Método automático y adaptativo para la detección de anomalías en la actividad física mediante un sensor de aceleración no invasivo

Resumen

In recent years the number of people over 65 has increased considerably and it is expected to double in the coming years. This progressive ageing causes an increase in costs both in intervention and healthcare. In this context, information and communication technologies (ICTs) are emerging as a good opportunity for the design of new value-added services for monitoring elderly people as well as decreasing costs and increasing the quality of life. These kinds of solutions promote autonomy and home care as well as the decrease in other factors such as health care assistance or hospital readmission rate. This thesis addresses the problem of progressive ageing proposing a new method of user monitoring that can be used for the detection, control and tracking of diseases. The proposed method can further improve health care assistance and intervention supporting the informal caregivers work. This work identifies the main requirements needed to get a fully functional solution that provides a broad acceptance by end-users. The method proposed in this work has been validated in a real environment with four users monitored for three months. The method designed and validated in this thesis is composed of three separate sub-methods: an activity measurement method based on physical activity, a behaviour modelling method for identifying routines and an anomaly detection method. The combination of these methods creates a novel automatic and adaptive monitoring solution to detect anomalous human behaviour. The proposed activity measurement method is based on the measurement of physical activity using a wearable accelerometer. This method is based on a new and more efficient physical activity estimator than the existing so far. It can be implemented on any device that integrates an accelerometer. It has been validated with users with different activity profiles. By processing the measured physical activity, the proposed modelling method identifies the user routine automatically and adaptively. Unlike the main existing approaches, the proposed method requires no training or labelling. The routine is built from historical measures taking into account both "intra-daily" measures (measures within the same day) and "inter-daily" measures (measures of previous days). The proposed method has been validated in a real environment by monitoring users with different routines. The relationship between physical activity and the calculated routine is analysed in the anomaly detection method to detect anomalous behaviour. The term anomalous behaviour refers to the detection of unexpected levels of activity according to the expected user routine. A valuation function is built adaptively for each activity sample using fuzzy logic techniques. This valuation function gives each activity sample a score corresponding to the anomaly degree. The anomaly degree is calculated depending on the relationship between the routine and the measured activity. A final stage of filtering allows the customization of the method to detect only the relevant anomalous behaviour. The proposed method has been tested in a real environment by monitoring users with different profiles and it is able to identify automatically temporal sections of anomalous behaviour. The proposed methods have been validated in a real environment using a non-invasive accelerometer integrated in a device with a watch-appearance. Each of these methods has been individually validated and besides all of them have been validated as a whole by monitoring four older people for three months. The validation was based on the identified functional and non-functional requirements: reliability, invasiveness, autonomy, adaptability and automation. The proposed method has been further validated with a chronic patient under medical supervision. The proposed method automatically and adaptively detects anomalous behaviour by measuring physical activity. It has a high rate of accuracy and sensitivity and it is based on the use of a non-invasive and autonomous acceleration sensor.