responder, given the patient is a responder (by having a score of at least 7 at the 2-year visit). Specificity measures the probability that the patient was predicted as a non-responder given that the patient is a non-responder. For the interpretation of the curves in Fig. 2, one can note that the data set for model 1 had a total number of actual non-responders at the 2-year visit of 126 (16%) and model 2 had 63 (14%). From the figures we can read that sensitivity and specificity curves intersects at around 65% for model 1 and 72% for model 2 at a predicted response probability of around 0.85. Positive predictive value is relatively high for all predicted response probabilities, whereas the negative predictive value decreases rapidly with increasing cut-off of the predicted response probability. In both models, a predicted response probability of 0.9 was already associated with a specificity and positive predictive value of more than 90%. A predicted response probability of 0.7 was associated with a sensitivity of 90% but a negative predictive value of only 40%. A predictive response score of 0.5 was also associated with negative predictive values of around 90%.

Figure 2
Figure 2

Sensitivity, specificity, positive predictive value and negative predictive value by response cutoff predictive probability for Model 1 (panel A) and Model 2 (Panel B), respectively. A patient is classified as a responder if the calculated model predictive probability is equal or greater than a given cutoff probability, and as a non-responder if the calculated predictive probability is lower than that, based on model 1 and 2, respectively. The procedure is repeated for all patients in the respective model data set and for some selected cutoff probablities. Different measures shown in the graphs were calculated comparing ‘predicted’ response (at baseline, model 1; and after 6 months of treatment, Model 2) with actual observed response after 2 years of GH treatment. Model 1 incorporates variables total QoL AGHDA score, waist circumference, total cholesterol, and IGF1SDS at baseline to predict 2 year response. Model 2 incorporates additional information on change in waist circumference after 6 months of GH treatment as predictor. As baseline IGF1SDS was no longer significant (P=0.66) as predictor after 6 months of GH treatment it was excluded from model 2.

Citation: European Journal of Endocrinology 173, 6; 10.1530/EJE-15-0305

Outlier and performance analysis

Assessment of goodness of fit of the models using Hosmer-Lemeshow tests showed relatively good agreement between observed and expected numbers (model 1: χ2 (8)=12.21, P=0.14; model 2: χ2 (8)=7.37, P=0.50). Observations with a delta deviance (ΔD) value greater or equal to 3.84 (5% percentile for χ2 (1)) were defined as outliers. A percentage of 5% or less outliers is generally considered expected. Model 1 had 37 outliers (4.7% of observations) and model 2 had 17 outliers (3.7%). All outliers (in both models 1 and 2) were observed non-responders with high predicted probabilities of being a responder. In model 1, 18 of 37 outliers (49%) and 391 of 754 non-outliers (52%) were reported with adverse events with diagnosis date before or around the 2 year visit, respectively. For model 2, the corresponding numbers were 11/17 (65%) for outliers and 259/437 (60%) for non-outliers respectively. Therefore, the adverse event rates were similar among outliers and non-outliers. Also, the rates of serious adverse events were similar among outliers and non-outliers (data not shown). One outlier patient was prescribed a zero GH dose at 2 years and one patient had an off-treatment period during the first 2 years of treatment (0.1 years, or approximately 1 month).

Discussion

We propose a simple clinical response score based on three clinical parameters that are routinely assessed during follow-up of GHRT. We additionally established two simple prediction models that allow the prediction of clinical response after 2 years, defined by a sum score of delta change in total cholesterol, waist circumference and QoL-AGHDA. In the baseline model, four variables were significant predictors of clinical response. These included the three variables that also compose the response score, namely, cholesterol, QoL-AGHDA, waist circumference, and IGF1 SDS. Worse values on baseline variables (higher cholesterol, AGHDA scores or waist circumference and lower IGF1 SDS) were associated with higher response probabilities. The combined model (model 2) with baseline variables QoL-AGHDA, waist circumference and cholesterol and 6-month changes in waist circumference predicted clinical response well.

In secondary analyses, we showed that using alternative cut-offs reduced the number of significant predictors and decreased the AUCs of the models. This shows that the chosen cut-off is not only clinically meaningful but also statistically the most valuable cut-off.

Both models showed a good goodness of fit with similar observed and expected values according to the Hosmer-Lemeshow test. The discriminatory power of model 1 was fair, with an AUC of >0.7, and the discriminatory power of model 2 was good, with an AUC >0.8. Both models were robust and based on relatively high numbers of non-missing values (n=791 and n=454 for models 1 and 2 respectively).

Effects of GHRT on cardiovascular risk markers, body composition and quality of life have been well documented in both clinical trials and observational studies (1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12). Additionally, in a previous analysis of the German KIMS data, we have shown that cardiovascular risk assessed by conventional risk scores such as the Framingham score are beneficially influenced by GHRT (10). However, to our knowledge, until now there have been no data combining different dimensions of GH effects such as body composition, lipid markers and quality of life into a unified marker of clinical response.

Our data confirm previous studies showing that those patients with worse baseline characteristics respond best to GHRT (2, 5, 8, 11). In a previous KIMS analysis by Feldt-Rasmussen et al., sex, age and baseline IGF1 were predictors of IGF1 response and clinical changes to GHRT (11) In another study, sex was also the best predictor of favourable change in body fat and IGF1 response (9). In these studies, the association of different clinical baseline variables with IGF1 response and changes in clinical outcome variables such as weight, BMI, cholesterol and quality of life or body fat were analyzed. Our study confirmed IGF1, but not age or sex, as baseline predictors of clinical response. Both the different statistical approaches and different outcome variables might account for this difference. In the study by Feldt-Rasmussen et al., the main outcome variable was change in IGF1 SDS, and clinical outcomes were only studied in univariate analyses.

We analyzed combined clinical response variables using multiple logistic regression methods. Moreover, we intentionally did not include changes in IGF1 into our response score, as our aim was to establish a pure clinical response model, based on easy-to-measure characteristics of the patient. Also, GH doses are generally titrated according to IGF1 response. This might have additionally blunted the sex differences, as women were likely treated with higher GH doses in our cohort.

We used a model with a dichotomized response. We are aware that this approach may cause some loss of information. However, we think that this approach is most useful for the purpose of our study, as we aimed for a simple and clinically useful response model.

Our study extends previous work by proposing a unified clinical response score and by providing two simple prediction models for clinical response: a baseline model that can be helpful for the decision to initiate GH treatment (model 1) and a combined model that helps to decide on continuation of GH treatment 6 months after GHRT start (model 2).

The results provided in this study might be useful for future clinical research and clinical practice in several aspects. The clinical response score was established on a large data set of 1417–2110 patients. This score can be a helpful tool in assessing clinical response in clinical studies on the effect of GHRT. It can also be used in more detailed future studies on response by subgroups such as sex, genetic background, baseline IGF1 levels, GH peak or others.

The prediction models established here allow predicting individual 2-year response both by baseline characteristics and 6-month changes using only a small number of variables. The combined prediction model with baseline and 6-months data showed a particularly good predictive ability and is easy to use in every-day clinical practice, as it only requires the routine measurements on waist circumference, cholesterol, the common and simple quality-of-life tool QoL-AGHDA and 6-months changes in waist circumference. It is of note that our prediction models should only be used after the other possible deficiencies are adequately replaced.

In clinical practice, for model 1, the four baseline prediction parameters waist circumference, QoL-AGHDA score, total cholesterol and IGF1 can be inputted in the regression equation provided. The regression equation yields a probability value between 0 and 1. Correspondingly, model 2 can be used after 6 months GH replacement. A probability value of 0.9 or higher indicates that the patient is highly likely to respond to treatment, whereas a probability of 0.5 or lower indicates that the patient is unlikely to respond to treatment.

Some limitations of our study need addressing. We do not know if our data are generalizable to patients with Cushing's disease, craniopharyngioma or previous acromegaly, as they were excluded from our study. Also, patients from the UK were excluded, as QoL-AGHDA data from UK might be biased since prescription depends on baseline QoL-AGHDA score. Therefore, we do not know if our data are generalizable to patients from the UK as well as from other countries that did not contribute to KIMS.

As in most observational studies, the number of missing values was high. We cannot exclude a potential exclusion bias. However, we expect this to be marginal, as only height, sex distribution and AGHDA score differed marginally. There were no differences in the other baseline values and in the 2-year changes of the response score variables.

Quality of life was measured with the QoL-AGHDA questionnaire. There are several other additional measurement tools of health-related quality of life. We do not know if our data can be generalized to cohorts with different measures of quality of life. In general, we do not know if our data can be generalized to populations treated with other GH formulations. However, we think that large differences are unlikely, as the effects of all GH formulations can be expected to be comparable. Also, possible changes in eating habits, physical activities and socio-economics status in the patients during the follow up period might affect accuracy of our prediction model.

Our data were not validated in an external cohort. However, using the leave-one-out method, we had less than 5% acceptable outliers and thus a good validation. In both models, the number of outliers was in the acceptable range below 5%. However, all outliers tended into one direction, being non-responders with high predicted response probabilities. The number of adverse events was similar among outliers and non-outliers. Therefore it is unlikely that this was influential. Additional factors such as non-adherence may also play a part.

In summary, we have proposed a clinical response score that allows individual clinical monitoring based on simple clinically meaningful variables. Moreover, we have established two simple clinical prediction models that allow predicting response in an individualized approach. This can be directly useful in helping to decide whether to initiate and whether to continue GH therapy.

Declaration of interest

H J Schneider received an independent research grant and statistical support for this study. During the work on this study, H J Schneider, M Buchfelder, H Wallaschofski and Peter H Kann were members of the German KIMS board, M Buchfelder, A Luger and G Johannsson were members of the KIMS strategic advisory board, and A Mattsson is an employee of Pfizer, Inc., Sweden. A Luger has received honoraria for presentations and/or participating in advisory boards from Pfizer, Ipsen, Novo Nordisk and Merck/Serono.

Funding

This study was supported by an independent research grant and by statistical support from Pfizer.

Acknowledgements

We thank all KIMS recruiting centres for contributing their data and all KIMS patients for allowing their data to be used. KIMS was funded by Pfizer, Inc.

References

  • 1

    Maison P, Griffin S, Nicoue-Beglah M, Haddad N, Balkau B, Chanson P. Impact of growth hormone (GH) treatment on cardiovascular risk factors in GH-deficient adults: a metaanalysis of blinded. randomized, placebo-controlled trials. Journal of Clinical Endocrinology and Metabolism 2004 89 21922199. (doi:10.1210/jc.2003-030840).

    • Search Google Scholar
    • Export Citation
  • 2

    Bengtsson BA, Abs R, Bennmarker H, Monson JP, Feldt-Rasmussen U, Hernberg-Stahl E, Westberg B, Wilton P, Wuster C. The effects of treatment and the individual responsiveness to growth hormone (GH) replacement therapy in 665 GH-deficient adults. KIMS Study Group and the KIMS International Board. Journal of Clinical Endocrinology and Metabolism 1999 84 39293935. (doi:10.1210/jcem.84.11.6088).

    • Search Google Scholar
    • Export Citation
  • 3

    Johannsson G, Bjarnason R, Bramnert M, Carlsson LM, Degerblad M, Manhem P, Rosen T, Thoren M, Bengtsson BA. The individual responsiveness to growth hormone (GH) treatment in GH-deficient adults is dependent on the level of GH-binding protein, body mass index, age, and gender. Journal of Clinical Endocrinology and Metabolism 1996 81 15751581.

    • Search Google Scholar
    • Export Citation
  • 4

    Burman P, Johansson AG, Siegbahn A, Vessby B, Karlsson FA. Growth hormone (GH)-deficient men are more responsive to GH replacement therapy than women. Journal of Clinical Endocrinology and Metabolism 1997 82 550555.

    • Search Google Scholar
    • Export Citation
  • 5

    Attanasio AF, Bates PC, Ho KK, Webb SM, Ross RJ, Strasburger CJ, Bouillon R, Crowe B, Selander K, Valle D et al.. Human growth hormone replacement in adult hypopituitary patients: long-term effects on body composition and lipid status–3-year results from the HypoCCS Database. Journal of Clinical Endocrinology and Metabolism 2002 87 16001606.

    • Search Google Scholar
    • Export Citation
  • 6

    Feldt-Rasmussen U, Abs R, Bengtsson BA, Bennmarker H, Bramnert M, Hernberg-Stahl E, Monson JP, Westberg B, Wilton P, Wuster C. Growth hormone deficiency and replacement in hypopituitary patients previously treated for acromegaly or Cushing's disease. European Journal of Endocrinology 2002 146 6774. (doi:10.1530/eje.0.1460067).

    • Search Google Scholar
    • Export Citation
  • 7

    Rosilio M, Blum WF, Edwards DJ, Shavrikova EP, Valle D, Lamberts SW, Erfurth EM, Webb SM, Ross RJ, Chihara K et al.. Long-term improvement of quality of life during growth hormone (GH) replacement therapy in adults with GH deficiency, as measured by questions on life satisfaction-hypopituitarism (QLS-H). Journal of Clinical Endocrinology and Metabolism 2004 89 16841693. (doi:10.1210/jc.2003-030134).

    • Search Google Scholar
    • Export Citation
  • 8

    Verhelst J, Kendall-Taylor P, Erfurth EM, Price DA, Geffner M, Koltowska-Haggstrom M, Jonsson PJ, Wilton P, Abs R. Baseline characteristics and response to 2 years of growth hormone (GH) replacement of hypopituitary patients with GH deficiency due to adult-onset craniopharyngioma in comparison with patients with nonfunctioning pituitary adenoma: data from KIMS (Pfizer International Metabolic Database). Journal of Clinical Endocrinology and Metabolism 2005 90 46364643. (doi:10.1210/jc.2005-0185).

    • Search Google Scholar
    • Export Citation
  • 9

    Barbosa EJ, Koranyi J, Filipsson H, Bengtsson BA, Boguszewski CL, Johannsson G. Models to predict changes in serum IGF1 and body composition in response to GH replacement therapy in GH-deficient adults. European Journal of Endocrinology 2010 162 869878. (doi:10.1530/EJE-09-0973).

    • Search Google Scholar
    • Export Citation
  • 10

    Schneider HJ, Klotsche J, Wittchen HU, Stalla GK, Schopohl J, Kann PH, Kreitschmann-Andermahr I, Wallaschofski H, KIMS board and of the DETECT study . Effects of growth hormone replacement within the KIMS survey on estimated cardiovascular risk and predictors of risk reduction in patients with growth hormone deficiency. Clinical Endocrinology 2011 75 825830. (doi:10.1111/j.1365-2265.2011.04137.x).

    • Search Google Scholar
    • Export Citation
  • 11

    Feldt-Rasmussen U, Brabant G, Maiter D, Jonsson B, Toogood A, Koltowska-Haggstrom M, Rasmussen AK, Buchfelder M, Saller B, Biller BM. Response to GH treatment in adult GH deficiency is predicted by gender, age, and IGF1 SDS but not by stimulated GH-peak. European Journal of Endocrinology 2013 168 733743. (doi:10.1530/EJE-13-0005).

    • Search Google Scholar
    • Export Citation
  • 12

    Appelman-Dijkstra NM, Claessen KM, Roelfsema F, Pereira AM, Biermasz NR. Long-term effects of recombinant human GH replacement in adults with GH deficiency: a systematic review. European Journal of Endocrinology 2013 169 R114. (doi:10.1530/EJE-12-1088).

    • Search Google Scholar
    • Export Citation
  • 13

    Ranke MB, Lindberg A, Chatelain P, Wilton P, Cutfield W, Albertsson-Wikland K, Price DA. Derivation and validation of a mathematical model for predicting the response to exogenous recombinant human growth hormone (GH) in prepubertal children with idiopathic GH deficiency. KIGS International Board. Kabi Pharmacia International Growth Study. Journal of Clinical Endocrinology and Metabolism 1999 84 11741183. (doi:10.1210/jcem.84.4.5634).

    • Search Google Scholar
    • Export Citation
  • 14

    Ranke MB, Lindberg A, Albertsson-Wikland K, Wilton P, Price DA, Reiter EO. Increased response, but lower responsiveness, to growth hormone (GH) in very young children (aged 0–3 years) with idiopathic GH deficiency: analysis of data from KIGS. Journal of Clinical Endocrinology and Metabolism 2005 90 19661971. (doi:10.1210/jc.2004-1051).

    • Search Google Scholar
    • Export Citation
  • 15

    Dahlgren J, Kristrom B, Niklasson A, Nierop AF, Rosberg S, Albertsson-Wikland K. Models predicting the growth response to growth hormone treatment in short children independent of GH status, birth size and gestational age. BMC Medical Informatics and Decision Making 2007 7 40. (doi:10.1186/1472-6947-7-40).

    • Search Google Scholar
    • Export Citation
  • 16

    Gutierrez LP, Koltowska-Haggstrom M, Jonsson PJ, Mattsson AF, Svensson D, Westberg B, Luger A. Registries as a tool in evidence-based medicine: example of KIMS (Pfizer International Metabolic Database). Pharmacoepidemiology and Drug Safety 2008 17 90102. (doi:10.1002/pds.1510).

    • Search Google Scholar
    • Export Citation
  • 17

    Akobeng AK. Understanding diagnostic tests 3: Receiver operating characteristic curves. Acta Paediatrica 2007 96 644647. (doi:10.1111/j.1651-2227.2006.00178.x).

    • Search Google Scholar
    • Export Citation
  • 18

    Lemeshow S, Hosmer DW Jr. A review of goodness of fit statistics for use in the development of logistic regression models. American Journal of Epidemiology 1982 115 92106.

    • Search Google Scholar
    • Export Citation
  • 19

    Gönen M. Receiver Operating Characteristic (ROC) Curves. Proceedings of the Thirty-first Annual SAS® Users Group International Conference 2006 pp 210–231; Cary, NC: SAS Institute Inc

  • 20

    Sarkar SK, Midi H, Rana S. Detection of outliers and influential observations in binary logistic regression: an empirical study. Journal of Applied Sciences 2011 11 2635. (doi:10.3923/jas.2011.26.35).

    • Search Google Scholar
    • Export Citation
  • 21

    Arboretti Giancristofaro R , Salmaso L. Model performance analysis and model validation in logistic regression. Statistica 2003 2 375396.

 

     European Society of Endocrinology

Sept 2018 onwards Past Year Past 30 Days
Abstract Views 0 0 0
Full Text Views 900 55 1
PDF Downloads 358 101 9
  • View in gallery

    ROC curves for prediction of 2-year response by baseline QoL-AGHDA, waist circumference, total cholesterol, and IGF1 SDS (Model 1) and for prediction of 2-year response by baseline QoL-AGHDA, waist circumference and total cholesterol and 6-month delta waist circumference (Model 2).

  • View in gallery

    Sensitivity, specificity, positive predictive value and negative predictive value by response cutoff predictive probability for Model 1 (panel A) and Model 2 (Panel B), respectively. A patient is classified as a responder if the calculated model predictive probability is equal or greater than a given cutoff probability, and as a non-responder if the calculated predictive probability is lower than that, based on model 1 and 2, respectively. The procedure is repeated for all patients in the respective model data set and for some selected cutoff probablities. Different measures shown in the graphs were calculated comparing ‘predicted’ response (at baseline, model 1; and after 6 months of treatment, Model 2) with actual observed response after 2 years of GH treatment. Model 1 incorporates variables total QoL AGHDA score, waist circumference, total cholesterol, and IGF1SDS at baseline to predict 2 year response. Model 2 incorporates additional information on change in waist circumference after 6 months of GH treatment as predictor. As baseline IGF1SDS was no longer significant (P=0.66) as predictor after 6 months of GH treatment it was excluded from model 2.

  • 1

    Maison P, Griffin S, Nicoue-Beglah M, Haddad N, Balkau B, Chanson P. Impact of growth hormone (GH) treatment on cardiovascular risk factors in GH-deficient adults: a metaanalysis of blinded. randomized, placebo-controlled trials. Journal of Clinical Endocrinology and Metabolism 2004 89 21922199. (doi:10.1210/jc.2003-030840).

    • Search Google Scholar
    • Export Citation
  • 2

    Bengtsson BA, Abs R, Bennmarker H, Monson JP, Feldt-Rasmussen U, Hernberg-Stahl E, Westberg B, Wilton P, Wuster C. The effects of treatment and the individual responsiveness to growth hormone (GH) replacement therapy in 665 GH-deficient adults. KIMS Study Group and the KIMS International Board. Journal of Clinical Endocrinology and Metabolism 1999 84 39293935. (doi:10.1210/jcem.84.11.6088).

    • Search Google Scholar
    • Export Citation
  • 3

    Johannsson G, Bjarnason R, Bramnert M, Carlsson LM, Degerblad M, Manhem P, Rosen T, Thoren M, Bengtsson BA. The individual responsiveness to growth hormone (GH) treatment in GH-deficient adults is dependent on the level of GH-binding protein, body mass index, age, and gender. Journal of Clinical Endocrinology and Metabolism 1996 81 15751581.

    • Search Google Scholar
    • Export Citation
  • 4

    Burman P, Johansson AG, Siegbahn A, Vessby B, Karlsson FA. Growth hormone (GH)-deficient men are more responsive to GH replacement therapy than women. Journal of Clinical Endocrinology and Metabolism 1997 82 550555.

    • Search Google Scholar
    • Export Citation
  • 5

    Attanasio AF, Bates PC, Ho KK, Webb SM, Ross RJ, Strasburger CJ, Bouillon R, Crowe B, Selander K, Valle D et al.. Human growth hormone replacement in adult hypopituitary patients: long-term effects on body composition and lipid status–3-year results from the HypoCCS Database. Journal of Clinical Endocrinology and Metabolism 2002 87 16001606.

    • Search Google Scholar
    • Export Citation
  • 6

    Feldt-Rasmussen U, Abs R, Bengtsson BA, Bennmarker H, Bramnert M, Hernberg-Stahl E, Monson JP, Westberg B, Wilton P, Wuster C. Growth hormone deficiency and replacement in hypopituitary patients previously treated for acromegaly or Cushing's disease. European Journal of Endocrinology 2002 146 6774. (doi:10.1530/eje.0.1460067).

    • Search Google Scholar
    • Export Citation
  • 7

    Rosilio M, Blum WF, Edwards DJ, Shavrikova EP, Valle D, Lamberts SW, Erfurth EM, Webb SM, Ross RJ, Chihara K et al.. Long-term improvement of quality of life during growth hormone (GH) replacement therapy in adults with GH deficiency, as measured by questions on life satisfaction-hypopituitarism (QLS-H). Journal of Clinical Endocrinology and Metabolism 2004 89 16841693. (doi:10.1210/jc.2003-030134).

    • Search Google Scholar
    • Export Citation
  • 8

    Verhelst J, Kendall-Taylor P, Erfurth EM, Price DA, Geffner M, Koltowska-Haggstrom M, Jonsson PJ, Wilton P, Abs R. Baseline characteristics and response to 2 years of growth hormone (GH) replacement of hypopituitary patients with GH deficiency due to adult-onset craniopharyngioma in comparison with patients with nonfunctioning pituitary adenoma: data from KIMS (Pfizer International Metabolic Database). Journal of Clinical Endocrinology and Metabolism 2005 90 46364643. (doi:10.1210/jc.2005-0185).

    • Search Google Scholar
    • Export Citation
  • 9

    Barbosa EJ, Koranyi J, Filipsson H, Bengtsson BA, Boguszewski CL, Johannsson G. Models to predict changes in serum IGF1 and body composition in response to GH replacement therapy in GH-deficient adults. European Journal of Endocrinology 2010 162 869878. (doi:10.1530/EJE-09-0973).

    • Search Google Scholar
    • Export Citation
  • 10

    Schneider HJ, Klotsche J, Wittchen HU, Stalla GK, Schopohl J, Kann PH, Kreitschmann-Andermahr I, Wallaschofski H, KIMS board and of the DETECT study . Effects of growth hormone replacement within the KIMS survey on estimated cardiovascular risk and predictors of risk reduction in patients with growth hormone deficiency. Clinical Endocrinology 2011 75 825830. (doi:10.1111/j.1365-2265.2011.04137.x).

    • Search Google Scholar
    • Export Citation
  • 11

    Feldt-Rasmussen U, Brabant G, Maiter D, Jonsson B, Toogood A, Koltowska-Haggstrom M, Rasmussen AK, Buchfelder M, Saller B, Biller BM. Response to GH treatment in adult GH deficiency is predicted by gender, age, and IGF1 SDS but not by stimulated GH-peak. European Journal of Endocrinology 2013 168 733743. (doi:10.1530/EJE-13-0005).

    • Search Google Scholar
    • Export Citation
  • 12

    Appelman-Dijkstra NM, Claessen KM, Roelfsema F, Pereira AM, Biermasz NR. Long-term effects of recombinant human GH replacement in adults with GH deficiency: a systematic review. European Journal of Endocrinology 2013 169 R114. (doi:10.1530/EJE-12-1088).

    • Search Google Scholar
    • Export Citation
  • 13

    Ranke MB, Lindberg A, Chatelain P, Wilton P, Cutfield W, Albertsson-Wikland K, Price DA. Derivation and validation of a mathematical model for predicting the response to exogenous recombinant human growth hormone (GH) in prepubertal children with idiopathic GH deficiency. KIGS International Board. Kabi Pharmacia International Growth Study. Journal of Clinical Endocrinology and Metabolism 1999 84 11741183. (doi:10.1210/jcem.84.4.5634).

    • Search Google Scholar
    • Export Citation
  • 14

    Ranke MB, Lindberg A, Albertsson-Wikland K, Wilton P, Price DA, Reiter EO. Increased response, but lower responsiveness, to growth hormone (GH) in very young children (aged 0–3 years) with idiopathic GH deficiency: analysis of data from KIGS. Journal of Clinical Endocrinology and Metabolism 2005 90 19661971. (doi:10.1210/jc.2004-1051).

    • Search Google Scholar
    • Export Citation
  • 15

    Dahlgren J, Kristrom B, Niklasson A, Nierop AF, Rosberg S, Albertsson-Wikland K. Models predicting the growth response to growth hormone treatment in short children independent of GH status, birth size and gestational age. BMC Medical Informatics and Decision Making 2007 7 40. (doi:10.1186/1472-6947-7-40).

    • Search Google Scholar
    • Export Citation
  • 16

    Gutierrez LP, Koltowska-Haggstrom M, Jonsson PJ, Mattsson AF, Svensson D, Westberg B, Luger A. Registries as a tool in evidence-based medicine: example of KIMS (Pfizer International Metabolic Database). Pharmacoepidemiology and Drug Safety 2008 17 90102. (doi:10.1002/pds.1510).

    • Search Google Scholar
    • Export Citation
  • 17

    Akobeng AK. Understanding diagnostic tests 3: Receiver operating characteristic curves. Acta Paediatrica 2007 96 644647. (doi:10.1111/j.1651-2227.2006.00178.x).

    • Search Google Scholar
    • Export Citation
  • 18

    Lemeshow S, Hosmer DW Jr. A review of goodness of fit statistics for use in the development of logistic regression models. American Journal of Epidemiology 1982 115 92106.

    • Search Google Scholar
    • Export Citation
  • 19

    Gönen M. Receiver Operating Characteristic (ROC) Curves. Proceedings of the Thirty-first Annual SAS® Users Group International Conference 2006 pp 210–231; Cary, NC: SAS Institute Inc

  • 20

    Sarkar SK, Midi H, Rana S. Detection of outliers and influential observations in binary logistic regression: an empirical study. Journal of Applied Sciences 2011 11 2635. (doi:10.3923/jas.2011.26.35).

    • Search Google Scholar
    • Export Citation
  • 21

    Arboretti Giancristofaro R , Salmaso L. Model performance analysis and model validation in logistic regression. Statistica 2003 2 375396.