Early identification of PRISm from chest CT: a multimodal for three-class stratification of normal, PRISm, and COPD.
Authors
Affiliations (8)
Affiliations (8)
- Department of Pulmonary and Critical Care Medicine, Huadong Hospital, Fudan University, Shanghai, China.
- School of Computer and Information Engineering, Shanghai Polytechnic University, Shanghai, China.
- Department of Pulmonary and Critical Care Medicine, Kaiyuan People's Hospital, Kaiyuan, Yunnan, China.
- Department of Pulmonary and Critical Care Medicine, Jinggu County People's Hospital, Pu'er, Yunnan, China.
- Department of Radiology, Jinggu County People's Hospital, Pu'er, Yunnan, China.
- Department of Radiology, Huadong Hospital, Fudan University, Shanghai, China.
- Huadong Hospital, Fudan University, Shanghai, China.
- School of Intelligent Medical and Health Engineering, Shanghai Polytechnic University, Shanghai, China.
Abstract
Preserved ratio impaired spirometry (PRISm), a precursor to chronic obstructive pulmonary disease (COPD), is difficult to diagnose, primarily due to its subtle and inconspicuous changes in lung structure. This study aims to develop an attention-driven multi-instance learning framework with ResNet34 (MIL-R34) that integrates CT imaging with clinical data to automatically classify normal, PRISm, and COPD, thereby addressing current limitations in early PRISm detection. A total of 1,063 participants from two centers who underwent thin-slice chest CT (<1 mm) and pulmonary function tests within 2 weeks were retrospectively analyzed. After screening, 966 participants were included (Center A: 776; Center B: 190 for external testing). Participants were divided into normal, PRISm, and COPD groups. For each subject, 20 CT slices were selected and processed using a denoising strategy. Four deep learning models (ResNet34, CNN, Swin Transformer, EfficientNet) were compared. Clinical data were encoded and combined with imaging features to construct a multimodal model. Performance was evaluated using AUC, accuracy, sensitivity, and specificity. Among CT-only models, ResNet34 achieved the best performance (AUC 0.873, accuracy 0.699). The clinical-only model showed limited performance (AUC 0.656, accuracy 0.423), whereas the multimodal model further improved performance (AUC 0.879, accuracy 0.724), outperforming both unimodal approaches. External validation yielded an AUC of 0.808 and an accuracy of 0.642, demonstrating reasonable generalizability despite performance decline across centers. This study demonstrates that chest CT imaging, when combined with clinical demographic data, can reliably identify PRISm as a transitional state between normal lung function and COPD. Furthermore, the proposed multimodal provides interpretable imaging biomarkers that enhance COPD risk stratification and address current diagnostic limitations in the early detection of PRISm.