Endocrine Abstracts

Introduction: The present study investigates the efficacy of an autonomous diagnostic system that employs artificial intelligence (AI) for the prediction of malignancy in thyroid nodules. The study focuses on evaluating the performance of this AI-based system in detecting malignant thyroid nodules, with the aim of improving diagnostic accuracy and patient outcomes. AI, with its ability to analyze large amounts of complex data and identify patterns that may be difficult for humans to detect, can potentially improve the accuracy and speed of thyroid carcinoma diagnosis.

Materials and methods: A retrospective single-center study was carried out, where 1104 patients who required surgical management of thyroid nodules were included. We collected a new clinical dataset of 1104 clinical health records with their respective nodule histopathology; after that, we trained ten machine learning models on this dataset and estimated their prediction performance by cross-validation. The data is optimized, and each artificial intelligence algorithm is evaluated independently and through cross-validation to obtain the lowest possible error, the highest accuracy, and the highest precision with the shortest response time. Also, we conducted a variable importance analysis to examine the relevance of nodule characteristics on the model performance.

Results: In this study, 1104 subjects were enrolled, and 925 (83.78%) were women. The mean age was 47.7 years (24-82 y); the histopathology analysis diagnosed 753 (68.21%) nodules with malignancy. The accuracy of the recurrent neural network model after hyperparameter turning was as high as 99.06%, and the F1-Score was 0.971, whereas its sensitivity and specificity varied significantly in detecting malignancy with different algorithms. The sensitivity-specificity to detect any malignant nodule for the three most accurate algorithms was: Recurrent neural network (RNN) 95.06%-99.19%, Support Vector Machine (SVM) 88.17%-94.33%, and Random Forest (RF) 78.86%-93.28%. Also, our analysis of variable importance helps us identify three key variables in diagnosing malignant thyroid nodules. The variables that have been recognized are the existence of calcification, the cyst’s composition, and the nodule’s size. Each of these factors is a powerful indicator of malignancy in nodules and aligns with the findings of clinical research and prior modeling studies.

Conclusion: After conducting cross-validation, we’ve successfully created an AI model that accurately detects the potential risk of thyroid carcinoma. Nonetheless, since the models were trained based on a small dataset, more patient data would be required, and external validation through prospective studies is necessary.