ESPE Abstracts

Introduction: Klinefelter syndrome (KS) is the most common sex-chromosome disorder and cause of infertility and hypogonadism in males. However, KS remains an underdiagnosed condition with the majority of expected cases escaping clinical diagnosis and follow-up. Generally, the mid-puberty endocrine profile associated with KS is characterized by elevated levels of gonadotropins due to diminished testosterone feedback.

Objective: To devise a machine learning model to identify cases of KS amongst healthy controls based on inter-individual endocrine profiles, and test its predictive value.

Methods: The current project used hospital records of hormone levels in n = 15 genomically confirmed and untreated cases of KS (47,XXY) in the age interval 6 - 13 years. Sex and age-adjusted hormone z-score equivalents of serum follicle-stimulating hormone (FSH), luteinizing hormone (LH), anti-Müllerian hormone (AMH), dehydroepiandrosterone sulfate (DHEAS), sex hormone-binding globulin (SHBG), 4-androstenedione, testosterone and 17-hydroxyprogesterone constituted the current ‘endocrine profile’. The random forest machine learning algorithm was trained and applied to classify cases of KS relative to the endocrine profiles of n = 107 healthy, prepubertal (testicular vol. < 4 ml) and age-matched boys. The machine learning model was trained using 75% of the total observations and subsequently applied to classify the remaining test data using the ‘randomForest’ and ‘caret’ packages in R.

Results: Classification performance of the machine learning model exhibited an accuracy of 0.97 (95% CI, 0.96 to 0.98), Kappa of agreement of 0.92 (95% CI, 0.89 to 0.94) and average sensitivity of 0.97 and specificity of 0.96. Corresponding confusion matrices confirmed the generally correct classification of KS and healthy subjects. Serum levels of FSH for age were the best singular hormone to classify KS, but this marker exhibited an accuracy of merely 60% in this respect. Clustering between healthy and KS endocrine profiles were also observed by principal component analysis (PCA), where LH and FSH levels primarily contributed to the variance between the two groups.

Conclusion: Machine learning applied to biochemical data was able to make valid predictions to accurately classify cases of KS in this relatively small retrospective cohort. The model may be improved by addition of anthropometric (e.g. height-for-age) and other biochemical markers with relevance for the KS phenotype.