Differentiating Emphysema From Emphysema-Dominated COPD Patients with CT Imaging Feature and Machine Learning

Abstract

Background: Differentiating between emphysema and emphysema-dominant chronic obstructive pulmonary disease (COPD) remains challenging but crucial for appropriate management. Quantitative computed tomography (QCT) offers potential for improved characterization, yet its optimal application in conjunction with machine learning for this differentiation is not fully established.

Methods: This prospective study enrolled 476 participants (99 with emphysema, 377 with emphysema-dominant COPD) aged 34-88 years. All participants underwent spirometry and chest CT scans. QCT features including emphysema index, mean lung density, airway measurements, and vessel measurements were extracted. A random forest model was developed using these QCT features to differentiate between the two groups. The model's performance was assessed using area under the receiver operating characteristic curve (AUC-ROC). Correlations between QCT parameters and pulmonary function tests were analyzed.

Results: The model achieved an AUC-ROC of 0.97 (95% CI: 0.96-0.99) in differentiating emphysema from emphysema-dominant COPD. Emphysema index and airway wall thickness were the most important features for classification. QCT-derived emphysema index showed strong negative correlation with FEV1/FVC (ρ = -0.54, p<0.001) in the emphysema-dominant COPD group, but no significant correlation in the emphysema group (ρ = 0.001, p=0.993). Mean lung density was significantly lower in the emphysema-dominant COPD group compared to the isolated emphysema group (p<0.001).

Conclusion: Machine learning analysis of QCT features can accurately differentiate emphysema from emphysema-dominant COPD. The differing relationships between QCT parameters and lung function in these two groups suggest distinct pathophysiological processes. These findings may contribute to improved diagnosis, phenotyping, and management strategies in emphysema and COPD.

Keywords: chronic obstructive pulmonary disease; computed tomography; emphysema; emphysema-dominant COPD; quantitative computed tomography.

Figure 1

Correlation coefficients between QCT measurements and lung function measurements in (A) the emphysema group and (B) the emphysema-dominant COPD group. The figure highlights the relationships between quantitative imaging features (eg, low attenuation volume, mean lung density) and physiological measures (eg, FEV1, FEV1/FVC ratio). Strong positive or negative correlations suggest potential imaging biomarkers for disease assessment.

Figure 2

Confusion matrix of the random forest model for classifying emphysema and emphysema-dominant COPD. The matrix displays the number of correctly and incorrectly classified cases, with accuracy, sensitivity, and specificity highlighted.

Figure 3

ROC curve of the random forest model in classifying emphysema and emphysema-dominant COPD, with an area under the curve (AUC) of 0.97 (95% CI: [0.96, 0.99]). The ROC curve demonstrates the model’s excellent diagnostic accuracy, with a high true positive rate and low false positive rate across different thresholds.

Figure 4

SHAP bar plot showing the average feature impact on the random forest model’s classification of emphysema and emphysema-dominant COPD. Features such as WA% and airway wall thickness are shown to have the highest contribution to the classification decision, providing insights into the most influential imaging parameters.

Figure 5

SHAP value plot illustrating individual feature contributions for specific cases in classifying emphysema and emphysema-dominant COPD. The plot highlights how key features, such as WA%, airway wall thickness, and Pi10, influence the model’s predictions for individual patients, offering a more patient-specific interpretability.

Figure 6

Decision curve analysis of the random forest model evaluating the clinical net benefit of classifying emphysema and emphysema-dominant COPD. The curve compares the net benefit of the model with default strategies (treat-all vs treat-none) across a range of threshold probabilities, demonstrating its potential utility in clinical decision-making.