Machine learning methods in differential diagnosis of ACTH-dependent hypercortisolism

• Published in: Problemy ėndokrinologii 2024-02, Vol.70 (1), p.18-29
• Main Authors: Golounina, O. O., Belaya, Zh. E., Voronov, K. A., Solodovnikov, A. G., Rozhinskaya, L. Ya, Melnichenko, G. A., Mokrysheva, N. G., Dedov, I. I.
• Format: Article
• Language: eng ; rus
• ISSN: 0375-9660; 2308-1430
• DOI: 10.14341/probl13342

AIM : To develop a noninvasive method of differential diagnosis of ACTH-dependent hypercortisolism, as well as to evaluate the effectiveness of an optimal algorithm for predicting the probability of ectopic ACTH syndrome (EAS) obtained using machine learning methods based on the analysis of clinical data. MATERIALS AND METHODS : As part of a single-center, one-stage, cohort study, a retrospective prediction of the probability of EAS among patients with ACTH-dependent hypercortisolism was carried out. Patients were randomly stratified into 2 samples: training (80%) and test (20%). Eleven machine learning algorithms were used to develop predictive models: Linear Discriminant Analysis, Logistic Regression, elastic network (GLMNET), Support Vector machine (SVM Radial), k-nearest neighbors (kNN), Naive Bayes, binary decision tree (CART), C5.0 decision tree algorithms, Bagged CART, Random Forest, Gradient Boosting (Stochastic Gradient Boosting, GBM). RESULTS : The study included 223 patients (163 women, 60 men) with ACTH-dependent hypercortisolism, of which 175 patients with Cushing’s disease (CD), 48 — with EAS. As a result of preliminary data processing and selection of the most informative signs, the final variables for the classification and prediction of EAS were selected: ACTH level at 08:00 hours, potassium level (the minimum value of potassium in the active stage of the disease), 24-h urinary free cortisol, late-night serum cortisol, late-night salivary cortisol, the largest size of pituitary adenoma according to MRI of the brain. The best predictive ability in a training sample of all trained machine learning models for all three final metrics (ROC-AUC (0.867), sensitivity (90%), specificity (56.4%)) demonstrated a model of gradient boosting (Generalized Boosted Modeling, GBM). In the test sample, the AUC, sensitivity and specificity of the model in predicting EAS were 0.920; 77.8% and 97.1%, respectively. CONCLUSION : The prognostic model based on machine learning methods makes it possible to differentiate patients with EAS and CD based on basic clinical results and can be used as a primary screening of patients with ACTH-dependent hypercortisolism.