Data-Driven Thyroid Nodule Diagnosis Using Belief Rule Base

Abstract

Doctors’ diagnosis preferences are different, which makes them adopt different assumptions in medical decision making. Taking the diagnosis of thyroid nodules as an example, this study compares three assumptions, namely deletion, imputation based on the distribution (distribution), and benign by default (benign). For deletion, which is the most used assumption, the clinical reports with missing features would be deleted. For distribution, the missing features would be replaced with a distribution of features with respective probabilities. Besides the two assumptions, certain doctors have also stated that they leave benign features unrecorded because they think that such benign features are irrelevant to the final diagnosis. Under the benign assumption, the missing features would be replaced with benign features. The three assumptions are tested comparatively. Moreover, the belief rule base (BRB) is used to construct the diagnostic model under the three assumptions since it is essentially a white-box approach that can provide good interpretability and direct access to doctors and patients. A total of 3766 clinical reports on thyroid nodule diagnosis were collected from ten radiologists over a seven-year period. Case study results validate that the benign by default assumption has produced the optimal results, although different doctors could present varied tendencies towards different assumptions. Guidance and suggestions for doctors’ practical work have been made based on the study results to improve work efficiency and diagnostic accuracy.