Exploratory machine learning approach in routine steroid profiles of rare adrenal disorders

Alexandra E. Kullei; Robert Markewitz; Paul-Martin Holterhus

doi:10.1530/endoabs.118.PO27

PO27

Prev Next

Section Contents Cite

Endocrine Abstracts (2026) 118 PO27 | DOI: 10.1530/endoabs.118.PO27

IDSD2026 Poster Abstracts Poster Abstracts (93 abstracts)

Exploratory machine learning approach in routine steroid profiles of rare adrenal disorders

Alexandra E. Kullei ^1,2 , Robert Markewitz ² & Paul-Martin Holterhus ^1,2

Author affiliations

¹Department of Paediatric Oncology and Rheumatology; Paediatric Endocrinology and Diabetes, University Hospital of Schleswig-Holstein, UKSH, Campus Kiel, Germany; ²Institute of Clinical Chemistry, University Hospital of Schleswig-Holstein, Kiel/Lübeck, Germany. Correspondence to: [email protected]

Background: Routine steroid profiling is an established part of the diagnostic workup of rare adrenal disorders. While individual steroid markers are well characterized, the combined interpretation of multi analyte profiles remains challenging in clinical practice. Building on previous work by Zalas et al., this study explores the potential of machine learning methods to extract diagnostic patterns from measured steroid panels.

Methods: Steroid profiles from individuals with 21 hydroxylase deficiency (21OHD, n = 40), 11 hydroxylase deficiency (11OHD, n = 10), adrenocortical carcinoma (ACC, n = 10), and obese persons (n = 60) were analyzed. All steroid concentrations were Z log transformed. A Random Forest model was used as an exploratory tool to assess group separation and identify informative steroid markers. Feature importance, proximity analysis, and surrogate decision trees were applied to support interpretation.

Results: The Random Forest analysis revealed distinct steroid patterns across the diagnostic groups. 21OHD was characterized by elevated 17 hydroxyprogesterone and androstenedione. 11OHD showed increased 11 deoxycortisol and 11 deoxycorticosterone. ACC displayed broader and more variable alterations, involving multiple precursors and downstream metabolites. Obese persons showed comparatively uniform profiles with lower variability. Across all groups, the model consistently highlighted combinations of precursors and downstream metabolites that contributed to separation and aligned with known biosynthetic pathways. Proximity and surrogate tree analyses indicated subgroup structures and occasional atypical profiles within the dataset.

Conclusions: This exploratory work suggests that data driven analytical approaches may complement the interpretation of routine steroid profiles in endocrine disorders. While not intended to replace established diagnostic reasoning, such methods can help uncover multidimensional biochemical signatures, support hypothesis generation, and inform future analytical frameworks for complex steroid datasets.