A decision tree for predicting the causative pathogens of community-acquired pneumonia from thin-section computed tomography

Abstract

Purpose: To determine whether decision trees are useful for predicting organisms that cause community-acquired pneumonia (CAP).

Materials and methods: We developed a decision tree for predicting the organisms that cause CAP based on previously reported characteristic computed tomography findings. Sixteen readers (two student doctors, six residents, and eight radiologists) separately diagnosed 68 randomly selected cases of CAP using chest computed tomography. The first, second, and third most likely causative organisms were estimated for each case, and the percentages of correct answers were evaluated for consistency with the isolated organisms. The same 68 cases were then read again using the decision tree, with the first three most likely organisms again being estimated, and the percentage of agreement was evaluated as the percentage of correct responses after using the decision tree.

Results: For student doctors, residents, and radiologists, the percentage of correct responses increased significantly (p < 0.0001) when the decision tree was used to predict the first, second, and third most probable causative organism. The radiologists all obtained an accuracy rate of 80% or higher when estimating up to three candidate organisms using the decision tree. The organism for which the decision tree was most useful was Mycoplasma pneumoniae, followed by Haemophilus influenzae and Chlamydophila pneumoniae (p < 0.001).

Conclusion: Use of the decision tree made it possible to estimate the organisms responsible for CAP with a high correct response rate.

Keywords: Chest; Community-acquired pneumonia; Computed tomography; Decision tree.

Fig. 1

The decision tree for estimating the causative organisms of community-acquired pneumonia

Fig. 2

Percentage of correct answers among the first, second, and third candidates for student doctors, residents, and radiologists

Fig. 3

a and b Mycoplasma pneumoniae pneumonia in a 21-year-old man (Case 4). Transverse thin-section computed tomography (CT) at the level of the division of the middle lobe bronchus showed ground-glass opacity with a segmental distribution, bronchial wall thickening, and centrilobular nodules (arrows). These findings strongly suggest M. pneumoniae as the causative pathogen, because the pathogen is likely to be one that causes bronchopneumonia and the patient is under 60 years of age. The correct response rates for the first candidate for student doctors + residents, as well as radiologists, were 0% and 37.5%, respectively, before using the decision tree and 100% and 87.5% after using the decision tree

Fig. 4

a and b Chlamydophila pneumoniae pneumonia in a 72-year-old woman (Case 18). Transverse thin-section CT at the level of the right lower lobe showed ground-glass opacity and consolidation with a segmental distribution. Bronchial wall thickening or centrilobular nodules could not be seen. These findings suggest that the causative pathogen was likely to be a pathogen causing lobar pneumonia. Additionally, because an acinar pattern can be observed, C. pneumoniae is suggested. The correct response rates for the first candidate for the student doctors + residents, as well as radiologists, were both 0% before using the decision tree, and both 62.5% after using the decision tree

Fig. 5

a and b Streptococcus pneumoniae pneumonia in a 66-year-old woman (Case 53). Transverse thin-section CT at the level of the division of the right B¹⁰ showed consolidation and ground-glass opacity with a non-segmental distribution. Bronchial wall thickening or centrilobular nodules cannot be seen. These findings suggest that the causative pathogen was likely to be one causing lobar pneumonia. According to its statistical frequency, S. pneumoniae was suggested as the pneumonia pathogen. If the distribution was diagnosed as segmental, S. pneumoniae could also be the first candidate because of the absence of evident bronchial wall thickening and centrilobular nodules. The correct response rates for the first candidate for student doctors + residents, as well as radiologists, were 12.5% and 87.5%, respectively, before using the decision tree, and 87.5% and 100% after using the decision tree