Effects of Implementing a Tree Model of Diagnosis into a Bayesian Diagnostic Inference System

To estimate a diagnostic probability similarly to experts using answers to interviews, we developed a system that fundamentally behaves as a Bayesian model. For predefined interviews, we defined the sensitivity and specificity related to one or more diagnoses. Additionally, we used a predefined parent-child relation between diagnoses to decrease the number of parameters to set. After calculating the disease probability, we trained the model using the difference of post-test probability between computer calculations and three experts' opinions. We evaluated the effects of setting up tree structures. When using a tree structure, the model trained faster and produced better fitting results than the model without tree structure. Training with multiple raters' training data confused the model. The scores worsened in later epochs. Herein, we present the new method's benefits and characteristics.