Endocrine Abstracts

Objectives: Thyroid nodules are common and mostly benign, but around 7-15% of them account for thyroid cancers. In recent years, a significant increase in the detection rate of thyroid nodules has been observed. Ultrasonography (US) has become a valuable tool in thyroid nodules malignancy risk assessment, however, it might lead to thyroid cancer overestimation and unnecessary biopsies. The aim of our study was to create a machine-learning prognostic model for thyroid nodules malignancy risk assessment, based on sonographic characteristics, fine-needle aspiration biopsy (FNAB) and blood tests results, verified by histopathology reports. The intention was to compare the accuracy of thyroid malignancy detection with the best prognostic model and EU-TIRADS reporting system.

Methods: This was a prospective study of machine learning thyroid malignancy risk prediction. We analyzed data of patients who underwent thyroidectomy due to the results of preoperative US and FNAB cytology reports. 422 patients with nodules (193 malignant) were included. They were split into training and test sets (3:1). The models were developed on the training set and evaluated on the test set. A variety of algorithms were explored.

Results: Due to its explainability and good performance, a random forest was selected as a base model. Test set ROC AUC at 71%. In our study, the model was better correlated with thyroid nodule malignancy than the EU-TIRADS reporting system. Statistically significant thyroid nodule risk factors were microcalcifications (P<0.001) and age (lower age correlated with greater malignancy risk, P<0.001). The larger the anteroposterior dimension, the greater the risk of malignancy (P<0.001). Otherwise, the larger the longitudinal dimension (P<0.001) or transverse dimension (P<0.001), the lower the risk of thyroid nodule malignancy. Using statistically important features the new index of anteroposterior-to-longitudinal dimension was created (ROC AUC 66% (95%CI 62.0 – 71.0)). The created index is easy to use and therefore may serve as a help for clinicians.

Conclusions: In the era of thyroid nodule overdetection there is a need for accurate approaches in thyroid malignancy prediction. The model created by our group is characterised by good accuracy and might be more effective than systems based only on sonographic characteristics. However, future development of these data-driven methods is limited by sample size.