RespiroDynamics: A Multifaceted Dataset for Enhanced Lung Health Assessment Using Deep Learning

Abstract

Advancements in lung health assessment, a critical component in diagnosing respiratory conditions, have gained prominence in medical research. This is especially true with the advent of non-invasive techniques such as spirometry. Central to this diagnostic method are three key metrics: Forced Vital Capacity (FVC), Forced Expiratory Volume in 1 second (FEV1), and Peak Expiratory Flow (PEF). In light of the increasing need for accurate and reliable assessment tools, the development of comprehensive datasets is imperative for advancing research in this field. Our paper presents RespiroDynamics: A Comprehensive Multimodal Respiratory Dataset, compiled from 60 participants, recorded in two sessions labelled ’rest’ and ’exercise’. This dataset incorporates a variety of data types, including Red-Green-Blue (RGB) and Thermal videos, Heart Rate (HR), ECG readings and metadata, all synchronized with observed respiratory activities. Additionally, these data are enriched with reference values from the NHANES III (Hankinson-1999) distribution. We elaborate on the data collection methodology, post-processing techniques employed, and the analytical approach to extract meaningful patterns and insights. Our evaluation, using a 5-fold cross-validation method on binary classification models for FEV1/FVC, revealed remarkable accuracies: 99.7% for the RGB model and a perfect 100% for the thermal model. For the PEF 3-class model, accuracies were 97.14% for both RGB and 96.0% for thermal models. This study aims to underscore the dataset’s potential in enhancing and establishing the first robust deep-learning model for lung health diagnostics.