Monitoring non-ambulatory posture and activity using moment statistics and Bayesian classifier

Providing an appropriate care plan for the elderly to participate in everyday life activities is important, as is developing the technological means to support physical assessments. Current clinical protocol is generally periodic and subjective stemming from expert assessments and interviews. This paper aims to present a probabilistic framework to efficiently model the posture and activity of non-ambulatory elderly patients, provide evidence for quantitative measurements and assist in the health related quality of life (HRQL) assessment. Pressure distribution data gathered from a developed sensor pad were projected and parameterized using moment statistics as features for static and dynamic activity modeling. The effect of posture angle was reduced by estimation of linear regression and a rotation matrix is used to realign the orientation of posture. A Bayesian classifier with Gaussian mixture model was adopted for posture recognition. A robust decision method based on minimum classification error was applied for the parameter estimation. Several objective evaluations and field trials were performed to investigate the detection performance of posture and activity. Our proposed approach outperformed vector quantization and shows encouraging potential for the development of HRQL indicators.