A machine learning model using routinely available biochemical test results differentiated Cushing disease from ectopic Cushing syndrome, with discrimination improving when pituitary magnetic resonance imaging findings were added, according to a retrospective study published in the European Journal of Endocrinology. 

The study evaluated whether routine baseline biochemical tests used to diagnose Cushing syndrome could also help differentiate its adrenocorticotropin (ACTH)-dependent etiologies. Investigators compared individual biochemical tests, a composite scoring system, and a balanced random forest (BRF) machine learning model, with and without pituitary magnetic resonance imaging (MRI).  

Investigators reviewed 509 patients with Cushing disease and 57 with ectopic Cushing syndrome diagnosed between 1994 and 2025 at six centers in Germany, Austria, and Italy. Diagnoses were confirmed by histopathology in 474 patients and by postoperative adrenal insufficiency or normalization of cortisol excess in 92.  

The biochemical inputs were morning plasma ACTH, morning and midnight serum cortisol, the 1-mg dexamethasone suppression test, 24-hour urinary free cortisol (UFC), and late-night salivary cortisol. Because participating centers used different assays and reference ranges, test results were expressed as multiples of the upper limit of normal. 

For the machine learning analysis, investigators used a BRF classifier, an ensemble of decision trees adapted to account for the imbalance between the 509 patients with Cushing disease and 57 with ectopic Cushing syndrome. The model handled missing values internally and addressed class imbalance by oversampling observed minority-class cases within each bootstrap sample. 

Investigators evaluated the model using leave-one-center-out cross-validation, training it on data from all but one center and testing it on the held-out center, with each center serving as the test set once. The study did not use a separate external validation cohort.  

At the cutoff selected using Youden's index, the BRF model using biochemical tests had 71% sensitivity and 84% specificity, with an area under the receiver operating characteristic curve (AUROC) of 0.853. When pituitary MRI findings were incorporated, sensitivity increased to 76% and specificity to 91%, with an AUROC of 0.902. 

Feature-importance analysis identified 24-hour UFC as the most influential predictor in the biochemical model, followed by morning serum cortisol, morning plasma ACTH, and the 1-mg dexamethasone suppression test. When MRI was included, 24-hour UFC ranked first and MRI second, followed by the 1-mg dexamethasone suppression test. 

Investigators separately developed a composite score for patients with at least three of the six biochemical tests available. Each result received zero, one, or two points according to whether it fell below the 25th percentile, between the 25th and 75th percentiles, or above the 75th percentile. The points were summed and divided by the number of tests available to calculate a mean biochemical composite score. 

Among 488 patients with at least three available tests, the biochemical composite score had an AUROC of 0.865. At an optimal cutoff of 1.5, sensitivity was 78% and specificity was 92%. Unlike the balanced random forest model, however, the composite score was fitted and evaluated in the same dataset without a held-out validation set. The authors said its reported performance therefore reflects in-sample accuracy rather than generalizability, and that its performance in an entirely new cohort remains unknown.  

Pituitary MRI findings and at least three biochemical tests were available for 445 patients. Investigators assigned zero points for a visible pituitary lesion larger than 6 mm, one point for a lesion smaller than 6 mm or an equivocal finding, and two points when no lesion was visible. Combining the biochemical and MRI scores yielded an AUROC of 0.931, with sensitivity of 88% and specificity of 85%. 

When the approaches were compared within the same cohort, only the composite score combined with MRI showed significantly greater discrimination than 24-hour UFC alone. 

Among the individual biochemical tests, 24-hour UFC showed the highest discrimination, with an AUROC of 0.854. At the optimal cutoff of 5.9 times the upper limit of normal, sensitivity was 72% and specificity was 83%. All 122 patients with UFC below 1.7 times the upper limit of normal had Cushing disease, while among those with values above 5.2 times the upper limit of normal, 34 (29%) had ectopic Cushing syndrome. 

The 1-mg dexamethasone suppression test had an AUROC of 0.828, with 70% sensitivity and 86% specificity at an optimal cutoff of 12.8 times the upper limit of normal. Late-night salivary cortisol had an AUROC of 0.802, with 60% sensitivity and 90% specificity, while morning plasma ACTH had an AUROC of 0.776, with 74% sensitivity and 77% specificity. 

A sensitivity analysis restricted to patients with histopathological confirmation showed no material differences from the overall cohort, which also included patients whose diagnoses were confirmed using postoperative biochemical criteria. 

The authors cited several study limitations. The study was retrospective, biochemical test availability varied among patients, centers used different assays and reference ranges, and pituitary MRI scans were interpreted by different radiologists, introducing potential measurement and observer variability. The authors also noted that the BRF carries an inherent risk of overfitting and that its greater complexity comes at the cost of reduced interpretability compared with the composite score. 

The authors concluded that composite scoring and machine learning may offer additional approaches for differentiating ACTH-dependent Cushing syndrome, particularly where corticotropin-releasing hormone testing is unavailable, but said both require further validation before routine use. "Both approaches require validation in larger and prospective studies before they can be considered for integration into routine clinical decision-making,” they wrote.  

The authors declared no conflicts of interest.