Prediction of the Intensity of Physical Movement of Aged People

Real-time health condition monitoring can bring huge benefits to the well-being of aged people. These technologies can be made using video cameras and other highly precise sensing systems. However, this poses a potential threat to personal privacy and high installation and maintenance costs. Research communities have investigated low-cost gas sensor systems to detect the physical movement of elderly people. Therefore, health conditions can be monitored remotely. The key challenge for these systems is to measure the physical movements using the sensor data which are highly prone to noise. In this paper, we explored various machine-learning techniques to predict the movement intensity of aged people from sensor data. In particular, we investigated i) feature generations with statistical and time domain behaviours of sensor values, and ii) removing noise from sensor data before applying to predict the movement intensity. For noise reduction, three different approaches were tested including rolling average of sensor data, sensor true value prediction, and Gaussian estimate of sensor values. Processed data is then used to generate around 700 features which provided 10-I5 strongly correlated features to build prediction models using LR, DT, RF, XGboost, GNB, and ensemble method. All five prediction models also built a voting based ensemble model. The lowest prediction accuracy 67% was found when the LR model was trained with raw data while high accuracy 78% was achieved when true sensor value prediction technique were used for the ensemble model. This study shows that sensor based movement intensity prediction accuracy is significantly improved by the machine learning models when noise reduction technique is used.