Participants

A total of 140 hospitalized patients who suffered their first stroke were enrolled according to the following criteria: (1) stroke confirmed on MRI or CT; (2) ability to walk independently (with/without assistive device) at least 10 meters; (3) no fall history within 1 year before stroke onset; and (4) written informed consent.

Only patients who met the above entry criteria were included to specifically identify the fall risk factors related to stroke.

A ten-meter walk test was included because it is a valid, reliable assessment for predicting falls in subjects with stroke[12,18]. This test collected dynamic gait parameters as predictors of fall. Therefore, subjects who walked with a person’s assistance had an external supportive force that may interfere with the assessment of gait, and these patients were not included in the present study.

No fall history within 1 year before study entry was selected because subjects who experienced falls may exhibit subsequent factors while walking independently, including fear of a fall, decreased mobility, and changes in gait pattern. Subjects with a history of falling previously may also have their movements closely monitored by their family members or caregivers [12]. These conditions may confound the relationship between variables and fall risk assessment.

Subjects dropped out during the follow-up for the following reasons: nursing home residency (11 subjects), unstable internal disease (8 subjects), unable to complete the interview due to dementia or severe cognitive impairment (4 subjects), loss of contact due to residence address changes (3 subjects), and epilepsy (2 subjects). Therefore, 112 subjects completed the study, and these subjects were further divided into two groups, faller or non-faller, depending on whether a fall occurred during the study period.

Baseline measurements

Initial assessments were performed at baseline including demographic data and a standardized recording of history and clinical examinations. Baseline physiological and psychological assessments were performed before subjects were discharged from the hospital (approximately 1 month after stroke). These baseline measurements of physical and psychological parameters were used as fall risk factors to develop the fall risk prediction model.

Physical assessments included the Modified Ashworth Scale (MAS), which was used to assess muscle tone in the elbow flexor, knee extensor, and ankle plantar flexor [19]. The level of a patient’s disability was assessed according to the Functional Independence Measure (FIM) [20]. Performance of activities in daily life was assessed during hospitalization according to eighteen items, including bathing, dressing, toileting, transferring, urinary continence, cognitive comprehension and social interaction. This assessment is widely used to measure and predict outcome [21].

The objective measurements of gait were completed using computerized systems with wearable inertial sensors. Subjects wore customized shoes (Ultraflex, Infotronic, the Netherlands) with eight load sensors in each shoe to measure the forces under the foot and detect temporal events in the gait cycle prior to the gait measurements. Data were sampled at the rate of 100 Hz and stored in a portable (Walkman size), lightweight data logger that was carried on the lower back of each subject. Several practice tests were performed before actual data were collected. Subjects walked at a self-selected speed over a 10-m hallway. The mean values of the two tests were used. Gait parameters were normalized to the subject’s body height to account for possible effects of anthropometrics [22,23]. The temporal asymmetry ratios (ASY) for single support time (ASY_ss), double support time (ASY_ds), single swing time (ASY_swing), stance time (ASY_stance), and step time (ASY_step) were quantified using the following equation [24]: A greater value of this ratio indicates higher asymmetry between the two sides.

The objective computerized measurements of balance were completed while the subjects stood on a Stabilo-platform (Ultralfex, Infotronic, the Netherlands) in a comfortable position without footwear or ankle foot orthoses. Subjects kept their eyes open and arms at their sides and were instructed to maintain their balance for 20 seconds [25,26]. Three tests were performed with a 30-second rest between tests. The mean value of three tests is presented. Subjects’ performances were recorded as the center of pressure (COP) trajectory paths. Data were sampled at the rate of 100 Hz, and COP stability was calculated as the standard deviation of the anterioposterior (COP_ap) and mediolateral (COP_ml) directions of the points obtained during measurement. The sway area (COP_area) was calculated as the square root of the sum of squares of the COP_ap and COP_ml.

Psychological evaluations included the Mini-Mental State Examination (MMSE) [27] and Chinese translated version of the Geriatric Depression Scale (GDS) [28], which were used to screen for cognition and depression, respectively. The modified Falls Efficacy Scale (mFES) was used to evaluate the fear of falling in stroke patients [29].

Assessment of falls

Falls were defined as incidents when the subject came to rest on the floor due to an unexpected loss of balance. All subjects were followed up for 6 months after the first assessment to collect the record of falls. Trained research nurses visited the subjects at home 4, 12, and 24 weeks after discharge from the hospital or rehabilitation ward. Phone reports from subjects were also encouraged in this study to prevent errors from retrospective data collection.

Statistics

Descriptive analysis was used for all variables, and results are presented as the means, standard deviations and percentages. Significant differences between fallers and non-fallers were assessed using independent Student’s t-test for continuous variables and χ² test analysis for categorical variables. The Mann-Whitney U test was used to detect mean differences between groups when variable distributions were not normal. Linear correlations between continuous variables were calculated using Pearson’s correlation test. Multivariate logistic regression (MLR) analysis was performed using a forward stepwise method with an entry criteria of P = 0.1 to identify the factors that were independently associated with falls. Two models were developed based on variables with statistical significance from bivariate analysis and clinical interests. Adjusted Odds Ratios (aOR) were acquired from the estimated coefficients and presented with the corresponding 95% confidence interval (CI) of the ratio.

The predictive accuracy of the model in discriminating fallers and non-fallers was assessed using sensitivity and specificity. The optimal cutoff point with the highest sensitivity and specificity for each model was defined as the Youden index [30,31].

A receiver operating characteristics (ROC) curve was plotted to assess the discrimination of the generated multivariate logistic models. The area under the curve (AUC) of the ROC was also calculated for each model to determine the fitness of individual MLR analysis. An AUC value below 0.5 was considered no discrimination, 0.7≦AUC≦0.8 was considered acceptable discrimination, 0.8≦AUC≦0.9 was considered excellent discrimination, and 0.9≦AUC≦1.0 was considered outstanding discrimination [32]. Commercial statistical software, SPSS version 13.0, was used, and a two-tailed P < .05 was considered significant.

Assessment of falls

Previous studies reported that the “gold standard” for collecting information on falls (e.g., prospective collection with calendars or postcards, regular reminders, and follow-up telephone calls) was prone to errors (e.g., memory, forgetting to write diaries and ambiguous definitions of fall) [14]. To minimize these types of errors in this study, falls were recorded regularly by nurses during home visits 4, 12, and 24 weeks post-discharge and by subjects’ self-report. Recordings of fall history, environmental risk exam, and medical consultations were performed during the interviews with each subject. One advantage of the interview was to provide better interaction between subjects and research team workers. Therefore, subjects could fully understand the risk of falls and the ultimate goal of this study to prevent fall occurrence.

Balance and gait performance

Poor postural balance was linked to increased fall risk in previous studies [11,35,36]. Mediolateral COP displacement during normal standing may be used as an indicator of accidental falls in the elderly because it was significantly associated with future falls [37]. The results of this study also demonstrated that postural sway in the mediolateral direction and area were greater in fallers compared to non-fallers.

A hemiparetic gait is described as slow and asymmetrical [38,39]. Walking velocity and cadence were lower in the faller group than the non-faller group in the present study. Gait speed is generally selected as the outcome measurement in clinical practice and a predictor of fall after a patient has a stroke, but gait speed is often confounded with balance, motor function, and endurance [40]. The current study adopted a quantitative gait analysis to help assess the risk of fall and further describe gait performance adequately.

A previous study reported that temporal gait symmetry measurement appeared to better reflect components related to weight shift, and it was superior to spatial symmetry ratios for identifying the risk of falls in impaired ambulators [38]. An “asymmetry ratio” was used to represent the level of temporal asymmetry (ASY_ss, ASY_ds, and ASY_step) in the present study, which was significantly different between fallers and non-fallers. An increase of one standard deviation in ASY_ss was associated with a 2.2 and 3.6 times higher fall risk in models I and II, respectively.

Spasticity related to falls

An asymmetrical gait pattern caused by impaired balance and abnormal muscle tone is commonly seen in stroke patients. The present results demonstrated that the severity of spasticity in the upper and lower extremities was markedly higher in the faller group compared to the non-faller group. These findings are consistent with another study that also reported spasticity as a risk factor for predicting falls in chronic stroke patients. Motor control and functional status of stroke patients declined with increasing spasticity [13]. Logistic regression model I in this study also demonstrated that spasticity of the gastrocnemius was a predictor of fall in stroke patients. Another study found that the degree of spasticity of the affected ankle plantar flexors primarily influenced gait asymmetry [24]. A spastic gait in stroke patients diminished power generation, decreased hip and knee flexion during the swing phase, and reduced stability during the stance phase due to the affected hip flexors, knee extensors, and ankle plantar flexors [41]. The present study revealed that the risk of fall increased 3.2 times when the severity of spasticity in the gastrocnemius increased by one grade. The results also support that the combination of gait asymmetry and abnormal muscle tone may increase falls in the stroke population.

Previous studies reported that spasticity reached a peak within 1–3 months after a stroke [41–45]. Thirty-nine percent of patients who suffered a first stroke exhibited sustained spasticity after 12 months [19]. Therefore, the early detection of spasticity and improvement in motor dysfunction using specific interventions, such as stretching, splinting, electrical stimulation, and botulinum toxin injection, may be crucial to reduce accidental falls [46].

Effects of functional performance

The Functional Independence Measure (FIM) is widely used to evaluate the performance of a patient’s daily activity to determine the level of a patient’s disability. All functional performance assessment results were significantly higher in non-fallers than fallers in our study. The fall risk decreased by 10% when the FIM score increased by one point. Previous studies also reported the significant correlation between FIM and fall occurrence [47–49]. A previous 10-year retrospective study also demonstrated the same correlation between the FIM score and fall risk [50], which is consistent with present results. However, the FIM score as a single variable may not be sufficient to accurately predict fall risk because falls generally resulted from multiple factors. The finding is also consistent with our MLR model, which enhanced the sensitivity and specificity of fall prediction.

Effects of depression in stroke patients

Depressive symptoms are common in the acute phase after stroke, and symptoms are associated with the persistence of depression and mortality after 12 months [51]. The current MLR model in this study demonstrated that the risk of fall increased 1.4 times with a one-unit increase in GDS. Moreover, fall risk may cumulatively increase due to a high cognitive load if the patient also had multiple motor impairments and depression combined with gait asymmetry and spasticity.

Fall prediction model

A bivariate correlation between risk predictors of accident fall in stroke subjects was performed to determine which variables to include in the MLR analysis. The results of bivariate correlation test revealed that the FIM and GDS exhibited the highest strength of negative correlation (Table 4). Therefore, two logistic regression models, including FIM or GDS, were developed in this study to determine the best fit of fall predictive factors. Notably, model I, which included GDS, gait asymmetry, and spasticity, exhibited slightly higher specificity, sensitivity, and Youden index than model II. Gait asymmetry, spasticity, and depression represented the functional, physical, and psychological domains of the subject’s impairments in function, respectively. Therefore, these results suggest that model I provides more comprehensive fall prediction than model II. ROC analysis further verified the discrimination of fitness of model I with a slightly greater AUC value (0.856) than model II (AUC value 0.815).

A previous study model with six predictors, including the Berg Balance score and functional performance, exhibited high predictive values (AUC = 0.712) in community stroke patients [52]. Both models in the present study used three predictors and demonstrated AUC values greater than 0.8. Several differences were observed between the two studies, including the race of enrolled subjects, time of assessment of falls post-stroke, and the selection of variables for MLR analysis. The present study included computerized gait asymmetry as a predictor in model II in addition to balance and functional predictors. The predictive values of model II reached AUC = 0.815 despite the inclusion of only three predictors. Computerized gait assessment was included in both models, and the computerized system may provide a more objective and accurate evaluation. Overall, these findings suggest that gait asymmetry is an important factor for the prediction of falls.

An earlier study also found that sideways fall was an independent risk factor (aOR, 2.5; 95% CI, 1.6–3.9) for hip fractures in the elderly, in which 20% of their population had a history of stroke. Therefore, preventing sideways fall may decrease the occurrence of hip fractures in the elderly [53]. Both prediction models in this study included gait asymmetry as one predictor of a fall. Model II included gait asymmetry and balance factors. Therefore, the fall prediction models, including computerized gait and balance assessments, may be used in stroke patients and the elderly for preventing possible falls.

Study limitations

Several limitations may result from the present study design. First, the present study results cannot be extrapolated to all people with stroke, particularly patients at lower functional levels with walking disability or severe cognitive impairments. Second, the subjects were not separated into a construction data set and a validation data set to test the multivariate logistic regression function because of the small number of subjects enrolled.