ESPE Abstracts (2015) 84 P-2-418

ESPE2015 Poster Category 2 GH & IGF (40 abstracts)

Random Forest Classification Predicts Response to Recombinant GH in GH Deficient Children Using Baseline Clinical Parameters and Genetic Markers

Adam Stevens a , Philip Murray a , Jerome Wojcik b , John Raelson c , Ekaterina Koledova d , Pierre Chatelain e & Peter Clayton a


aFaculty of Medical and Human Sciences, Institute of Human Development, University of Manchester and Manchester Academic Health Science Centre, Royal Manchester Children’s Hospital, Manchester, UK; bQuartz Bio, Geneva, Switzerland; cGenizon BioSciences, St Laurent, Quebec, Canada; dMerck Serono, Darmstadt, Germany; eDepartment Pediatrie, Hôpital Mère-Enfant – Université Claude Bernard, Lyon, France


Background: Prediction of response to recombinant GH (r-GH) is currently based on regression modelling. This approach generates a prediction equation which can be applied to data from an individual child. However this method can underestimate the effect of inter-dependent variables. Random forest classification (RFC) is an alternative prediction method based on decision trees that is not sensitive to the relationships between variables.

Objective and hypotheses: To assess the predictive value of RFC in GH deficient (GHD) children.

Method: We used pre-pubertal GHD children (peak GH (GH) <10 μg/l) from the PREDICT long-term follow-up study (NCT00699855, n=113) and the PREDICT validation study (NCT01419249, n=293). Single nucleotide polymorphisms (SNP) associated with 1st year growth response to r-GH (n=22) were genotyped. RFC was undertaken to identify variables associated with growth response (change in height (cm)) using the baseline clinical variables of gender, age, GH peak, GH dose, distance to target height SDS and mid-parental height SDS. Accuracy ((true positives+true negatives)/total population) of the RFC models was assessed and a variable importance score (VIS) calculated by permutation.

Results: RFC demonstrated that basal clinical variables could predict growth response with high accuracy (80.6%, P<1.1×10−39). The variables were ranked by VIS as follows: 1/GH peak 2/gender 3/age 4/mid-parental height SDS and 5/distance to target height SDS. The addition of genetic data could not improve prediction (accuracy 80.7%, P<2.8×10−38); however SNPs alone could act as weak but distinct predictors of growth response, accuracy (accuracy 65.4%, P<1.9×10−13). The SNPs with predictive value were rs1024531 (GRB10) and rs7101 (FOS).

Conclusion: The Ranke regression model1 predicts 65% of the variability in first year response in GHD with GH peak as the most significant variable. RFC also predicts response and identifies GH peak as the most important variable. Interestingly, two genetic markers alone can provide a level of prediction.

Conflict of interest: Dr Adam Stevens has received honoraria as an investigator from Merck Serono.

Funding: The PREDICT study was supported by Merck Serono S.A - Geneva, Switzerland.

Article tools

My recent searches

No recent searches.