Re-Defining High Risk COPD with Parameter Response Mapping Based on Machine Learning Models

Area of Science:

• Pulmonary Medicine
• Radiology
• Artificial Intelligence

Background:

• Chronic obstructive pulmonary disease (COPD) diagnosis relies on FEV1% predicted value (FEV1%pre), but optimal thresholds for high-risk identification require refinement.
• Parameter response mapping (PRM) derived from CT scans offers quantitative insights into small airway remodeling, a key feature of COPD.

Purpose of the Study:

• To determine the optimal FEV1%pre threshold for identifying high-risk COPD using a machine learning classification model based on PRM.
• To establish novel, PFT-parameter-based high-risk criteria for COPD that align with machine learning classifications.

Main Methods:

• Retrospective analysis of 561 non-COPD subjects with complete PFT, CT scans, and questionnaires.
• Utilized 72 PRM quantitative parameters from CT scans to assess small airway remodeling.
• Established random forest classification models for 80 different FEV1%pre thresholds (50%-129%) to find the optimal cutoff for distinguishing high-risk COPD.

Main Results:

• Machine learning-based PRM demonstrated superior consistency between PRM parameters and PFT in distinguishing high-risk COPD.
• An area under the curve (AUC) of 0.84 was achieved with an FEV1%pre threshold of 72%, indicating optimal performance.
• Lower AUCs of 0.72 and 0.64 were observed at thresholds of 80% and 95%, respectively.

Conclusions:

• Machine learning-based PRM is a feasible method for redefining high-risk COPD.
• Establishing an optimal FEV1%pre threshold is crucial for improving the accuracy and consistency of high-risk COPD diagnosis.