Infrared thermography and machine learning based patient health monitoring

The respiration rate (RR) plays an important role in the determination of the human health condition. However, the presently used conventional RR techniques are contact-based processes that cause discomfort, skin damage, epidermal stripping, etc. They often pose problems for babies having delicate skin, making them vulnerable to skin infections. Also, the present day neonatal intensive care units are dark from the inside, which limits the use of optical technologies in the same. Infrared Thermography (IRT) is a safe and non-contact alternative, which overcomes these issues. This paper presents the application of passive IRT in monitoring the human RR. The breathing signals obtained are noisy and are filtered using the Butterworth filter. The "Ensemble of regression trees" computer vision algorithm is used to automate the tracking of nostrils in real-time, during object occlusion, and random head motion. The "Logistic regression classifier" is implemented to characterize the respiration rate of the volunteers as normal, abnormal, Bradypnea (slow breathing), or Tachypnea (fast breathing). The Validation accuracy, Training accuracy, and Testing accuracy of the classifier are obtained as 97.5%, 98%, and 95%, respectively. The Sensitivity, Specificity, Precision, G-mean, and F-measure are also computed. Further, the Standard deviation of the classier is obtained as 0.02.