Non-contact monitoring of human respiration using infrared thermography and machine learning

Abstract

The respiration rate (RR) is an important indicator of human health condition. The well-established RR monitoring techniques are mostly contact processes, which have many limitations, especially for premature neonates with fragile skin. The present study reports application of Infrared Thermography for continuous and reliable monitoring of the RR, through a non-invasive and non-contact approach. Here, the temperature modulation across the nostrils during the respiration process is monitored using a thermal camera. The signal to noise ratio (SNR) of the obtained breathing waveform, is improved by subjecting it to suitable filtering techniques. The performance of the filters is compared using various parameters such as SNR, Mean square error, magnitude response, and group delay. Further, a computer vision algorithm “Ensemble of regression trees” is implemented to track the nostrils (region of interest) in the presence of significant head movement as well as object occlusion, in an automated manner. A novel “Breath detection algorithm” (BDA) is also developed to differentiate normal and abnormal breaths by predefined thresholds and obtain the breaths per minute, without any manual intervention. The robustness of the proposed algorithm is tested by implementing it on 80 breathing waveforms under various conditions, such as constant respiration, slow and fast breathing, different focal conditions, and presence and absence of air conditioner and fan. The performance of the algorithm is determined by computing its Precision, Sensitivity, Spurious cycle rate, and Missed cycle rate values, which are 98.76%, 99.07%, 0.92%, and 1.23% respectively. The parameters obtained from the proposed BDA are further fed to a 10-fold cross-validation k-Nearest Neighbour (k-NN) classifier, which uses multi-class classification to decide whether the human volunteer has normal or abnormal respiration in general, or is suffering from Bradypnea (slow breathing) or Tachypnea (fast breathing) in particular. The robustness and performance of the k-NN classifier is determined by computing its Training accuracy, Validation accuracy, and Testing accuracy, obtained as 98.59%, 99.5% and 98%, respectively. Other performance metrics such as Sensitivity, Specificity, Precision, and F-measure values are calculated as well, for each class separately. Finally, a standard deviation value of 0.0102 is obtained from the outputs of the 10-fold cross-validation method. Further, the pattern between the data points fed to the k-NN classifier is observed using the t-Stochastic Neighbour Embedding algorithm.