A Machine Learning-based Approach for Predicting Pulmonary Function Testing Parameters Using Spectral Oscillometry

Abstract Rationale Plethysmography as part of a full pulmonary function test (PFT) is the reference standard in determining the presence and severity of hyperinflation and gas trapping. The performance of full PFTs is very challenging for patients with COPD, requires a significant time and staff resources, and remains a limited resource in most practice settings. On the other hand, oscillometry is increasingly used in daily clinical practice as a non-invasive method to characterize the mechanical properties of the respiratory system by measuring respiratory impedance during quiet tidal breathing. The present study sought to determine whether spectral oscillometry measurements could be used to determine hyperinflation and gas trapping in lieu of a full PFT. Methods Data from a cross-sectional observational study (ClinicalTrials.govNCT05913323) which included adult participants with spirometry-confirmed “mild” (GOLD 1) or “very severe” (GOLD 4) COPD, was used in the present study. Standard spectral oscillometry parameters, namely, resistance at frequencies of 5Hz and 19Hz (R5 and R19, respectively), reactance at 5Hz (X5), as well as the area under the reactance curve (AX), were measured using oscillometry. Each participant performed a full PFT including total lung capacity (TLC, a marker of hyperinflation) and the ratio of residual volume (RV) to TLC (RV/TLC, a marker of gas trapping). A Convolution Neural Network (CNN) was developed to predict TLC (hyperinflation) and RV/TLC (gas trapping) from the spectral oscillometry parameters. The CNN featured an input layer with 10 nodes, two hidden layers with 64 and 32 nodes, and an output layer with 2 nodes. Results In 17 participants with COPD, 52.94% were female, with mean age of 66.4±8.3 years and mean FEV1 of 1.5L (57.84% predicted). The CNN model was trained and validated using an 80:20 ratio and run for 1000 epochs. Model hyperparameters, i.e., the number of neurons and learning rate, were optimized through cross-validation, resulting in an overall prediction accuracy of 60%. Incorporating demographic information (age, weight, height, sex, smoking history, and disease severity) alongside oscillometry parameters proved to be essential to order to achieve high, repeatable accuracy. The optimized and trained model successfully predicted TLC and RV/TLC with accuracies of 92% and 88%, respectively (see Figure 1). Conclusions These results confirm that combining oscillometry parameters with machine learning techniques are comparable to full pulmonary function tests, which are not always readily accessible, in the determination of hyperinflation and gas trapping in patients with COPD.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.