Fall Risk Prediction via Classification of Lower Extremity Strength in Older Adults with Exergame-Collected Data

. Objective : The goal of this article is to present and evaluate a sensor-based falling risk estimation system. The system consists of an array of Wii Balance Boards (WBB) and an exergame that estimates if the player is at an increased falling risk by predicting the result of the 30 Second Chair-Stand Test (30CST). Methods : 16 participants recruited at a nursing home performed the 30CST and then played the exergame as often as desired during a period of two weeks. For each session, features related to how they walk and stand on the WBBs while playing the exergame were collected. Different classifier algorithms were used to predict the result of the 30CST on a binary basis (able or unable to maintain physical independence). Results : We achieved a maximum accuracy of 91% when attempting to estimate if the player’s 30CST score will be over or under a threshold of 12 points using a Logistic Model Tree. We also believe it is feasible to predict age- and sex-adjusted cutoff scores. Conclusion : An array of WBBs seems to be a viable solution to estimate lower extremity strength and with it the falling risk. In addition, data extracted while playing may form a basis to perform a general screening to identify elderly at an increased falling risk.


Introduction 1
Among older adults, falls are an important cause of mortality and early placement in nursing homes. The 2 main causes of falls are accidental and environment-related (31%) or caused by gait imbalances (17%). 3 Approximately 30 to 60% of older adults fall each year. Out of these falls, 10 to 20% result in injury, 4 hospitalization, or death. Among the most relevant factors to prevent these falls are risk assessment and 5 exercise [1]. The role of sensor-based solutions in regards to falling risk has traditionally been focused on 6 detecting said falls. Both wearable and smartphone-based solutions for fall detection are readily available 7 [2]. 8 9 In the field of training and health, exergames, active video games that incorporate physical movements, 10 aim to combine physical exercise with the fun associated with gaming. The main advantage of this approach 11 is that it increases motivation and thus adherence to training [3]. In addition, designing games that perform 12 fall risk prevention exercises or collect clinically meaningful data in the background, using sensors similar 13 to the ones already used to detect falls, is a positive factor [4]. An additional advantage of this approach, in 14 comparison to traditional exercise, is the possibility to adapt the exergame to the specific needs of the user 15 in real time and without external intervention based on game data [5]. 16

17
The possibility of using the Wii Balance Board (WBB) to estimate whether the player is at an increased 18 falling risk has been identified [6]. However, related publications mention the need for additional studies, 19 particularly in finding direct relationships between sensor data and clinically meaningful falling risk 20 estimation methods. Studies show that WBB data contain information that allow a discrimination between 21 elderly who previously fell and others who did not [7]. This study achieved an accuracy of 76.6% in this 22 binary classification on 12 participants. Early evidence also shows that the WBB could be used to train 23 balance in the elderly [8], and that there are statistically significant differences in the way elderly at falling 24 risk interact with the WBB. These differences actually correlate with clinical fall risk tests. Yamada et al.
[9] 25 found statistically significant differences and moderate correlations (r=0.69) in a study with 45 participants. 26 A limitation of the WBB is that, due to its small surface, it can only be used to estimate balance while 27 standing, and not in movement. 28 In a previous article, we presented PDDanceCity, a city map exergame with the goal of providing dual-29 tasked cognitive and physical rehabilitation [10]. The game is controlled with an array of six WBBs, which 30 we call Extended Balance Board (EBB) [11]. Thanks to its extended surface, EBB data can be used to 31 estimate the balance of the player both while standing and walking. We believe the data extracted from the 32 EBB could be used to estimate the balance and gait skills of the player in the background, without the need 33 to actively perform any specific test, or for any caregiver to be present. 34

35
In order to do so, this study aims to analyze the possibility of training a classifier to predict the falling risk of 36 a player based on EBB data collected in the background while playing PDDanceCity. This can be achieved 37 by attempting to predict the score of a standardized test that can be used to assess the falling risk. There 38 are several such tests to measure lower extremity strength, for example, the 30 second Chair-Stand test 39 (30CST) [12], which is part of the Fullerton Fitness Test Battery, and is fairly easy to administer. The 40 Fullerton Fitness Test Battery is commonly employed in older adults in community settings and can 41 measure physical patterns of physical decline in advanced ages. Evidence suggests it could also be used 42 as a screening test to estimate the falling risk and balance impairment in older adults [13,14]. The 30CST 43 classifies participants as subjects able or unable to maintain physical independence depending on whether 44 their test score is above or below an age-and sex-adjusted cutoff. We believe this binary prediction could 45 be achieved with a classifier algorithm using data extracted from the EBB. 46 47 Thus, the goal of this study is to determine the feasibility of classifying EBB-extracted data to perform a 48 binary prediction (player is able or unable to maintain physical independence). This prediction could be 49 used to detect when residents at an elderly nursing home may be losing physical indepencend and could 50 be more likely to fall in the near future. We validate this estimation basing the result on a prediction of the 51 30CST score. Data is collected while users are playing PDDanceCity to provide a very simple background 52 screening process determining whether the player may be at an increased falling risk. 53 54

Methods 55
PDDanceCity [10] is a labyrinth navigation exergame designed for dual-tasking rehabilitation. The goal of 56 the game is to navigate a labyrinth, representing a city map to reach a goal, where only two-dimensional 57 movements are possible (up, down, right and left). As an additional requirement, players are encouraged 58 to reach the target with the least possible number of steps. In addition, they may be required to visit a given 59 number of points of interest (for example a museum, monument or café) which may or may not be directly 60 on the shortest path ( Figure 1). The game offers a dual-tasking rehabilitation task, training visuospatial 61 function, memory, balance, and physical coordination. In order to use EBB data as a basis to control PDDanceCity, the center of mass ( ) is calculated as 70 follows. We define as the 6x4 matrix of sensor values (six WBB boards and four sensors per board), and 71 i,j ( ) as the value of sensor ( , ) of in instant . We define as the matrix of ( , ) coordinate vectors i,j 72 assigned to each sensor (Figure 3), based on its position. We also define ( ) as the last total weight value 73 calculated by all boards, that is, the weight of the player.
( ) is calculated as the weight-normalized 74 bidimensional projection of sensor values as: 75 This results in a set of two minus one to one values ( , ) which can be used to determine 78 intentionality. To achieve this, we define a directional intention based on two conditions: the main directional 79 component must be equal to or greater than 0.5 in magnitude, and the other component must be equal to 80 or lesser than 0.1 in magnitude. As an example, (0.1, 0.9) represents an upwards step, and 81 (−0.8, 0.05) would represent a leftwards movement. Between each step, the player is always required to 82 return to the center (both values lower than or equal to 0.1 in magnitude). Figure 4 represents two examples 83 of this directional intention. We also define the instability factor ( ) as an approximation of the first order 84 differential of ( ). This parameter is a measure of how a player shifts their weight on the EBB. A very 85 fast weight shifting, causing a high value of ( ), would be an indicator of potential lack of balance (or loss 86 thereof) among older adults who are not expected to move quickly. This is calculated as: 87 well as the number of times that ( ) overcame different possible thresholds. In addition to these two 98 elements, we also consider features related to the time intervals between steps, and the standard deviation 99 of these intervals. A complete feature list is presented in Table 1. All features are calculated per playthrough, 100 with no windowing. We used the Matlab software to calculate these features [15]. 101 102 To evaluate our system, we recruited 16 participants (median age 73, 6 males) at a nursing home in 103 Darmstadt, Germany. A computer was installed in a common room, connected to a television and the EBB 104 ( Figure 2). Participants were invited to play PDDanceCity as often as they desired for a period of two weeks. 105 During the first session, nominal data (age and sex) was collected, and the 30CST was administered. The    a single training instance was obtained. The data of 6 of these levels had to be discarded due to data failure, 118 leaving 81 training instances for classification. Due to the reduced number of participants, and to minimize 119 the risk of overfitting based on age and sex, we attempted to classify if the player's predicted 30CST score 120 was above or below a cutoff score of 12 points, without using these nominal data (age and sex) as features. 121 We refer to players classified above this cutoff as fit, and those under the cutoff as not fit. This score was 122 chosen to even out both groups, as eight participants had a 30CST score of 11 or lower. We also explore 123 the possibility of predicting the adjusted cutoffs, which we discuss at the end of the results section. All 124 classification tasks were performed using Weka [16]. 125 126

127
The best classification results are presented in Table 2. This decision tree used the average time between 128 steps exclusively, with a score of 6.17 or lower indicating a participant without an increased falling risk. A 129 comparison of different classification algorithms is presented in Figure 5. In all cases, we performed our 130 classification using ten-fold cross-validation. Results of a feature selection analysis (information gain 131 attribute evaluation) are included in Table 3. No features were excluded for classification. 132 133  As a second potential scenario of analysis, we also aimed to predict the age-and sex-adjusted 30CST 137 cutoff scores. The resulting accuracy was very high (99%) but, as discussed in the previous section, we 138 suspect that to be due to overfitting to age and sex because of our limited sample size, as the classifier did 139 achieve 100% accuracy using exclusively age and sex as features. If we remove these two features in this 140 scenario, we achieve a classification accuracy of 86% predicting the age-and sex-adjusted 30CST 141 outcome. For this reason, we believe that provided a large (and diverse) enough sample size of participants 142 of a wide array of ages and different degrees of fitness, it should be possible to predict the age-and sex-143 adjusted 30CST binary result using the methods presented in this publication. 144 145 We complemented this classification with an analysis of the effect sizes of each feature between the fit and 149 not fit groups, measured on the basis of Hedges' g due to the low sample size and the disparity in standard 150 deviations. We also evaluated statistical significance using a t-test. These effect sizes are presented in 151 Table 4. Features related to the instability factor and the mean and standard deviation of the time between 152 steps seem to contain the most information related to the 30CST. In accordance with Cohen's rule of thumb 153 (0.2 is a small, 0.5 a medium and 0.8 a large effect size), the effect sizes of these features are large, with 154 the differences between the fit and not fit groups being in most cases very significant (p<0.001) or at least 155 significant (p<0.05).   Despite our limited number of participants and training instances, we obtained excellent classification 164 results. Generally, decision trees seem to provide the best performance in the proposed classification task. 165 Our study also presents a design limitation due to using the 30CST as a validation method. Although we 166 decided on using the 30CST to minimize the risk of falls in study participants while conducting the test, this 167 test is less correlated to the risk of falling than other options such as the Berg Balance Scale. A future study 168 with a larger cohort should consider using this method instead of the 30CST to further support the 169 hypothesis that EBB data can be used to accurately identify participants at an increased falling risk. In  which means they have to be manually connected for each play session. They also operate on batteries, 175 and when these are low the data received is not reliable anymore. Additionally, the EBB frame presents a 176 risk depending on how the EBB is placed in its surroundings: if it is not set against a wall behind it, a player 177 may fall when taking a step backwards. We aim to address these technical limitations in a future iteration 178 of the EBB by providing direct electrical supply to the WBBs, automating the Bluetooth synchronization 179 process and building a complete enclosure around the EBB. 180 181

Conclusions 182
In spite of the aforementioned limitations, we believe our results suggest that the EBB, as an extension of 183 the WBB, can be used to screen the elderly population for individuals with an increased risk of falling, as a 184 basis to perform therapeutic and rehabilitation adjustments. Nevertheless, a larger dataset is required to 185 determine the feasibility of predicting if a participant will be above or below their age-and sex-adjusted 186