Non-contact breathing rate monitoring using infrared thermography and machine learning

Monitoring vital physiological parameters such as breathing rate (BR) is crucial for assessing patient health. However, current contact-based measurement methods often cause discomfort, particularly in infants or burn patients. This study aims to develop a non-contact system for monitoring BR using infrared thermography (IRT). This approach permits to detects and tracks the nose from thermal video, extracts temperature variations into a breathing signal, and processes this signal to estimate BR. The estimated BR is then classified into three health categories (bradypnea/normal/tachypnea) using k-nearest neighbors (k-NN). To evaluate system accuracy and robustness, experiments were conducted under three conditions: (i) stationary breathing, (ii) breathing with head movements, and (iii) specific breathing patterns. Results demonstrated high consistency with contact-based photoplethysmography (PPG) measurements, achieving complement of the absolute normalized difference (CAND) index values of 94.57%, 93.71%, and 96.06% across the three conditions and mean absolute BR errors of 1.045 bpm, 1.259 bpm, and 0.607 bpm. The k-NN classifier demonstrated high performance with training, validation, and testing accuracies of 100%, 100%, and 99.2%, respectively. Sensitivity, specificity, precision, and F-measure results confirm system reliability for non-contact BR monitoring in clinical and practical settings.