Improving clinical disease subtyping and future events prediction through a chest CT-based deep learning approach

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonary Medicine

Background:

• Chronic obstructive pulmonary disease (COPD) is a major global health burden.
• Accurate assessment of COPD severity and prognosis is crucial for patient management.
• Current methods often rely on a combination of clinical, functional, and imaging data.

Purpose of the Study:

• To develop and evaluate a deep learning (DL) approach for extracting comprehensive information from high-resolution computed tomography (HRCT) in COPD patients.
• To create a DL model that learns a compact subject representation predictive of COPD severity and outcomes.
• To assess the model's ability to predict spirometric obstruction, emphysema severity, exacerbation risk, and mortality.

Main Methods:

• A DL-based model was developed to extract regional image features from HRCT scans.
• The model adaptively weighted these features to form an aggregate patient representation.
• The model was evaluated on 10,300 participants from the COPDGene cohort.

Main Results:

• The DL model strongly predicted spirometric obstruction (R² = 0.67) and GOLD stage (89.1% within one stage).
• It achieved 41.7% and 52.8% accuracy in stratifying centrilobular and paraseptal emphysema severity, respectively.
• The model predicted future exacerbations (AUC = 0.73) and outperformed the BODE index for all-cause mortality prediction (concordance 0.61 vs 0.56).

Conclusions:

• The DL model independently predicted key COPD outcomes from CT imaging alone.
• This AI-driven approach demonstrates potential for enhancing COPD research and clinical practice.
• The method offers a non-invasive way to assess disease severity and prognosis.