COPD and Asthma Differentiation using Quantitative CT Biomarkers by Hybrid Feature Selection and Machine Learning

Introduction: There are considerable similarities between symptoms in chronic obstructive pulmonary disease (COPD) and asthma, and misdiagnosis can lead to inappropriate treatment. Computed tomography (CT) imaging can quantify lung disease features, and previous studies show structural differences in the airways and parenchyma features between COPD and asthma. The objective of this study was discriminate COPD and asthma using CT quantitative features and machine learning. Methods: Asthma and COPD patients were recruited from Thoraxklinik at Heidelberg University Hospital (Heidelberg, Germany). CT images were analyzed using VIDA Diagnostics. A total of 89 CT imaging features were investigated. For dimension reduction, hybrid filter and wrapper-based feature selection were used. For filter-based, factor analysis based on principal component analysis was used to select features and in the wrapper phase, particle swarm optimization was coupled with support vector machine algorithm to select the top features. Result: A total 95 subjects were investigated; n=47 asthma and n=48 COPD. There was no significant difference between the asthma and COPD participants for age (p=0.25), BMI (p=0.31) or FEV₁ (p=0.43). A total of 7 imaging features were selected, and COPD and asthma were differentiated with 79% accuracy (PrecisionCOPD=87, RecallCOPD=76, F1-scoreCOPD=81, PrecisionAsthma=71, RecallAsthma=83, F1-scoreAsthma=77). Conclusion: Quantitative CT imaging can discriminate COPD and asthma patients using as few as 7 CT features with moderate accuracy. The hybrid feature selection significantly reduced the number of features and increased the machine learning performance.