2011

Pärkkä, Juha

Personal health monitoring refers to the long-term health monitoring that is performed in uncontrolled environments instead of a laboratory, for example, at home or by using wearable sensors. The monitoring is done by individuals alone, usually without guidance from health care professionals. Data produced by personal health monitoring (for example, actigraphy, heart rate, etc.) are currently used more in personal wellness monitoring rather than in clinical decision-making, because of challenges in the interpretation of the long-term and possibly unreliable data. Automatic analysis of long-term personal health monitoring data could be used for the continuous recognition of changes in individual's behavior and health status, and to point out which everyday selections have a negative effect on health and which have a positive effect. This can not be achieved by using sparse measurements in controlled environments. In this thesis, data analysis was carried out for the recognition of physical and mental load using data from wearable sensors and other self-measurements. Large, annotated data libraries were collected in real-life or realistic laboratory conditions for the purpose of the development of practical algorithms and the identification of the most information-rich sensors and signal interpretation methods. Time and frequency domain features were computed from raw sensor data for the correlation analysis and the automatic classification of the personal health monitoring data. The decision tree, artificial neural network, K-Nearest Neighbor and a hybrid of a decision tree and artificial neural network classifiers were used. Automatic activity recognition aims at recognizing individual's activities and postures using data from unobtrusive, wearable sensors. Similarly, the unobtrusive, wearable sensors can be used for the assessment of energy expenditure. The quantities measured in this thesis include acceleration, compass bearings, angular rate, ECG, heart rate, respiratory effort, illumination, temperature, humidity, GPS location, pulse plethysmogram, skin conductance and air pressure. The results indicate that several everyday activities, especially those with regular movements, can be recognized with good accuracy. The energy expenditure estimate obtained using movement sensors was found accurate in activities involving regular movements. The sensors that react to the change of activity type without delay were found the most useful for activity recognition. These include accelerometers, magnetometers, angular rate sensors and GPS location sensors. Automatic assessment of mental load aims at measuring the level of mental load during everyday activities using data from wearable sensors. The assessment of long-term stress aims at finding measures that reflect the perceived stress level, either directly or as observed through changes in behavior. Data were collected with people suffering from long-term work-related stress and participating in a rehabilitation program. Automatic measurements of recovery, measured with a bed sensor, actigraphy and bedroom illumination sensors were found to correlate best with the self-assessed stress level. Careful selection of sensor types, sensor locations and input features played a more critical role in successful classification than the selection of a classifier. Computational complexity of the classifier's classification phase has an impact on the power consumption of a hosting mobile terminal. Power consumption is one of the bottlenecks in long-term personal health monitoring solutions today.