Prospective Fall-Risk Prediction Models for Older Adults Based on Wearable Sensors

Abstract

Wearable sensors can provide quantitative, gait-based assessments that can translate to point-of-care environments. This investigation generated elderly fall-risk predictive models based on wearable-sensor-derived gait data and prospective fall occurrence, and identified the optimal sensor type, location, and combination for single and dual-task walking. 75 individuals who reported six month prospective fall occurrence (75.2 ± 6.6 years; 47 non-fallers and 28 fallers) walked 7.62 m under single-task and dual-task conditions while wearing pressure-sensinginsoles and tri-axial accelerometers at the head, pelvis, and left and right shanks. Fall-risk classificationmodels were assessed for all sensor combinations and three model types: neural network, naïve Bayesian, and support vector machine. The best performing model used a neural network, dual-task gait data, and input parameters from head, pelvis, and left shank accelerometers (accuracy = 57%, sensitivity = 43%, and specificity = 65%). The best single-sensor model used a neural network, dual-task gait data, and pelvis accelerometer parameters (accuracy = 54%, sensitivity = 35%, and specificity = 67%). Single-task and dual-task gait assessments provided similar fall-risk model performance. Fall-risk predictive models developed for point-of-care environments should use multi-sensor dual-task gait assessment with the pelvis location considered if assessment is limited to a single sensor.

Faller Classification in Older Adults Using Wearable Sensors Based on Turn and Straight-Walking Accelerometer-Based Features

“…The types of populations (retrospective or prospective fallers; single-fall or multiple-fall fallers) and methodologies vary, and differ from the current paper. The prospective fall prediction study in [26] permits comparison based on the identical older-adult population. The turn-feature based classification results in this paper (73.4% accuracy, 60.5% sensitivity, 82.0% specificity, and 0.44 MCC score) were better than the best straight-walking classification results in [26] (56.5% accuracy, 42.5% sensitivity, 65.4% specificity and 0.083 MCC score), based on similar accelerometer derived features for 25 ft walk single-task and dual-task tests, and similar cross validation with 10,000 random stratified splits.…”

“…The prospective fall prediction study in [26] permits comparison based on the identical older-adult population. The turn-feature based classification results in this paper (73.4% accuracy, 60.5% sensitivity, 82.0% specificity, and 0.44 MCC score) were better than the best straight-walking classification results in [26] (56.5% accuracy, 42.5% sensitivity, 65.4% specificity and 0.083 MCC score), based on similar accelerometer derived features for 25 ft walk single-task and dual-task tests, and similar cross validation with 10,000 random stratified splits. Those results for straight walking were similar to the straight-walking-based classification results in Test II of this paper (55.5% accuracy, 46.1% sensitivity, 61.8% specificity and 0.08 MCC).…”

“…Those results for straight walking were similar to the straight-walking-based classification results in Test II of this paper (55.5% accuracy, 46.1% sensitivity, 61.8% specificity and 0.08 MCC). Since the straight-walking classifier performances for [26] and this paper were similar, it is likely that the use of turn-based features was the main contributing factor to improved classification results (turn compared to straight), rather than the inclusion of more walking sections (6MWT in this paper compared to 25 ft walk test in [26]). This strongly suggests that turn-based features provide better information for prospective faller classification and thus faller prediction.…”

“…These features were carried forward in further research demonstrating their effectiveness for faller classification [25,26]. Temporal and acceleration descriptive statistics provided direct measures of body motion related to gait.…”

Analysis of Optimal Sensor Positions for Activity Classification and Application on a Different Data Collection Scenario

“…In [21], three feature selection algorithms were examined, i.e., Relief, Simba, and mRMR, and it was reported that these algorithms yielded similar feature importance, especially for the first four highest ranked features. Relief-F [47], which is an extended version of Relief [48], was reported as the best feature selection algorithm in [49] compared to Fast Correlation Based Filter and Correlation Based Feature Selection. It was one of the most widely used feature selection algorithms, with low computational time [50] and the ability to deal with incomplete and noisy data, and can be used for evaluating feature quality in multi-class problems [51].…”