Abstract:
Objective: To investigate the predictive validity of simple gait-related dual task tests in predicting falls in community-dwelling older adults.
Design: A validation-cohort study with six months follow-up.
Setting: General community.
Participants: Sixty-six independently ambulant community-dwelling adults aged 65 years or older, with normal cognitive function. Sixty-two completed the follow-up. No participants required frames for walking.
Interventions: Not applicable.
Main outcome measure: Occurrence of falls in the follow-up period and performance on primary and secondary tasks of eight dual and one triple task tests.
Results: A random forest classification analysis identified the top five predictors of a fall as 1) absolute difference in time between the timed up and go as a single task and while carrying a cup; 2) time required to complete the walking task in the triple task test; 3 & 4) walking and avoiding a moving obstacle as a single task and while carrying a cup and 5) performing the timed up and go while carrying a cup. Separate bivariate logistic regression analyses showed that performance on these tasks was significantly associated with falling (p < 0.01). Despite the random forest analysis being a more robust approach than multivariate logistic regression, it was not clinically useful for predicting falls.
Conclusions: This study identified the most important outcome measures in predicting falls using simple dual task tests. The results showed that measures of change in performance were not useful in a multivariate model when compared to an “allocated all to falls” rule.
Key words: Dual task, older adults, falls, gait, predictive validity
List of abbreviations:
DT: Dual Task
ST: Single Task
TT: Triple Task
BMI: Body Mass Index
MMSE: Mini-Mental State Examination
EXIT-15: abbreviated executive interview
HADS: Hospital Anxiety and Depression Scale
POMA: Performance Oriented Mobility Assessment
FES-I: Falls Efficacy Scale International
TUG: Timed Up and Go
WWT: Walking While Talking
ProFaNE: Prevention of Falls Network Europe
IQR: Interquartile Range
RF: Random Forest
OOB: Out-Of-Bag
Introduction:
Many Dual Task (DT) tests have been developed in the recent years. One important area of DT research is the assessment of falls’ risk, with poor gait performance in DT conditions predicting falls in older adults living in a variety of settings1. Dual-tasking is one form of executive functioning and deterioration in executive functioning has been associated with falling2.
In general, slowing down under DT conditions can predict falls in older adults1,3. Reviews conducted to assess the predictive validity of DT tests for falls showed conflicting results1,4. Moreover, DT tests seem to be better in predicting falls in frailer older adults 1,5. Studies conducted in the community setting have had conflicting results. Some have shown validity in predicting falls2,6,7,8,9,10,11, some have shown no added value over Single Task (ST) tests3 and some showed no predictive validity12,13,14,15. The contradictory results might be explained by the heterogeneity of the tasks used, the outcomes measured and the duration of falls’ follow-up periods1,4,16. Furthermore, most of the cognitive tasks used in the tests lack ecological validity17.
Research in the area of dual tasking has suffered from the scarcity of assessment of secondary task performance4. This is a problem if older adults vary in which tasks they prioritize under DT conditions. Some participants might concentrate on gait and potentially avoid falls. Others might prioritize the secondary task, which might increase the falls risk. The use of complex methods to measure gait or balance performance, such as force-platforms and electronic walkways,9,14,15 has been a feature of most studies, making these difficult to perform in the clinical setting. Most studies have assessed raw ST and/ or DT performance as a predictor of falls2,8,15. However, raw performance (e.g. walking speed) does not provide a measure of the specific impact of dual-tasking (since most of the variance in walking under DT conditions is accounted for by walking under ST conditions). The use of a measure of proportionate change in task performance is recommended to account for differences in baseline ST performance18. Unfortunately, in the community setting only Yamada et al.10 has used such a measure.
In previous studies, we have investigated the feasibility and reliability of a variety of DT combinations in community-dwelling older adults, and a set of 9 tests was identified19,20. The primary aim of this study was to assess the predictive validity of these simple gait-related DT and Triple Task (TT) tests in predicting falls in community-dwelling older adults. Uniquely, this study has performed a comprehensive assessment of the predictive validity of several clinically feasible DT tests, by measuring performance on both primary and secondary tasks and considering both the absolute and proportionate differences in performance to identify which tests and which measures would be the best predictors of falls in community-dwelling older adults.
Methods:
Participants
A convenience sample was recruited through a community falls prevention programme and friends and family of staff at a university. To be eligible to take part in the study, participants had to be aged 65 years or older, community-dwelling, speaking and understanding English, able to travel to the assessment laboratory, have a Mini-Mental State Examination (MMSE) score of ≥ 24 and be able to maintain their feet together and adopt the semi-tandem stance of the four-test balance scale for 10 seconds21. Exclusion criteria included the use of walking frames and uncorrected visual or hearing impairments.
Ethical approval
Ethical approval for this study was granted by the west of Scotland research ethics committee (Reference Number: 10/S0701/82) and the ethics committee at the university. The participants were provided with an information sheet explaining the study and asked to sign consent forms.
Procedures
Participants attended a single data collection session. Sample baseline characteristics were recorded using a self-reported health questionnaire adapted from Greig et al.22, Body Mass Index (BMI), MMSE23 , four-test balance scale21, abbreviated Executive Interview (EXIT-15)24, Hospital Anxiety and Depression Scale (HADS)25, Performance Oriented Mobility Assessment (POMA)26 and Falls Efficacy Scale-International (FES-I)27.
Eight DT and one TT tests were used to predict falls. These tests had previously been found reliable in a sample of community-dwelling fallers and non-fallers20. These tests were:
- Straight walking and visuospatial clock task
- Walking with turns and naming animals
- Walking with turns and counting backwards in 3’s
- Avoiding stationary obstacles and naming animals
- Avoiding a moving obstacle and carrying a cup
- Timed Up and Go (TUG) and carrying a cup
- Stair descent and naming animals
- Walking While Talking (WWT) complex6
- Straight walking, visuospatial clock task and carrying a cup (TT test)
The visuospatial task required deciding if the two hands of an imaginary clock are on the same or different sides of a line drawn between the 12 and 6 point28.
Primary task walking time was measured with a stopwatch, secondary task performance was recorded and assessed as performance speed (total answers/ second) and accuracy (errors/ total answers). Participants were instructed to walk at their preferred speed and to perform both tasks as well as they could. A rest period was offered after each test to prevent fatigue. Participants performed the tasks as STs first. Secondary STs were performed in a seated position for 30 s and ST performance was adjusted individually for each task combination to be equivalent to the DT time. The order of the STs and DTs was chosen randomly to avoid performance bias. The cup task was assessed dichotomously as no spill or spill.
Follow-up
A six month follow-up period was implemented to collect incidence of falls. Participants were provided with monthly falls’ diaries as recommended by the Prevention of Falls Network Europe (ProFaNE)29. These diaries were sent back at the end of each month via pre-paid envelopes. A brief telephone interview was conducted in the event of a fall to inquire about the circumstances and consequences of the fall.
Data processing and analysis
In addition to recording ST and DT performance, the proportionate difference for primary tasks was calculated using the following equation:
Proportionate difference = ((Performance on ST - Performance on DT)/Performance on ST)*100%
It was not always possible to calculate the proportionate difference for secondary tasks as frequently subjects made no errors and sometimes gave no answers in ST. In that situation, the absolute difference was used instead.
Absolute difference = Performance on ST - Performance on DT
Number of falls25 was collected and participants were divided into fallers (one or more fall) and non-fallers (no falls). Primary task and secondary task performance in ST, DT and the proportionate and absolute differences were used to predict falls. Falls in the previous year, age, gender, performance on FES-I, EXIT-15, HADS and POMA were also used as predictors. In total there were 96 predictors.
SPSS version 18 for windows and R version 2.15.0 were used to analyse the data. Demographic characteristics are presented as median and interquartile ranges (IQR). Normality of the data was checked using the Kolmogorov-Smirnov test. Differences between fallers and non-fallers in terms of sample characteristics was assessed using independent sample t-tests and Mann-Whitney U tests. Nominal data such as gender were assessed using the Chi-square test.
Regression analysis was originally planned for examining predictors. However, the limitations of small sample size (caused by slower than anticipated recruitment and consequent time and funding limitations), relative to the large number of predictors, meant that instead, Random Forest (RF) classification analysis was conducted. RF classification can robustly estimate parameters by repeated simulation on “training” subsets of the data (67%) to grow the “forest”, which are then assessed on a separate “test” subset of the data (33%)30. Classifications trees were used to predict a target variable (e.g. fallers/ non-fallers) based on multiple input variables. RFs provide two explanatory variable utility measures: mean decrease in accuracy and mean decrease in Gini, which are used for ranking variables31. Mean decrease in accuracy computes the importance of a variable for classification by measuring the change in the accuracy when this variable is omitted, and similarly mean decrease in Gini impurity is a measure of allocation error when a variable is removed. Mean decrease in Gini is a more reliable than mean decrease in accuracy31. The error rate in the RF classification is called Out-Of-Bag (OOB) error rate.
Only the top five variables were tested using binary logistic regression against the outcome. The differences between fallers and non-fallers on the five top variables were assessed. Using Bonferroni’s adjustment (0.05/5) the level of statistical significance was set at p < 0.01.
Results:
Sample
A sample of sixty-six community-dwelling older adults took part. Two subjects dropped out due to health-related factors (i.e. Stroke and cancer diagnosis) and two subjects expressed lack of interest in completion. One subject passed away after completing four months of the follow-up period and was included in the analysis. The final sample, therefore, consisted of sixty-two subjects. The falls rate was 21% of the sample (n = 13), with a total of 25 observed falls. Most falls occurred in the house. Only one fall required medical attention for prescription of pain medication.
The sample’s characteristics are described in table 1. At baseline fallers were older (borderline significant), with greater fear of falling (FES-I), poorer executive function (EXIT-15), poorer balance and gait (POMA) than non-fallers. Fallers maintained the single leg stance of the four-test balance scale for a shorter duration (median = 8.00 seconds, IQR = 5.00) than non-fallers (median = 10.00 seconds, IQR = 0.50) (p = 0.027). Eighteen of the non-fallers (36.7%) and ten of the fallers (76.9%) (p = 0.01) had a history of falls in the previous 12 months.
Random Forest (RF) classification
The predictions reported here were based on 3000 trees. Table 2 shows a “confusion matrix” explaining the details of the error rate in this classification model. The OOB estimate of error rate was 27.4% indicating a correct classification rate of 72.6%.
The five most useful variables in this classification were: The time required to complete TT, avoiding a moving obstacle and cup, TUG and cup, ST avoiding a moving obstacle and the absolute difference in time between ST TUG and DT TUG (see figure 1).
Although the RF model only had an OOB error rate of 27.4%, this is still worse than the error rate of a model that would simply ignore the data and allocate all subjects to the non-faller group (13/62 = 20.9%) irrespective of their measurements.
Logistic regression:
The results of the regression analysis for the top five variables are presented in table 3. In general there was an agreement between the RF classification results and the logistic regression analysis. The Odd Ratios (ORs) for the variables were statistically significant at p < 0.01. The ORs indicate that a longer walking time in TT, avoiding a moving obstacle as an ST or DT and performing the TUG while carrying a cup are associated with a higher risk of falling. An absolute difference in TUG time in the negative direction, indicating slower DT performance, was associated with higher risk of falling. There was a statistically significant difference (p < 0.01) between fallers and non-fallers on the five top variables identified through the RF classfications (see figure 2).
Discussion:
This study aimed to investigate the ability of DT and TT tests to predict future falls in community-dwelling older adults. The results showed that simple DT and TT tests might be useful in predicting falls in community-dwelling older adults.
The RF classification analysis identified the five candidate variables with the greatest impact on the accuracy of the model. The OOB estimation of error rate for this model was high (27.4%) when compared to the error rate of the model allocating every subject to the non-fallers group (13/62 = 21%). This indicates that this analysis was unable to find a useful predictive tool on this limited dataset. Nevertheless, it allowed for a more valid and robust analysis method.
The ORs produced by the multivariate logistic regression varied from 0.61 to 1.29 and were statistically significant (see Table 3). When comparing the ORs from the five variables with those of other falls predictors, it seems that they are slightly smaller but still within the range of what has been previously found. In a review by Rubenstein et al.32 visual deficits had an OR ranging from 1.1 to 7.4 and postural hypotension and OR ranging from 1.0 to 3.4. In terms of other DT studies, the ORs from the current study seem to be in line with the results of Faulkner et al.33 who reported and OR of 1.34 (95% CI 1.04-1.74) for walking time in a DT condition.
It is perhaps not surprising that the time required to complete the TT test had the greatest impact of on the accuracy of the RF model. Older adults perceived the performance of more than two tasks simultaneously as more difficult and risky than dual tasking34. The results for the TT test contradict the results by Makizako et al.8, who showed that a TT did not predict future falls. Makizako et al.8 used a probe reaction time task, stepping in place and counting backwards in 1’s. This test might have had a greater cognitive than motor load as two of the tasks were cognitive. The current TT test had a higher motor load as it combined gait, upper limb motor and visuospatial tasks.
The time required to avoid a moving obstacle in ST and DT were amongst the top variables in both the regression and RF analysis. Older adults identified avoiding moving obstacles while walking as difficult and fall inducing34. This task involved an unexpected interruption to walking, simulating everyday life situations35. It might be that avoiding a moving obstacle is equally valuable in both ST and DT conditions as the ST version includes elements of dual tasking, as it required the avoidance of a ball that crossed the walking pathway.
The TUG combined with a motor cup task was the only DT that has been previously used to predict falls. Lundin-Olsson et al.36 found that an absolute difference in TUG time between TUG and cup and ST TUG of ≥ 4.5 seconds had an OR 0.21 (=1/4.7) 95%CI = 0.07-0.66 for predicting falls in older adults living in sheltered accommodation. A sample size of forty-two subjects was tested by Lundin-Olsson et al.36 and thirteen subjects (31%) had one or more falls in the six month follow-up period. Although a similar follow-up period was applied in the current study, the observed falls rate was lower (21%). However, the greater falls rate in the study by Lundin-Olsson et al.36 might have been due to the subjects being more frail than those who took part in the current study. The participants, in the current study, with a previous history of falls were taking part in falls exercise classes and although the classes did not involve specific DT elements, the nature of the group exercises might have improved DT performance. The OR for prediction of falls using the TUG combined with a motor task in this study (0.61) compared to that previously reported (OR 0.2136) seem to corroborate the differences in functional ability in the two groups of participants.