PulmoClass-3DAtt: A Self-Attention Network for Classification of COPD, PRISm and Normal

Background: Chronic obstructive pulmonary disease (COPD) is the third leading cause of death worldwide, making its accurate diagnosis critical. It is difficult to distinguish COPD from preserved ratio impaired spirometry (PRISm), due to their shared small-airway pathology. This study develops a novel deep-learning framework that combines chest CT images with clinical variables to discriminate between COPD, PRISm, and normal categories. Methods: In this retrospective study, consecutive subjects were enrolled from a university-affiliated tertiary hospital between January 2018 and June 2024. After random split at an 8:2 ratio, the training cohort was used to develop a convolutional encoder that extracts imaging features and simultaneously predicts five spirometric parameters (FEV 1 , FVC, FEV 1 /FVC, FEV 1 %pred, FVC %pred). The extracted features were then fed into the self-attention deep-learning model PulmoClass-3DAtt, which was trained and fine-tuned through ten-fold cross-validation to classify disease status. Classification performance was quantified by the area under the receiver-operating-characteristic curve (AUC), overall accuracy (ACC), precision (Pre), sensitivity (Se) and specificity (Sp). Regression performance for spirometric indices was assessed with mean-squared error (MSE), mean absolute error (MAE) and concordance correlation coefficient (CCC). Results: The cohort comprised 1918 participants (1362 COPD, 174 PRISm, 382 normal controls). The model achieved robust overall performance (AUC = 0.86, ACC = 0.87). For COPD, Normal and PRISm the respective values are ACC 0.87, 0.88, 0.87; Se 0.90, 0.88, 0.42; Sp 0.84, 0.88, 0.97; Pre 0.85, 0.79, 0.72; and F1-score 0.88, 0.85, 0.53. In the validation set, the five spirometric metrics were predicted with MSE 0.03– 0.35, MAE 0.14– 0.46 and CCC 0.56– 0.83. Conclusion: By integrating multimodal imaging–function data through a self-attention framework, PulmoClass-3DAtt reliably discriminates among COPD, PRISm and normal status, providing an immediately applicable tool for clinical decision support and the delivery of precision pulmonary medicine. Plain Language Summary: This study furnishes the first feasibility evidence that a self-attention deep-learning architecture can discriminate among chronic respiratory diseases, establishing a reusable paradigm for future investigations while directly informing clinical decision-making and disease stewardship.By supplying an image-based, equipment-agnostic classifier, the model offers a globally deployable, low-cost solution to the widespread under-diagnosis and misclassification of lung-function impairment.Simultaneous prediction of spirometric indices from routine CT and readily available clinical data enables real-time grading of disease severity and individualised planning of follow-up therapy without additional patient burden.