Endocrine Abstracts

Introduction: Thyroid nodules (TN) are common entities usually discovered during a physical exam or by chance with ultrasound (US) procedures performed for reasons different from thyroid check. The majority of nodules are benign, notwithstanding the growing incidence of thyroid cancer (TC), which increases much largely as compared to all other tumours each year and will be, presumably, the most incident malignancy in women after breast cancer. The clinical challenge relies on the accurate identification of malignant nodules needing attention since the very beginning of the diagnostic process from nodules that will follow an indolent course of disease. Therefore, pre-surgical assessment of TN should improve in order to advance into the era of personalized medicine and identify the best way to manage what has been defined as a ‘tsunami of thyroid nodules’.

Aims: The present study has focused on the development of an algorithm capable of inferring TN state of malignancy in order to improve patient’ management and avoid inappropriate use of diagnostic procedures.

Materials and methods: Information regarding epidemiological, clinical, biochemical, ultrasound, molecular and cytological data of over 12.000 TN with known histological diagnosis was collected and analyzed with Machine Learning techniques. Specifically, TN were grouped into clusters with similar characteristics using partition clustering algorithms such as K-means e K-medoids, being the latter more flexible to outliers deviations.

Results: The designed algorithm was able to accurately identify malignant and benign TN showing malignant and benign US characteristics, respectively, confirming US high predictability for these categories. Moreover, the majority of TN belonging to malignant or suspected malignant cytological classes showed malignant histology.

Discussion: Our algorithm did not achieve the capability of inferring TN malignancy belonging to the grey zone of indeterminate nodules, where US and/or cytology are not accurate enough. However, it was able to confirm that US is essential since it provides the majority of information and is highly predictive for benign TN and for TN having a cytological diagnosis associated to US characteristics highly suspicious of malignancy. In conclusion, the mathematical model built herein brings insight for future tools that could be used to improve TN management and estimate malignancy probability that underlies the discovery of a TN.