Hypertension Detection in Facial Image of Visible and Near-Infrared Bands Using Sparse Coding

Abstract

Hypertension is a risk factor for cardiac and cerebrovascular disorders, and routine blood pressure monitoring is important for its early detection. The previous study that attempted to detect hypertension by applying CNN to facial visible images found that the problem was that features other than physiological responses, such as facial expressions, were mixed. Thus, we applied sparse coding to the facial visible images. However, we were able to extract features related to acute blood pressure fluctuations, we were unable to obtain sufficient accuracy. One of the reasons for the low accuracy is that the visible band is a wavelength band that is easily affected by ambient light. In contrast, the near-infrared (NIR) band is highly permeable to biological tissues and reduces the influence of external light. In this study, we attempted to detect hypertension by applying sparse coding to facial NIR images, which can capture blood flow fluctuations deep inside the body, in addition to facial visible images. By using different wavelength bands, information from the surface to the depth of the living body can be obtained, which is expected to improve the accuracy of hypertension detection. Besides, the dimensionality reduction methods, principal component analysis (PCA) and t-distributed stochastic neighbor embedding (t-SNE), were used to compare with sparse coding. As a result, a hypertension detection accuracy of 81.0% was obtained when visible images and images obtained from a Si NIR camera sensitive to 760 to 900 nm were used together. This result suggested that the detection accuracy can be improved by using multiple wavelength bands together.