Optimizing thyroid AUS nodules malignancy prediction: a comprehensive study of logistic regression and machine...

Area of Science:

• Endocrinology
• Oncology
• Medical Informatics

Background:

• Accurate diagnosis of thyroid nodules with indeterminate cytology, especially atypia of undetermined significance (AUS), remains a clinical challenge.
• Predicting malignancy risk in AUS nodules is crucial for appropriate patient management.
• Distinguishing benign from malignant nodules avoids unnecessary invasive procedures.

Purpose of the Study:

• To compare the performance of machine learning (ML) and logistic regression (LR) models in predicting malignancy risk for thyroid nodules in the AUS category.
• To identify the most effective model for accurate risk stratification.

Main Methods:

• A retrospective analysis of 356 AUS nodules from 342 patients who underwent thyroid surgery.
• Data including clinical, ultrasonographic, and molecular features were collected and split into training (70%) and validation (30%) sets.
• Two ML models (random forest, XGBoost) and three LR models (lasso, best subset, backward stepwise) were developed and evaluated using AUC, calibration, and clinical utility.

Main Results:

• Of the 356 AUS nodules, 90% were malignant, primarily papillary thyroid carcinoma with frequent BRAF V600E mutations.
• The LR model using backward stepwise regression demonstrated superior performance with AUCs of 0.83 (training) and 0.80 (validation).
• This LR model showed good calibration and clinical utility, outperforming the ML models.

Conclusions:

• Logistic regression, specifically the backward stepwise method, is superior to ML models for predicting malignancy in AUS thyroid nodules.
• A nomogram derived from the best LR model provides a practical tool for personalized risk assessment.
• This approach can aid clinical decision-making, potentially reducing overtreatment and improving patient outcomes.