Machine Learning-Based Decision Tree Model for the Diagnosis of Congenital Disorders of Adrenal Steroidogenesis Using LC-MS/MS-based Plasma Steroid Hormone Profiles

Erçetin Atam Noyan; Tosun Busra Gurpinar; Tarik Kirkgoz; Abali Zehra Yavas; Mehmet Eltan; Azad Akbarzade; Ali Yaman; Serap Turan; Abdullah Bereket; Arga Kazim Yalcin; Tulay Guran

RFC4.6

Prev Next

Section Contents Cite

ESPE Abstracts (2024) 98 RFC4.6

ESPE2024 Rapid Free Communications Adrenals and HPA Axis 1 (6 abstracts)

Machine Learning-Based Decision Tree Model for the Diagnosis of Congenital Disorders of Adrenal Steroidogenesis Using LC-MS/MS-based Plasma Steroid Hormone Profiles

Atam Noyan Erçetin ¹ , Busra Gurpinar Tosun ² , Tarik Kirkgoz ² , Zehra Yavas Abali ² , Mehmet Eltan ² , Azad Akbarzade ² , Ali Yaman ³ , Serap Turan ² , Abdullah Bereket ² , Kazim Yalcin Arga ^1,4,5 & Tulay Guran ²

Views

Author affiliations

¹Department of Bioengineering, Marmara University, Istanbul, Turkey. ²Department of Pediatric Endocrinology and Diabetes, Marmara University, Istanbul, Turkey. ³Department of Biochemistry, Marmara University, Istanbul, Turkey. ⁴Genetic and Metabolic Diseases Research and Investigation Center, Marmara University, Istanbul, Turkey. ⁵Health Biotechnology Joint Research and Application Center of Excellence, Istanbul, Turkey

Background: Congenital disorders of adrenal steroidogenesis (CDAS) result in various adverse clinical outcomes including adrenal insufficiency, hypertension, androgen excess, or differences in sex development, depending on the specific enzyme deficiency involved. Rapid and precise diagnosis of these conditions facilitates improved short-term and long-term clinical management. The requirement for clinical expertise may lead to delays in the diagnostic process.

Aim: We aimed to develop a machine learning-based decision tree model to facilitate the differential diagnosis of nine CDAS using plasma steroid hormone profiles.

Participants and Methods: Our study cohort included data from healthy individuals (n = 702) and patients genetically diagnosed with one of 9 CDAS (n = 328). Eighteen plasma steroids measured simultaneously as a panel by liquid chromatography-mass spectrometry (LC-MS/MS) were included in the analyses. A machine-learning-based decision tree model was constructed to classify CDAS subtypes. The LightGBM algorithm was used to identify the most discriminative hormone markers, and cut-off scores were estimated via logistic regression for each decision step of the model.

Results: The overall accuracy of the model in distinguishing healthy individuals from patients was 97%, specificity was 93.7%, and sensitivity was 99.6%. The model performance was assessed using cross-validation techniques and achieved high accuracy (98.16% to 100%), sensitivity (65.4% to 100%), and specificity (89% to 100%) in discriminating CDAS subtypes. In particular, our approach enabled the identification of important steroid hormone profiles that contribute significantly to the classification of CDAS.

Conclusion: Our machine learning-based decision tree model is a promising tool for the differential diagnosis of CDAS, helping clinicians to treat patients in a timely and accurate manner. Furthermore, our results emphasize the potential of using steroid hormone profiles in conjunction with advanced computational methods to improve diagnostic accuracy in rare endocrine diseases such as CDAS.