Abstract
BACKGROUND: Chronic obstructive pulmonary disease (COPD), the third leading cause of death worldwide, demands prompt and precise identification and phenotyping for effective management. This study aims to develop a multimodal multi-task learning framework that concurrently performs automated detection and classification of COPD. METHODS: Retrospective multi-task model fusing computed tomography (CT) and clinical data (n = 2320) at a tertiary hospital. Predictive performance for lung-function metrics was assessed using the concordance correlation coefficient (CCC) and mean absolute error (MAE). Classification efficacy was evaluated via the area under the receiver operating characteristic curve (AUC), accuracy (ACC), precision, recall, and F1-score. Generalisability was further verified by replicating the experiments on three distinct backbone networks. RESULTS: This study included 1624 patients for model training, 348 patients for the validation set, and an additional 348 patients for the independent test set. The optimal model achieved a maximum CCC of 0.75 for forced vital capacity (FVC), corresponding to an MAE of 0.37, and a maximum CCC of 0.77 for forced expiratory volume in one second (FEV1), corresponding to an MAE of 0.33. For the binary classification task (COPD/Non-COPD), the highest AUC achieved through multi-task learning was 0.88, with a maximum ACC of 0.83. In the ternary classification task (COPD/preserved ratio impaired spirometry (PRISm)/Normal), the highest AUC reached 0.87, with a maximum ACC of 0.79. CONCLUSIONS: Multitask-learning models that integrate chest CT images with basic clinical variables outperform their single-task counterparts in both the identification and classification of COPD. This approach supports evidence-based clinical decision-making and advances the delivery of precision medicine.