Growth is a good indicator of a child’s health, and growth failure prompts the pediatrician to search for nutrition disorders, subclinical chronic diseases, or hormone deficiencies. Growth hormone deficiency (GHD) is characterized by the insufficient production of growth hormone (GH), which leads to deficient insulin-like growth factor 1 (IGF1) synthesis and secretion leading to growth failure in children. An accurate diagnosis is crucial for timely initiation of treatment in order to optimize child growth and adult height and to avoid co-morbidities resulting in impaired quality of life (1).

Auxologic evaluation lies at the basis of the diagnosis of GHD in children (2–4), and the Growth Hormone Research Society clearly defined in its consensus guidelines released in year 2000 the clinical criteria that should prompt immediate investigation of GHD in childhood and adolescence (5). Many national endocrine societies have set up procedures to diagnose GHD, and the health authorities of several countries have established national or regional boards that review and monitor GH prescriptions (2, 6). Multiple GH stimulation tests (GHSTs) have been designed to evaluate GH sufficiency in children in an attempt to reach the most accurate diagnosis of GHD (7, 8). Although they have limitations, GHSTs are still used as the gold standard for the diagnosis of pediatric GHD in most countries (1, 8).

The guidelines of the US Pediatric Endocrine Society advocate for restricting GH testing (1). For instance, in newborns with hypoglycemia and/or neonatal cholestasis the diagnosis of GHD is an emergency, and GHSTs may be dangerous (9). In a child with growth failure, the presence of micropenis and cryptorchidism or of craniofacial midline abnormalities are other putative predictors of GH deficiency (1, 5). Acquired GHD may be suspected in patients with intracranial tumors, severe traumatic brain injury or cranial radiotherapy in whom a common co-morbidity is hypothalamic obesity, associated with blunted response during GHSTs (10). However, none of these studies provide predictive values that can guide medical decisions. Therefore, due to the lack of sufficient evidence, the guidelines cannot recommend establishing the diagnosis of GHD without GHSTs in patients with these conditions (1, 11).

Pharmacological GHSTs remain a standard practice in pediatric patients—usually due to requirements from health systems (6)—even in children with clearly identifiable potential risk factors for GHD (12–14) in whom the implementation of GHSTs might be considered redundant (1, 6, 15). The aim of the present study was to assess predictors of GHD in children with growth failure by analyzing a large cohort of pediatric patients in whom GHSTs had been performed in a tertiary referral center. Our primary objective was to develop and validate an accurate clinical prediction rule with high enough specificity to allow confirmation of the diagnosis of GHD in children without recurring to GHSTs. We, therefore, designed and validated a multivariable prediction model in accordance with the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) statement (16).

A total of 1,006 GHSTs were eligible for evaluation: 770 out of the 834 in the 2004–2014 cohort used to build the prediction model, and 161 of the 172 in the 2017–2019 cohort used to validate the model, could be analyzed ( Figure 1 ).

Clinical characteristics of the 2004–2014 cohort are summarized in Table 2 . It includes GHSTs performed in 491 boys and 279 girls (1.8:1), 79% were prepubertal, with a median age of 7.74 years and median height SDS of −2.51. Of the 770 GHSTs analyzed, 150 (19.5%) yielded results defined as GHD. The auxologic features were similar in patients without GHD (controls) and patients with GHD (cases), and within the latter, clinical features were similar in those predicted by the model and those not predicted. Both groups included congenital and acquired conditions ( Supplementary Table 4 ). An MRI, used to define “pituitary dysgenesis” ( Table 1 ) was available in 218 of the 770 patients (120 of 150 with GHD and 98 of 620 without GHD),

In this cohort of 770 patients, after tuning the classification model trained using the 15 potential dichotomic predictors, we selected 9 variables of interest, with an odds ratio (OR) >5 and a p-value <0.0001 ( Table 3 ); “cranial radiotherapy,” with an OR=4.4, was also selected given its clinical relevance. The continuous variables height, weight, and serum levels of IGF1 and IGFBP3 were not informative enough to be considered in the model ( Table 2 ). We used a gain ratio, entropy reducing algorithm to automatically build a decision tree from the 10 selected variables, without reduce error pruning or limits on minimum number of records. Anterior pituitary hormone (TSH, ACTH, prolactin) deficiencies were categorized as 0 (no deficiency), 1 (any one deficiency), or ≥1 (multiple pituitary hormone deficiency) ( Supplementary Figure 3 ). This model reached a 99.2% specificity ( Table 4 , “decision tree” column). Its sensitivity (49.3%) was above the required threshold, and overall accuracy was 89.5%, with a resulting F-measure = 0.646. Cross validation, with a test set of 20% of the cases randomly selected using stratified sampling, showed similar results, which remained consistent after repeating the random selection and changing the parameters (for instance, Gini index instead of gain ratio, or by using pruning). Alternative models, such as Naïve Bayes, also gave similar results.

The conditions “pituitary dysgenesis” and “≥1 anterior pituitary hormone (TSH, ACTH or prolactin) deficiency” were selected by the entropy reduction algorithms as the first variables to analyze. Interestingly, this selection was consistent with relevant clinical criteria. We also built a random forest (37) to derive an analysis of feature relevance, that reinforced the robustness of the set of conditions selected ( Supplementary Table 3 ). Finally, we also calculated conditional probabilities for the cases where two variables could give similar information. The resulting prediction rule stated that a patient, who met the criteria required for performing GHSTs according to the Summary Statement of the Growth Hormone Research Society (5), was a case (GHD associated to risk factors) if (s)he met the following conditions:

The proposed criteria were very conservative for specificity ( Table 4 , “prediction rule” column), and missed 89 cases of GHD of the 2004−2014 cohort, which could then be diagnosed using GHSTs. Therefore, the proposed predictive rule diagnosed 61 GHD cases, which represents 40.7% of all GHD patients.

We validated the predictive rule in an independent cohort of patients (2017−2019). Clinical characteristics of this second cohort are summarized in Table 5 . It includes GHSTs of 111 boys and 50 girls (2.2:1), 87% prepubertal, with a median age of 7.67 years and median height SDS of −2.57. This validation group was similar to the 2004−2014 cohort in terms of age, gender, or proportion of pathological tests compared, as well as in IGF1 serum levels. GHD was diagnosed in 36 patients (22.4%). An MRI was available in 46 of the 162 patients (27 of 36 with GHD and 19 of 126 without GHD), In this validation cohort (2017−2019), the prediction rule showed a specificity of 99.2%, and its positive predictive value was 95.2% ( Table 4 , “validation” column). The false-positive case according to our rule was a 10-year-old boy with height at −3.98 DS, IGF1 level 78 ng/ml (reference for age 60−370), who showed one peak GH level of 6.71 ng/ml at GHST, very close to the cutoff value, and no other remarkable feature. Given the severe growth deficiency, a therapeutic trial with GH treatment resulted in 1.06 DS gain in height after 1 year. GHD could not be ruled out in this patient despite the GHST result. The positive likelihood ratio of the prediction rule was 69.4, and the number needed to test with the rule was 1.19.

Finally, we tested significance following the methodology presented. The a priori probability of a GHD case in the second cohort was 36/161 (22.4%). We can safely assume that individual patients are independent cases, so we have a set of Bernoulli Trials under a binomial distribution. Thus, the p-value given by the cumulative function was <0.000001.

In this study, we identified clinically relevant risk factors for GHD in children, which were applied to build a robust clinical prediction rule to diagnose GHD, which could avoid resorting to GHSTs, in pediatric patients with growth failure and comorbidities. Our conclusion is based on the scientific evidence provided by the use of strict diagnostic criteria and clearly defined and accurately measured exposure variables in the analysis of a large cohort of GHSTs performed in a tertiary pediatric hospital.

The Endocrine Society clinical practice guideline on “Hypothalamic–Pituitary and Growth Disorders in Survivors of Childhood Cancer” advises using the same provocative testing to diagnose growth hormone deficiency in childhood cancer survivors as are used for diagnosing growth hormone deficiency in the non-cancer population as an ungraded good practice statement (11). On the other hand, the “Guidelines for the treatment of children with GHD, idiopathic short stature, and primary insulin-like growth factor 1 deficiency” of the Pediatric Endocrine Society (PES) suggest that in patients with auxological criteria, hypothalamic-pituitary defects and deficiency of at least one additional pituitary hormone, GHD diagnosis could potentially be established without performing GHSTs (1). However, due to the insufficient level of evidence according to the Grading of Recommendations, Assessment, Development, and Evaluation (GRADE) consensus (39), these recommendations could only be considered as conditional. Our study provides evidence to increase the strength of these recommendations.

We built a model on the knowledge generated by experts over the last 20 years (1, 5, 39)(and references therein), and applied a rigorous mathematical and machine-learning approach. Feature selection was based on the combination of statistical analyses with clinical criteria, used to refine the model in an iterative way until the prediction rule was satisfactory from both the statistical analysis and the clinical criteria, in a cohort of 770 GHSTs performed in our center between 2004 and 2014. Since the prediction rule was intended to diagnose GHD without the need for a GHST, and therefore type I errors would result in a false diagnosis of GHD leading to GH treatment in real world practice, we set goals of high specificity and positive predictive value for our rule. Of all the potential risk factors considered during model building, we identified the presence of pituitary dysgenesis on MRI or the existence of two or more anterior pituitary hormone deficiencies (TSH, ACTH, or prolactin) as specific enough to diagnose GHD without resorting to GHSTs in children meeting the criteria required for GHST by the Summary Statement of the Growth Hormone Research Society (5). Alternatively, if only one (TSH, ACTH, or prolactin) deficiency was present, the coexistence of central diabetes insipidus, neonatal symptoms of pituitary deficiency (hypoglycemia or hypogenitalism), sellar or suprasellar surgery or tumor (excluding microadenomas), clinical or radiological craniofacial midline abnormalities, or cranial radiotherapy ≥18 Gy, also led to a safe diagnosis. Gonadotropin deficiency was not considered in the analysis because its ascertainment may prove challenging in prepubertal patients.

To test the clinical applicability of the prediction rule, we validated our results using an independent cohort of 161 GHSTs performed between 2017 and 2019. Auspiciously, specificity was 99.2% in this second cohort, supporting the safety of our prediction rule. As expected, sensitivity was relatively low, reaching 55.6% in the validation cohort, indicating that almost half of the patients with GHD would only be identified after referring to GHSTs. Nonetheless, the sensitivity of the prediction rule applied to children meeting the criteria required for GHST by the Summary Statement of the Growth Hormone Research Society (5) would reduce in approximately half of the cases with GHD the need to perform a relatively invasive endocrine test, which underscores the clinical relevance of our results.

An unexpected, clinically relevant result is that only 20% of the children undergoing provocative tests, due to a suspicion of GHD, proved to be GH deficient (150 of 770 in the first cohort and 36 of 161 in the validation cohort). This is particularly significant given that only patients meeting the rigorous criteria defined by the Growth Hormone Research Society for prescribing GHSTs to children (5) were included in our study. This may be explained by the stringent criteria used to ascertain GHD in our center. Indeed, the diagnosis of GHD was based on peak GH levels <6.1 ng/ml between 2004 and 2011 (17), or <4.7 ng/ml between 2012 and 2019 (18), according to previously validated cutoff values (19).

Key strengths of this study are the high number of patients included in the construction of the predictive model as well as in the independent validation sample. It should also be stressed that strict criteria were used to define cases and controls: as mentioned above, the diagnosis of GHD was based on stringent cut-off levels for peak GH in GHSTs. A meticulous analysis of inclusion and exclusion criteria was performed to avoid inclusion bias. The population sample is representative of patients seeking advice from pediatric endocrinologists at referral centers for the assessment of short stature, which renders our results widely applicable.

Our study also has some limitations related to its design. Frequently, observational studies are prone to missing information in their datasets. To minimize memory bias, we limited the assessment of potential predictors to risk factors that were routinely reported in the clinical charts or were available from electronic records of endocrine laboratory or imaging study in our hospital. However, we cannot exclude that risk factors have been missed and, therefore, not included in the prediction rule we generated. Particularly, very few patients had a confirmed genetic diagnosis. This may explain why the condition “familial or sporadic GHD of genetic etiology” was not prioritized by our model. Also, MRI studies were not available in all patients; nonetheless, false positive prediction of GHD due to pituitary dysgenesis did not occur in any of the 98 controls (no GHD, Table 3 ) who underwent MRI.

In summary, this study developed an algorithm that led to the construction of a predictive rule for decision-making in the diagnosis of GHD in children typically seeking advice from pediatricians for growth failure, on the basis of a reduced number of clinically relevant and easily identifiable risk factors. The application of this rule avoids the need for GHSTs in a significant proportion of the patients in whom testing to assess GHD is presently indicated, and it is especially important for a subgroup of labile or vulnerable patients, such as infants, very low weight patients and children with oncological conditions or other comorbidities.

The original contributions presented in the study are included in the article/ Supplementary Material . Further inquiries can be directed to the corresponding author.

The studies involving human participants were reviewed and approved by Comité de Ética en Investigación, Hospital de Niños Ricardo Gutiérrez, Buenos Aires, Argentina. Written informed consent from the participants’ legal guardian/next of kin was not required to participate in this study in accordance with the national legislation and the institutional requirements.

FC, RG, DY, IB, GF, and RR conceived the study and designed the analysis plan. FC, DY, and RR did the statistical analyses. FC, RG, DY, GF, and RR wrote the manuscript. SMB, MS, MR, MB, AK, DB, PP, IB, and GF contributed to obtain and interpret the data. All authors contributed to the article and approved the submitted version.

This work was partially supported by funding from: CONICET (Consejo Nacional de Investigaciones Científicas y Técnicas) grant PIP-11220130100687, and FONCYT (Fondo para la Investigación Científica y Tecnológica) grant PICT-2014-2490, Argentina.

DY is a founding partner of Practia S.A. GF is employed by Takeda Pharmaceutical. RR and IB report lecture honoraria from Novo Nordisk and non-financial support from Biosidus, Merck, Novo Nordisk, Pfizer, and Sandoz. RG reports lecture honoraria from Novo Nordisk and Raffo, and non-financial support from Merck, Novo Nordisk, Pfizer, and Sandoz.

The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.