JMIR Publications

ABSTRACT

Background:

Deep learning (DL) has attracted widespread attention in clinical practice. Its application to diagnosing chronic obstructive pulmonary disease (COPD) and grading COPD severity has been explored, but systematic evidence assessing diagnostic and grading accuracy remains limited, posing challenges for developing intelligent diagnostic tools.

Objective:

This study systematically estimated the diagnostic and grading accuracy of DL for COPD to inform the development or optimization of intelligent detection tools.

Methods:

The Cochrane Library, Embase, Web of Science, and PubMed were systematically searched for relevant studies published up to April 1, 2025. The risk of bias was assessed via the Quality Assessment of Diagnostic Accuracy Studies-2 tool. Subgroup analyses stratified by the method of validation set generation and the source of imaging data were conducted. Meta-analyses were carried out on the validation sets.

Results:

In total, 50 original studies involving 870,074 cases were included. The results demonstrated that, for the diagnostic binary classification of COPD, DL models yielded a pooled sensitivity of 0.87 (95% confidence interval [CI]: 0.84–0.90), specificity of 0.88 (95% CI: 0.83–0.92), negative likelihood ratio (NLR) of 0.15 (95% CI: 0.12–0.19), positive likelihood ratio (PLR) of 7.2 (95% CI: 5.0–10.4), diagnostic odds ratio (DOR) of 49 (95% CI: 29–86), and the area under the summary receiver operating characteristic (SROC) curve (AUC) of 0.93 (95% CI: 0.18–1.00). For computed tomography (CT)-based DL models, pooled sensitivity was 0.87 (95% CI: 0.84–0.89), specificity was 0.87 (95% CI: 0.82–0.91), NLR was 0.15 (95% CI: 0.12–0.19), PLR was 6.7 (95% CI: 4.8–9.4), DOR was 44 (95% CI: 26–73), and AUC was 0.92 (95% CI: 1.00-0.00). For respiratory sound–based models, sensitivity was 0.90 (95% CI: 0.81–0.95), specificity was 0.96 (95% CI: 0.90–0.98), NLR was 0.11 (95% CI: 0.06–0.20), PLR was 20.9 (95% CI: 8.5–51.4), DOR was 196 (95% CI: 63–612), and AUC was 0.97 (95% CI: 1.00-0.00). In the multiclass classification of COPD, DL models showed limited accuracy in discriminating among different GOLD stages: GOLD stage 0 (84.2%, 95% CI: 68.1%–95.4%), stage 1 (61.7%, 95% CI: 46.6%–75.9%), stage 2 (67.9%, 95% CI: 46.2%–86.2%), stage 3 (70.8%, 95% CI: 45.0%–91.1%), and stage 4 (66.2%, 95% CI: 30.1%–94.7%).

Conclusions:

DL models based on CT imaging or respiratory sounds demonstrated excellent diagnostic accuracy for COPD. These models hold promise for the development of intelligent diagnostic tools in clinical practice. Nonetheless, their diagnostic performance in the multiclass classification of GOLD stage was suboptimal. Furthermore, current studies have not adequately addressed the impact of varying imaging protocols and lack external validation. Future studies should refine study designs to verify the generalizability of these findings.