Baseline biochemical evaluation for etiologic differentiation in ACTH-dependent Cushing's syndrome: machine learning and composite score analysis in a multicenter cohort of 566 patients.

BACKGROUND

Limited availability of corticotropin-releasing hormone (CRH) currently complicates the differentiation of adrenocorticotropin (ACTH)-dependent Cushing's syndrome (CS). The diagnostic value of common screening tests in distinguishing Cushing's disease (CD) from ectopic CS (ECS) remains unclear.

OBJECTIVE

To assess the diagnostic performance of screening tests, alone and in combination, in differentiating CD from ECS.

METHODS

Retrospective multicenter study enrolling patients with confirmed ACTH-dependent CS and available screening tests at diagnosis. Data are expressed as multiples of upper limit of normal (×ULN). Optimal cut-offs were determined using Youden's Index. Combination with composite score models and machine learning algorithm were performed.

RESULTS

A total of 566 patients were included (509 [90%] with CD). The optimal morning ACTH cut-off was 1.8×ULN (sensitivity 74%, specificity 77%, AUC=0.776 [CI-95% .688-.853]). 24h-urinary free cortisol (24h-UFC) showed the best performance (cut-off 5.9×ULN, sensitivity 72%, specificity 83%, AUC=0.854 [.816-.923]), followed by the 1 mg dexamethasone suppression test (12.8×ULN, sensitivity 70%, specificity 86%, AUC=.828 [.740-.906]). Using a composite score, a cut-off of 1.5 yielded 78% sensitivity and 92% specificity (AUC=0.865 [.798-.933]). Combining this score with findings derived from pituitary magnetic resonance imaging (MRI), sensitivity and specificity were 88% and 85% (AUC=0.931 [.887-.974]). Machine learning algorithm (balanced random forest) including screening tests-results yielded a sensitivity of 71% and specificity of 84% (AUC=0.853 [.741-.966]), while including the MRI-findings variant achieved a sensitivity and specificity of 76% and 91% (AUC=0.902 [.805-.999]).

CONCLUSION

Combining screening tests in composite scoring and machine learning differentiates subtypes of ACTH-dependent CS and offers a potential diagnostic alternative to the CRH stimulation test.