Supervised machine learning scheme for electromyography-based pre-fall detection system

Falls are the leading cause of disability and death among the elderly. Over the years, several inertial-based wearable devices for automatic fall and pre-fall detection have been devised. Under controlled condition, these systems show a high performance for unbalance detection (up to 100% of specificity and sensitivity), however the mean lead time before the impact is about 200-400 ms. Although this period of time is enough to active an impact reduction system (i.e wearable airbag) to minimize injury, it is necessary to increase it so as to improve the system efficiency and reliability. A user's muscle behavior analysis could be more strategic than that of a kinematic evaluation one, permitting a rapid recognition of an imbalance event. This also holds true for several research studies on muscles response during a state of imbalance, whereas a limit number of them deal with the development of wearable electromyography (EMG)-based systems for human imbalance detection, suitable for predicting a lack of balance in real time. With respects these limitations, the main purpose of this work has been the development of a low computational cost expert system for real time and automatic fall risk detection. The analysis of this is achieved through lower limb muscles behavior monitoring. A Machine Learning scheme has been chosen in order to overcome the well-known drawbacks of threshold approaches widely used in pre-fall systems, in which the algorithm parameters have to be set according to the users' specific physical characteristics. Ten kinds of time-domain features, commonly used in the analysis of the lower-limb muscle activity, have been investigated. With a view to reducing the processing complexity, the Markov Random Field (MRF) based Fisher-Markov feature selector was tested. It showed a high degree of accuracy in the EMG-based features selection for lack of balance detection. The Co-Contraction Index, Integrated EMG and Willison Amplitude features have been also considered. The supervised classification phase has been obtained through a low computational cost and a high classification accuracy level Linear Discriminant Analysis. The developed system shows high performance in terms of sensitivity and specificity (about 90%) in controlled conditions, with a mean lead-time before the impact of about 775 ms. Therefore, the feasibility of a quick and wearable surface EMG-based unbalance detection system, by using Machine Learning methodology, has been demonstrated. The system may recognize a fall event during the initial phase, increasing the decision making time and minimizing the incorrect and inappropriate activations of the protection system, in real life scenario.