Rationale: Dyspnea is common in smokers with or without chronic obstructive pulmonary disease. Its multifactorial nature makes it challenging to identify specific factors causing dyspnea in smokers with and without chronic obstructive pulmonary disease.
Objectives: The study aims to identify associations between clinical history, spirometry, and computed tomography findings related to dyspnea in smokers, and to develop and compare dyspnea models using different variable combinations.
Methods: Dyspnea was defined as a self-reported modified Medical Research Council dyspnea scale score ≥ 2. Participants from the COPDGene Study dataset were utilized and split into training and testing samples (80%/20%) to develop and validate a predictive model. The ECLIPSE Study was used for external validation. Bivariable and multivariable logistic regression analyses were used to examine factors associated with dyspnea. Predictive models were developed using Elastic Net.
Main Results: The final prediction model demonstrated good predictive performance, achieving an area under the curve of 0.85 in the test set and 0.80 in the external dataset. We confirmed prior associations with dyspnea and identified novel interactions of multiple risk factors with chronic obstructive pulmonary disease severity.
Conclusions: Dyspnea in smokers with and without chronic obstructive pulmonary disease can be predicted with high accuracy using a model that utilizes clinical history, spirometry, and chest CT imaging. To make accurate predictions, data from at least two of the three variable domains (clinical history, spirometry, or chest CT imaging) was required.
Supplementary Information: The online version contains supplementary material available at 10.1007/s00408-026-00871-5.
Keywords: Machine learning; Quantitative chest computed tomography; Spirometry.