Endocrine Abstracts

Background: In 2018 we presented the first version (v1) of our microRNA-based algorithm for indeterminate thyroid nodules classification (mir-THYpe full), which was used in real-world clinical routine until recently, when an new optimized version (v2) was released using new machine learning techniques and combinating microRNA and DNA data.

Objective: Our aim was to simulate and analyze what the performance of v2 algorithm would have been, if it had been used in the classification of the same samples originally classified by the v1.

Methods: The cohort of this study was composed of microRNA and DNA data extracted from thyroid FNA smear slide of 1.718 nodules (from 1.687 patients who payed for the test) (945 Bethesda 3 and 773 Bethesda 4) that were originally classified by the v1 in real-world clinical routine and now were re-analyzed and classified by the new optimized v2 algorithm. The molecular analysis performed in the samples consisted of microRNA profile and DNA mutation analysis (BRAF V600E and pTERT C228T/C250T). From those, anatomopathological data was available for 329 nodules (112 benign and 217 malignant) and used to evaluate the performance of the v2. Due to the unrealistically high disease prevalence (66.0%), a real-world adjusted prevalence (32%) was performed based on Bayes´ theorem.

Results: When comparing the results of the v1 with the v2 version of the algorithm, 979 vs 1.175 samples were classified as negative for malignancy (Benign Call Rate/BCR - 57.0% vs 68.4%) and 739 vs 543 samples as positive. In this simulation, the real-world performance of the new v2 would be: 94.5% of sensibility, 75.9% of specificity, 64.8% of positive predictive value (PPV), 96.7% of negative predictive value (NPV) and 81.8% of accuracy.

Conclusion: The results show an BCR improvement in performance of v2 compared to v1, resulting in an increment of the number of cases that would have benefited from the test avoiding unnecessary diagnostic surgeries. The v2 also showed high PPV and NPV, suggesting that real-world performance for future tests will also be optimized.