Evaluation of Wearable Sensors as an Older Adult Fall Risk Assessment Tool
MetadataShow full item record
Falls are common in the geriatric population, with approximately one third of older adults falling each year. Falls can result in lasting physical and psychological consequences and cost approximately $20 billion per year in the United States. Wearable sensors can be used for quantitative, gait-based, point-of-care fall risk assessment that can be easily and quickly implemented in clinical care and older adult living environments. The objectives of this study were to evaluate eyes open and eyes closed static posturography in older adults; provide in-depth analysis of the differences between single-task and dual-task gait in elderly individuals and the relation to faller status; generate models for wearable-sensor-based fall risk classification in older adults and identify the optimal sensor type, location, combination, and modelling method for walking with and without a cognitive load task; and compare wearable-sensor-based fall risk classification performance to clinical assessment-based and posturography-based fall risk classification outcomes. A convenience sample of 100 older individuals (75.5 ± 6.7 years; 76 non-fallers, 24 fallers based on 6 month retrospective fall occurrence; 47 non-fallers, 28 fallers based on 6 month prospective fall occurrence with retrospective fallers excluded) walked 7.62 m under single-task (ST) and dual-task (DT) conditions while wearing pressure-sensing insoles and tri-axial accelerometers at the head, pelvis, and left and right shanks. Participants also completed the Activities-specific Balance Confidence scale, Community Health Activities Model Program for Seniors questionnaire, six minute walk test, static posturography with eyes open and closed, and ranked their fear of falling. Fall risk classification models were assessed for all sensor combinations and three model types: multi-layer perceptron neural network, naïve Bayesian, and support vector machine. Feature selection was performed using Relief-F, Fast Correlation-Based Filter (FCBF), and Correlation based Feature Selection (CFS). For static posturography, measures sensitive to anterior-posterior motion and medial-lateral centre of pressure (CoP) velocity were greater under eyes closed compared to eyes open conditions for prospective non-fallers, fallers, and multi-fallers. For prospective multi-fallers, medial-lateral range and root-mean square distance from the mean were also greater when visual input was removed, suggesting that assessment of medial-lateral balance control may be particularly important for evaluating the risk of multiple falls. Differences were found between prospective fallers and non-fallers for Romberg Quotient (RQ) anterior-posterior range and root-mean square distance from the mean. Differences between prospective multi-fallers and non-fallers were for eyes closed and RQ anterior-posterior and vector sum magnitude velocity. This suggests that RQ calculations are particularly relevant for elderly fall risk assessments. Measures that changed between ST and DT walking conditions, including non-temporal measures related to movement frequency and abnormal body segment movements, were identified. Increased gait variability under DT conditions was indicated by increased posterior CoP stance path deviations, medial-lateral CoP stance path deviation durations, and CoP stance path coefficient of variation; and decreased Fast Fourier Transform quartiles and ratio of even to odd harmonics. Decreased gait velocity and decreased pelvis and shank acceleration standard deviations (SD) could represent compensatory gait strategies to counter the increased gait variability and thus maintain stability. Differences between prospective fallers and prospective non-fallers were related to movement frequency and variability. Fall risk classification models that used Relief-F feature selection achieved the best performance. With feature selection, the best model for prospective faller classification contained ten features (four pressure-sensing insole features, six left shank accelerometer features) and used a support vector machine classifier. This model achieved an accuracy of 94%, F1 score of 0.923, and Matthew’s Correlation Coefficient (MCC) of 0.866. The posterior pelvis accelerometer provided strong single-sensor performance (83% accuracy, F1 score 0.769, MCC 0.645), although lower than the best multi-sensor model performance, and should be considered if a single-sensor system is necessary to reduce assessment cost and complexity at the point-of-care. Neural networks and support vector machines both achieved strong classification performance and outperformed naïve Bayesian classifiers. Sensor-based models outperformed clinical assessment-based models and posturography-based models for both retrospective and prospective fall risk classification. Wearable sensors provided strong fall risk classification performance and should be considered for point-of-care assessment of elderly fall risk.