Comparison of the Performances of Generalized Linear Model, Decision Tree, Random Forest, Naive Bayes, and Support Vector Machine Models From Machine Learning-Based Classification Algorithms in the Diagnosis of Cushing's Syndrome.

AIMS/

BACKGROUND

Various tests are used in the screening and diagnosis of Cushing's syndrome (CS). However, none of the existing approaches provide optimal sensitivity and specificity. This study aimed to analyze the accuracy of various tests used for diagnosing CS, including free cortisol measurements in urine, dexamethasone suppression tests (DSTs; 1, 2, and 8 mg), and nocturnal salivary cortisol measurements. The study aimed to identify the most accurate tests using advanced machine learning (ML) techniques.

METHODS

We performed binary classifications using data from 278 patients with CS and 220 controls without CS. We developed five ML classification models, namely, a generalized linear model, decision tree, random forest, naive Bayes, and support vector machine, and compared them using standard performance metrics. We employed Boruta feature selection to identify the most informative clinical and biochemical predictors.

RESULTS

The naive Bayes model achieved the highest predictive performance, with 100% classification accuracy and a kappa value of 1.0000, followed by the random forest and decision tree models, each with 99.01% accuracy. Boruta feature selection ranked as the most informative predictor post 2-mg DST cortisol, followed by midnight cortisol, post 8-mg DST cortisol, urinary free cortisol, and post 1-mg DST cortisol. The decision tree model identified a cohort-specific post 2-mg DST cortisol threshold of approximately 1.78 µg/dL for distinguishing CS-positive from CS-negative cases.

CONCLUSIONS

ML models, particularly naive Bayes, exhibited strong internal performance in distinguishing patients with CS from non-CS controls using routinely assessed biochemical parameters. The prominence of the post 2-mg DST cortisol predictor in the Boruta ranking and decision tree analysis indicates that this variable contributed substantially to the model-based classification in this cohort. However, the identified 1.78 µg/dL threshold should be interpreted as a cohort-specific, data-driven finding rather than as a clinical cutoff recommendation, and external validation in larger multicenter cohorts is required before clinical implementation.