Background: To assess the diagnostic value of combining plasma steroid profiling with machine learning (ML) in differentiating between mild autonomous cortisol secretion (MACS) and nonfunctioning adenoma (NFA) in patients with adrenal incidentalomas.
Methods: The plasma steroid profiles data in the laboratory information system were screened from January 2021 to December 2023. EXtreme Gradient Boosting (XGBoost) was applied to establish diagnostic models using plasma 24-steroid panels and/or clinical characteristics of the subjects. The SHapley Additive exPlanation (SHAP) method was used for explaining the model.
Results: 76 patients with MACS and 86 patients with NFA were included in the development and internal validation cohort while the external validation cohort consisted of 27 MACS and 21 NFA cases. Among five ML models evaluated, XGBoost demonstrated superior performance with an AUC of 0.77 using 24 steroid hormones. The SHAP method identified five steroids that exhibited optimal performance in distinguishing MACS from NFA, namely dehydroepiandrosterone (DHEA), 11-deoxycortisol, 11β-hydroxytestosterone, testosterone, and dehydroepiandrosteronesulfate (DHEAS). Upon incorporating clinical features into the model, the AUC increased to 0.88, with a sensitivity of 0.77 and specificity of 0.82. Furthermore, the results obtained through SHAP revealed that lower levels of testosterone, DHEA, LDL-c, BMI, and ACTH along with higher level of 11-deoxycortisol significantly contributed to the identification of MACS in the model.
Conclusions: We have elucidated the utilization of ML-based steroid profiling to discriminate between MACS and NFA in patients with adrenal incidentalomas. This approach holds promise for distinguishing these two entities through a single blood collection.
Keywords: adrenal incidentaloma; liquid chromatography-tandem mass spectrometry; machine learning; metabolomics; mild autonomous cortisol secretion; steroid profiling.
Copyright © 2024. Published by Elsevier Inc.