Endocrine Abstracts

Background: The gradual increase in the detection rate of adrenal incidentalomas makes them a common clinical problem. The vast majority of them are benign adrenocortical adenomas. Nevertheless every patient with adrenal incidentaloma requires performing number of tests to exclude pheochromocytoma, autonomous cortisol secretion, adrenal carcinoma and primary hyperaldosteronism. Evaluation of whether adrenal incidentalomas are malignant or functional and continuing patient follow-up to assess the necessity for surgery assumed important place in endocrinology practice.

Objective: The aim of the study was to compare several machine learning techniques in a qualification for a surgical treatment of adrenal tumors and choose the most accurate algorithm as a valuable adjunct tool for decision-making.

Methods: A retrospective, single-center study was performed on hospitalized patients with adrenal incidentaloma between 2017 and 2019. From a database comprising 264 patients with adrenal incidentaloma, clinical data for 30 patients who underwent adrenalectomy due to suspicion of primary aldosteronism, pheochromocytoma, Cushing’s syndrome, or adrenal cancer were extracted. All included patients underwent the endocrine work-up aimed to study the hormonal status of adrenal incidentalomas and every adrenal lesion was assessed with CT scan. On the basis of postoperative histopathological examinations, proper qualifications were confirmed in 20 out of 30 selected patients. Several machine learning algorithms, including Support Vector Machine, Multilayer Neural Network, C4.5 Decision Tree, Random Forest, k-Nearest Neighbours, Naïve Bayes, Zero R, One Rule, Logistic Regression, were trained to qualify the patients for an adrenalectomy. Finally, attribute selection technique was used to assess their usefulness in classification.

Results: The highest average accuracy was obtained for Support Vector Machine with linear kernel and soft margin – 90% of properly classified subjects. The Neural Network gave the second best result and was able to classify with an accuracy of 86%. Statistical evaluation using Pair-T Student modified for dependent samples was significantly better in comparison to baseline approach Zero-R (P<0.05). The most commonly selected by classifiers attributes were tumor homogeneity (100%), maximum diameter of the tumor (100%) and obesity (98%). Nevertheless prior attribute selection did not improve accuracy of trained algorithms.

Conclusions: Presented results show that application of machine learning methods in qualifying patients for an adrenalectomy may improve the decision process. The new training machine learning-based methods might be used to simplify making therapeutic decisions in adrenal incidentaloma patients and reduce the time from the initial identification of adrenal incidentaloma to the final decision about surgery.