Prediction of respiratory rate from schlieren images using artificial intelligence

Respiratory rate is a fundamental physiological parameter and one of the earliest indicators that plays a crucial role in assessing a patient's condition. However, existing respiratory rate measurement methods face challenges related to stability and invasiveness. In this study, we developed a noninvasive and stable method for measuring respiratory rate using the schlieren imaging technique combined with artificial intelligence (AI).

A system capable of accurately assessing respiratory rate was established using the schlieren imaging technique to visualize respiration-induced airflow, combined with AI. Schlieren-based respiratory images were acquired from 40 healthy subjects under three conditions: normal respiratory rate (2 min), fast respiratory rate (1 min), and slow respiratory rate (1 min). Using these images, three AI models (ResNet, InceptionNet, and EfficientNet) were trained to classify respiration and reconstruct respiratory signals by applying sine-fitting techniques to the classification results. Afterward, respiratory rates were estimated, and their performances were compared against those obtained using the standard tool called capnography.

The accuracies of ResNet, InceptionNet, and EfficientNet were 0.82, 0.91, and 0.92, respectively. Among them, EfficientNet demonstrated the superior performance (accuracy: 0.92; precision: 0.90; sensitivity: 0.93; F1-score: 0.92). Intra-class correlation values between the respiratory rates predicted by the three models (ResNet, InceptionNet, and EfficientNet) and the actual respiratory rate were 0.99, 0.99, and 0.98, respectively.

These results suggest that our AI-based respiratory monitoring system using the schlieren imaging technique provides a stable and non-invasive method for accurate respiratory assessment.