Development of an inertial system for gait analysis

Nowadays ageing population is increasing constantly so specific needs are required to be solved in order to improve the quality live of our elders. Frailty which affects ageing population is intrinsically related to falls and both have negative consequences like risk of hospitalization, institutionalization, reduction of quality of live and mortality. Moreover, frailty and falls share along other pathology’s an intrinsic relationship with gait. World Health Organization, suggest the implementation of preventive strategies, focused on reducing the fall risk and improve health status of our elders. The implementation of these strategies must be focus on firstly the detection and evaluation of those subjects at risk and secondly the implementation of a schedule for monitorization and rehabilitation of those subjects. Frailty and falls risk are currently assessed by specialist with traditional metrics performed in hospitals. However these metrics tend to be subjective due to the use of basic tools. Beside the great benefits of new technologies, they are not currently widespread for the assessment of gait and functional performance. This study propose the development of an inertial system for gait analysis in real time and synchronization capabilities between two devices. A wearable device for gait assessment in non controlled environments is proposed. In addition the complexity of gait makes it difficult for traditional algorithms to detect gait phases in a wide set of different subjects. To solve this problem an artificial neural network for gait phases classification with inertial data has been developed. Adaptive algorithms like artificial neural networks outperform traditional algorithms in duties of classification of a wide set of data and parameters. Different structures and hiper-parameters for the development of the artificial neural network are tested and optimized, to find the best suited solution for gait phase detection in real time. One key aspect of the training and validation of these networks is the pre-processing of data, feeding the network during the training process with high quality and diverse set of data to these networks, so as to avoid over-fitting problems. Final model for gait phase detection makes use of a sequential model with three structures stacked; convolution layers, LSTM layers and dense layers, for a final accuracy of 98.28%. Sensor performance is tested through a series of test to analyze delays in synchronization, or lost of synchronization through time. This synchronization between devices is necessary for the estimation of double support parameter. Synchronization results show that there are no delays in the synchronization event. In addition the synchronization is still maintained after a ten minute test, where the delay of the clock is below the sampling rate of the IMU (104Hz). Noise analysis and calibration process are also performed to study the error of the selected IMU model and improve estimations through calibration process. This noise analysis found that sensor specifications in the data-sheet are close to the values obtained. In addition the noise of this sensor was compared with other sensors used for gait assessment showing great results. Finally, algorithms for the estimation of gait parameters, that can be estimated with IMU placed on the foot, have been developed in order to carry out rich assessments of gait in real time