Data-Driven Thyroid Nodule Diagnosis Using Belief Rule Base.

Area of Science:

• Medical Decision Making
• Artificial Intelligence in Medicine
• Radiology

Background:

• Physician diagnostic preferences vary, influencing medical decision-making assumptions.
• Thyroid nodule diagnosis involves handling missing clinical features, with different assumptions impacting outcomes.

Purpose of the Study:

• To compare the effectiveness of three distinct assumptions (deletion, distribution imputation, benign by default) in thyroid nodule diagnosis.
• To develop and validate a diagnostic model using belief rule base (BRB) under these different assumptions.

Main Methods:

• Collected 3766 thyroid nodule diagnosis reports from ten radiologists over seven years.
• Constructed diagnostic models using belief rule base (BRB) for each assumption: deletion, distribution imputation, and benign by default.
• Compared the performance of the three models based on diagnostic accuracy and interpretability.

Main Results:

• The benign by default assumption yielded optimal diagnostic results compared to deletion and distribution imputation.
• Belief rule base modeling demonstrated good interpretability, facilitating understanding for clinicians and patients.
• Individual radiologist preferences showed varied tendencies towards different assumptions.

Conclusions:

• The benign by default assumption is recommended for improving thyroid nodule diagnostic accuracy and efficiency.
• The study provides practical guidance for physicians to enhance diagnostic performance.
• Interpretable AI models like BRB can support clinical decision-making in radiology.