Efficacy of machine learning model and EU-TIRADS-PL reporting system in predicting thyroid malignancy risk: a comparison.
Authors
Affiliations (6)
Affiliations (6)
- Chair and Department of Endocrinology, Jagiellonian University Medical College, Cracow, Poland.
- Institute of Microbiology, ETH Zürich, Switzerland.
- Chrzanów County Hospital, Chrzanów, Poland.
- Department of Pathomorphology, Jagiellonian University Medical College, Cracow, Poland.
- Department of Endocrine Surgery, Third Chair of General Surgery, Jagiellonian University Medical College, Cracow, Poland.
- Chair and Department of Endocrinology, Jagiellonian University Medical College, Cracow, Poland. Electronic address: [email protected].
Abstract
Thyroid nodules are common and mostly benign, but 5-15% are malignant. Increasing detection rates may lead to overdiagnosis and unnecessary fine-needle aspiration biopsies (FNAB). We aimed to develop a machine learning (ML) model for thyroid malignancy risk assessment using ultrasound, FNAB, laboratory data and to compare its performance with the EU-TIRADS-PL reporting system. This prospective study included 561 patients who underwent thyroidectomy based on preoperative ultrasound and FNAB results, including 261 malignant nodules (46.5%). Data were divided into training and test sets (3:1). Multiple ML algorithms were evaluated using clinical, sonographic, cytological and biochemical variables verified by histopathology. Logistic regression (LR) was selected as the final model because of its good performance and interpretability. The model achieved a test set ROC AUC of 74% and showed better correlation with thyroid malignancy than the EU-TIRADS-PL system. Significant predictors of malignancy included microcalcifications (p=0.002), younger age (p<0.001), male sex (p=0.001),and larger anteroposterior dimension (p=0.026). Larger longitudinal dimension was associated with lower malignancy risk (p<0.001). An anteroposterior-to-longitudinal dimension index was developed, with ROC AUC of 68% (95% CI 62.0-71.0). The proposed ML model showed good accuracy and may improve thyroid malignancy risk assessment compared with systems based only on sonographic features.