Evaluation of human standing balance using wearable inertial sensors: A machine learning approach

According to recent studies, mobility and balance problems represent a risk factor for elderly people, resulting into augmented fall probabilities with negative impact on medical, social and economical aspects. Posturographic analysis plays a major role in clinical applications, enabling the evaluation of standing balance and the quantification of specific risks. Wearable devices represent a viable option to support diagnostic practice, but several issues need to be addressed to design systems with the required reliability. In this work, we propose a solution that hinges upon signals that could be gathered in principle from sensors on board of wearable devices. In particular, we propose to extract meaningful features from the accelerometer and gyroscope data streams that could be processed and exploited to feed machine learning algorithms (namely Support Vector Machines) and recognize determined standing balance tasks. Specifically, we put forward to exploit time-domain, frequency-domain and structural features that are usually used for processing signals from plate force devices and we augment them with those obtained from the gyroscope. Experimental results underline the effectiveness of the proposed method, which could effectively classify subjects involved in standing balance exercises and also discriminate between specific tasks reaching average performance levels of 87% in terms of precision, 86% in terms of recall, 87% in terms of F1 score, and 92% for what regards accuracy. The proposed study confirms the possibility of developing hardware-software systems that could represent affordable, flexible, yet meaningful solutions to assist, therefore, monitoring activities at fine grain resolution.