Multimodal and fusion based analysis of wearable sensor data for human health monitoring

In this study, in the field of health informatics, intelligent methods based on multi-mode and fusion-based analyzes have been developed in order to obtain meaningful information from the data collected from different sensors in order to monitor the health status and daily activities of the elderly and people in need of care. Pulse, EKG, oxygen saturation is used as wearable sensors. In addition to arrhythmia, sleep apnea diseases and early diagnosis of the data collected from these sensors, the physical activity of the person who was monitored by accelerometers, magnetometer, gyroscope and inclinometer sensors was investigated. In the scope of the thesis, physical activity data were collected from 12 different people according to the presence of indoor and outdoor. With the data collected, it was found where the person's physical activity was done. Standard deviations and exponential moving average based new features are created at this stage. It is observed that the performance of classification has increased in the studies conducted with the new features. Subsequent to our own dataset, the performance of the classification was also tested in sleep apnea and physical activity datasets to test the performance of the attribute building methods found later. It has been observed that the performance of classification has increased in studies with new attributes used with different datasets. The CNN + LSTM architecture we developed within the scope of this thesis was compared with the traditional classification methods. In the classification analysis, it has been found that performance increase is achieved with CNN + LSTM architecture.