Development of a machine learning-based fine-grained risk stratification system for thyroid nodules using predefined...

Area of Science:

• Artificial Intelligence in Medicine
• Medical Imaging Analysis
• Oncology Diagnostics

Background:

• Thyroid nodules are common, and accurate malignancy risk stratification is crucial for appropriate management.
• Existing systems like Thyroid Imaging Reporting and Data System (TIRADS) have limitations in specificity and sensitivity.

Purpose of the Study:

• To develop and validate a machine learning (ML)-based model for estimating thyroid nodule malignancy risk.
• To evaluate the clinical utility and performance of the ML model compared to established TIRADS guidelines.

Main Methods:

• Constructed and validated an ML model using clinicoradiological features from 5708 thyroid nodules (4597 benign, 1111 malignant).
• Eight predictive models were assessed via nested 10-fold cross-validation; the best-performing model was externally validated.
• Compared the ML model's performance against multiple TIRADS interpretations (ACR, European, Korean, AACE/ACE/AME).

Main Results:

• The ML model achieved an Area Under the Receiver Operating Characteristic (AUROC) curve of 0.914, with 83.2% sensitivity and 89.2% specificity on the validation set.
• The model demonstrated significantly higher AUROC and specificity compared to TIRADS, with similar sensitivity.
• ElasticNet model performance on the external validation set: 83.2% sensitivity, 89.2% specificity, 81.8% PPV, 90.1% NPV.

Conclusions:

• A reliable ML-based predictive model for thyroid nodule malignancy risk stratification was successfully developed.
• The developed model offers enhanced specificity, contributing to more personalized and effective management of thyroid nodules.
• An interactive version of the AI algorithm is available at http://tirads.cdss.co.kr for clinical application.