Falls are the second leading cause of death from unintentional injuries, especially in adults aged 60 and over. The financial costs and contributory factors of falls are high. Many preventive measures exist, but more tailored solutions are needed. This MSc thesis focuses on the development of a personalized algorithm that serves as a proof-of-concept to classify an individual's walking pattern using an accelerometer on a smart bracelet. The system can distinguish deviant walking from reference walking. The MSc thesis includes the acquisition of a dataset with accelerometer signal data simulating the gait of elderly individuals, the development of a semi-supervised learning algorithm and the implementation of the algorithm in an embedded system. The use of the Mahalanobis distance metric was found to be the most appropriate method for separating deviant from non-deviant walking. The analysis showed that the algorithm could work for 82.35% of all persons in the recorded dataset. In a practical test of the implemented algorithm in an embedded system, the algorithm performed with an accuracy of 87.5%, specificity of 75% and sensitivity of 100%. The robustness of the algorithm is questionable, especially with large changes in walking speed or changes in hand movements. The MSc thesis has shown that it is possible to distinguish deviant walking from non-deviant walking. However, challenges remained in extracting domain features from the wrist-mounted accelerometer due to noise caused by hand movement. Further research should investigate methods to compute domain features from the wrist-mounted sensor and develop a more reliable algorithm for step detection. This technology could be extended into a system that would distinguish abnormal walking from normal walking in elderly users and, in the event of a significant deterioration in walking quality, warn healthcare professionals of an increased risk of falling.
|
