Heart Rate Assessment in a Pediatric ICU with Non-Contact Infrared Thermography and Machine Learning

Heart rate is one of the vital signs for monitoring health. Non-invasive, non-contact assessment of heart rate can lead to safe and potentially telemedicine based monitoring. Thermal videos as a modality for capturing heart rate has been underexplored. Regions with large vessels such as the face can capture the pulsatile change in temperature associated with the blood flow. The use of a machine learning-based approach to capture heart rate from continuous thermal videos is currently lacking. Our present clinical investigation comprises the continuous monitoring of heart rate from a smaller number of samples by using a combination of an efficient deep-learning-based segmentation followed by domain-knowledge-based feature calculation for estimating heart rate from 124 thermal imaging videos comprising 3,628,087 frames of 65 patients, admitted to the pediatric intensive care unit at AIIMS, New Delhi. We hypothesized that periodic fluctuations of thermal intensity over the face can capture heart rate. Frequency domain features for thermal time series were extracted followed by supervised learning using a battery of models. A random forest model yielded the best results with a root mean squared error of 24.54 and mean absolute percentage error of 16.129. Clinical profiling of the model showed a wide range of clinical conditions in the admitted children with acceptable model performance. Affordable and commercially available thermal cameras establish the feasibility and cost viability of exploring deployments for patient heart rate estimation in non-invasive and non-contact environments.