Physically Meaningful Surrogate Data for COPD.

Area of Science:

• Respiratory Medicine
• Biomedical Engineering
• Artificial Intelligence in Healthcare

Background:

• Rising prevalence of chronic obstructive pulmonary disease (COPD) necessitates AI integration in respiratory care.
• Deep learning models require extensive data, but patient data is costly and time-consuming to acquire.
• Existing methods for generating COPD data are insufficient for training robust AI models.

Purpose of the Study:

• To introduce a novel, easily replicable physical simulator for generating COPD-like data from healthy subjects.
• To enable rapid and effective data generation for training deep learning frameworks in respiratory healthcare.
• To validate the faithfulness of simulated data against real-world COPD metrics.

Main Methods:

• Development of a physical COPD simulator using healthy subjects.
• Analysis of generated flow and photoplethysmography (PPG) waveforms for physiological accuracy.
• Training a convolutional neural network (CNN) on simulated data for COPD detection.

Main Results:

• The simulator generates FEV₁/FVC obstruction ratios spanning the full spectrum of COPD, from >0.8 to <0.2.
• Waveform analysis confirmed the faithfulness of the generated COPD surrogates.
• A CNN trained solely on simulated data achieved an Area Under the Curve (AUC) of 0.75 for real-world COPD detection, outperforming a model trained solely on real-world data (AUC 0.63).

Conclusions:

• The novel COPD simulator effectively generates diverse, high-fidelity surrogate data from healthy individuals.
• AI models trained on this simulated data demonstrate strong performance in detecting real-world COPD.
• This approach offers a cost-effective and efficient solution for advancing AI in respiratory healthcare.