Abstract
The rapidly increasing prevalence of debilitating breathing disorders, such as chronic obstructive pulmonary disease (COPD), calls for a meaningful integration of artificial intelligence (AI) into respiratory healthcare. Deep learning techniques are "data hungry" whilst patient-based data is invariably expensive and time consuming to record. To this end, we introduce a novel COPD-simulator, a physical apparatus with an easy to replicate design which enables rapid and effective generation of a wide range of COPD-like data from healthy subjects, for enhanced training of deep learning frameworks. To ensure the faithfulness of our domain-aware COPD surrogates, the generated waveforms are examined through both flow waveforms and photoplethysmography (PPG) waveforms (as a proxy for intrathoracic pressure) in terms of duty cycle, sample entropy, FEV<sub>1</sub>/FVC ratios and flow-volume loops. The proposed simulator operates on healthy subjects and is able to generate FEV<sub>1</sub>/FVC obstruction ratios ranging from greater than 0.8 to less than 0.2, mirroring values that can observed in the full spectrum of real-world COPD. As a final stage of verification, a simple convolutional neural network is trained on surrogate data alone, and is used to accurately detect COPD in real-world patients. When training solely on surrogate data, and testing on real-world data, a comparison of true positive rate against false positive rate yields an area under the curve of 0.75, compared with 0.63 when training solely on real-world data.
Introduction
Deep learning models have shown immense potential in identifying patterns within physiological signals, particularly those captured by wearable sensors. However, the efficacy of these models is often limited by the availability of large, labeled datasets. In respiratory health, collecting clinical data from COPD patients is logistically challenging, expensive, and time-consuming. While synthetic data generated through computer simulations can mitigate some of these issues, it often fails to capture the complex, non-linear interactions between respiratory mechanics and systemic physiological responses, such as blood flow variations.
This paper addresses the "data hunger" problem by introducing a physical hardware simulator. This device allows healthy individuals to generate high-quality, physiologically accurate training data that mimics the airway obstruction seen in COPD patients, providing a "physically meaningful" surrogate for clinical data.
Methodology
The core of this research is a custom-built COPD-simulator. The apparatus is constructed from 3D-printed joiners, PVC tubing, and one-way valves to ensure unidirectional airflow during the respiratory cycle.
Apparatus Design
The device utilizes separate pathways for inspiration and expiration. Airway obstruction is simulated by varying the diameter of the expiratory tube from 25mm down to 3mm. This physical resistance forces the subject's body to adapt its breathing mechanics naturally, ensuring that the resulting signals are biologically authentic rather than purely synthetic.
Data Collection and AI Framework
The system records two primary modalities:
- Flow Waveforms: Standard spirometry data capturing volume and time.
- Photoplethysmography (PPG): Captured via an in-ear wearable, PPG serves as a proxy for intrathoracic pressure. The increased work of breathing against resistance causes pressure changes in the chest cavity that modulate venous return, which is reflected in the PPG signal's Respiration Induced Intensity Variation (RIIV).
A surrogate dataset was collected from 10 healthy subjects (ages 18–30) across four obstruction levels. A two-layer 1D Convolutional Neural Network (CNN) was then trained exclusively on this surrogate data to classify respiratory waveforms.
Results and Discussion
The simulator successfully replicated the full clinical spectrum of COPD, producing FEV₁/FVC ratios ranging from 0.8 (healthy) to less than 0.2 (severe obstruction).
Model Performance
When tested on a real-world dataset of 4 COPD patients and healthy controls, the model trained solely on surrogate data achieved an Area Under the Curve (AUC) of 0.75. In contrast, a model trained only on the limited real-world clinical data achieved an AUC of only 0.63. This suggests that the surrogate data provided a more balanced and comprehensive representation of the disease spectrum than the small clinical cohort.
Physiological Insights
The model correctly identified that a lower inspiratory duty cycle—spending a smaller fraction of the breath cycle inhaling—is a key indicator of obstructive disease. This aligns with clinical observations where COPD patients exhibit prolonged exhalation phases due to airway resistance.
Conclusion
Physical simulators offer a robust method for generating high-fidelity surrogate data for medical AI. By leveraging physical constraints to induce natural physiological responses, researchers can create clean, balanced datasets that outperform small clinical samples in training diagnostic models.
Future iterations of the simulator will aim to address current limitations, such as hardware air leakage and the need for dynamic (rather than static) obstruction to better mimic the concave flow-volume loops seen in clinical COPD. Additionally, the design could be adapted to simulate restrictive lung diseases by modifying inspiratory resistance.
Summary updated