Clinical validation of an artificial intelligence model for thyroid fine-needle aspiration biopsy indication: comparison with TI-RADS systems and human specialists in a Chilean public hospital

Abstract

Introduction Ultrasound evaluation of thyroid nodules shows significant interobserver variability, contributing to the overutilization of fine-needle aspiration biopsy (FNA). Standardized systems such as TI-RADS aim to reduce this variability; however, their performance in real-world clinical practice remains heterogeneous. Artificial intelligence (AI) has emerged as a potential tool to support clinical decision-making in this setting. Objective To evaluate the clinical performance of an artificial intelligence model for recommending FNA in thyroid nodules and to compare it with ACR TI-RADS and Horvath TI-RADS classifications applied by specialists. Materials and methods A retrospective cross-sectional study evaluating the clinical performance of an artificial intelligence model for FNA recommendation was conducted in adult patients who underwent thyroid ultrasound and FNA between January and December 2021. FNA recommendations generated by an AI model were compared with those issued by an expert radiologist and an expert endocrinologist using ACR TI-RADS and Horvath TI-RADS systems. The reference standard was cytology according to the Bethesda System, considering Bethesda ≥ III as a positive result, as an operational definition aligned with clinical decision-making for further diagnostic evaluation. Diagnostic performance metrics and concordance were calculated. Results A total of 101 patients were included (89.1% women), with a median age of 61 years and a median nodule size of 2.3 cm (IQR: 1.6–3.3). The AI model showed a sensitivity of 0.88 and a specificity of 0.41. Horvath TI-RADS, applied by an expert radiologist, demonstrated a sensitivity of 0.82 and a specificity of 0.69. Concordance between AI and TI-RADS–based methods were moderate. Conclusion In a real-world clinical setting, the AI model demonstrated performance comparable to that of human specialists for recommending FNA in thyroid nodules. These findings support its potential role as a complementary decision-support tool, particularly in settings with variability in ultrasound interpretation.

Ähnliche Arbeiten