Endocrine Abstracts

Background: Adrenal Incidentaloma’s (AI) incidence has been increased 10-fold in the past two decades and more specifically, up to 7% in the general population and up to 9% in elderly patients. Determining whether an adrenal mass is malignant or hormonally active is of great clinical and socioeconomical importance, in order to select the best medical approach and follow up. Artificial Intelligence techniques could provide useful clinical tools for predicting the nature of AI with acceptable accuracy.

Methods: A total of 96 patients diagnosed with AI were selected from two tertiary centres. Their clinical data were collected and randomly grouped into training and validation sets in a ratio of 4:1. Two different experimental approaches were adopted. The first approach, concerned the analysis of twelve routine clinical features [functionality, age at diagnosis, hypokalemia, lateralisation, size max, baseline cortisol morning levels, adrenocorticotropic hormone (ACTH), 17-OH progesterone levels, 24 h urinary free cortisol, 1 mg overnight dexamethasone suppression cortisol levels, Hb1Ac, Computed Tomography imaging (baseline non-contrast Hounsfield)] for developing a predictive model for benign and non-benign adrenal tumors. The second approach included eight more features [age, gender, hypertension, hyperlipidaemia, anti-lipidemic therapy, Diabetes Mellitus (DM), anti-DM therapy, dehydroepiandrosterone (DHEAS)] for classifying patients into the same groups (benign and non-benign tumors). Eleven Machine Learning (ML) models were applied and tested in each of the above-mentioned experimental approaches to construct the best performing predictive model. In order to evaluate the performance of the proposed classification systems, precision, sensitivity, F1-score and accuracy were examined.

Results: The first tested Machine Learning model, based on the combination of 12 features, resulted in predictions with an average accuracy of over 97% using a K-Neighbors classifier. Equally, the second tested model based on the combination of 20 features resulted in a prediction with an average accuracy of over 96%, using a Ridge classifier. The statistical comparison between the results of the two different experimental approaches did not show statistically significant difference, as verified through a Mann Whitney test.

Conclusion: This study proposes the use of Machine Learning methods to establish a prediction model for benign and non-benign adrenal tumors. A high predictive performance was achieved even based on relatively few clinical cases. This study provides a decision support model which could constitute a helpful and practical clinical tool, accelerating the decision process in the healthcare system resulting to early screening, which ultimately results in time and cost savings.