The Study About Fall Risk Assessment And Fall Detection Based On Statistical Analysis Methods

In recent years, more and more people have been paying attention to the aging problem. Due to the degradation of lower extremity function, reduction of reaction and balance ability, and effect of own disease, elderly people could frequently occur fall. However, fall brings them life-threatening and physical pain. What is worse, fall has deep psychological impact on elderly people, which leads to the reduction of activity ability and degradation of health. The aim of our study is to reduce the risk of falling and prevent the occurrence of fall injury in elderly population. The research work presented in this dissertation includes:(1) The elderly fall risk assessment model was constructed with gait characteristics based on the kinetic data quantified by sample entropy. A total of 101 subjects, whom belong to Malianwa Street, aged above 50 years old and participated in questionnaire survey. In addition, the data of ground reaction force and moment was record when they performed walking at comfortable state. Participants were classified into three groups (high, medium and low risk group) according to the score of elderly fall risk assessment scale. The demographic variables, sample entropy of ground reaction force (GRF) and ground reaction moment (GRM), and impulse difference of bilateral foot were considered as potential explanatory variables of risk assessment model. Firstly, we investigated whether different groups could present difference in sample entropy of every potential explanatory variable. Statistical difference was found for the following variables:age(p= 0.0171); impulse difference(p= 0.0039); sample entropy of GRF in vertical direction(p= 0.0144); sample entropy of GRM in anterior-posterior direction(p= 0.0387). Then, we use age, impulse difference and sample entropy of resultant moment as explanatory variables to established multivariate ordered Logit model for fall risk level.(2) To detect real-time fall behavior, hidden Markov models with inertial sensor signal collected by micro-electromechanical (MEMS) is proposed. First, imitate the process of fall backwards of elderly, collect original inertial signals and establish two HMM models based on resultant acceleration and angular velocity. Then, calculate posterior probability of observation sequence as the degree of matching of the HMM model with the fall signal. At last, support vector machine is used to determine the maximum separation boundary between ADL and fall by calculating the degree of matching. The real-time detection for falls is realized by the process of sliding window.18 subjects participated in the experiment of fall and ADL. Each participant was required to perform 20 times for falls and 5 times for ADL Then, the 180 cases of falls was regarded as the training data of HMM model,108 cases fall and 216 cases ADL as the training data of separation boundary,72 cases fall and 144 cases of ADL as the test samples, which were used to establish model and recognition. The results shown that, the Correct Rate is 94.91%, Sensitivity is 97.22% and the Specificity is 93.75%, which has a further improvement in sensitivity of recognition compared with establishing model just for resultant acceleration.