Comparing Posture Classification: A Human Lying Posture Pressure-Map Dataset

We discuss methods and algorithms for classifying the posture of a patient in their bed. The actual classification tasks are performed with a measurement chain including an in-house designed pressure mattress, a matrix of $30 \times 11$ sensing spots, a data concentrator, and a cloud-based service. Utilizing a survey of open-source datasets that facilitate such classifying operations, we designed a relevant experiment. A Human Lying Posture Pressure-Map dataset (HLPPDat) formed during the research is publicly available. Involving 20 subjects in 64 defined postures, the classification output was separated into four basic groups included prone posture. The data enabled us to analyze multiple classification methods developed with the state-of-the-art concepts of Machine Learning (ML), sparse representation, and artificial intelligence represented by Transfer Learning (TL). The analysis included both data measured and data experimentally corrupted in a manner that would most probably occur due to a partial measurement error. Regarding the techniques and options tested, feature extraction via the Histogram of Oriented Gradient (HOG) and the K-Nearest Neighbors (KNN) tools appeared to be the most beneficial, yielding an accuracy of over 99.5% in the leave-one-subject-out crossvalidation. The research confirmed that an accurate classification is feasible at a matrix sensor resolution markedly lower than the limits regularly presented in the literature. The system allows monitoring how long a bedridden person has remained in the same posture, and thus it has a potential to help prevent decubitus in both hospitals and the home care.