Development of a multivariable prognostic PREdiction model for 1-year risk of FALLing in a cohort of community-dwelling older adults aged 75 years and above (PREFALL)

This was a study conducted in a general population setting within a single municipality.

Community-dwellers aged 75 years or above residing in Hjørring Municipality, Northern Jutland, Denmark, were studied. We recruited participants through preventive home visits (PHV), senior activity centers (SAC), along with senior clubs and associations. Participants were assessed for eligibility and included consecutively in PHVs and by convenience in the remaining recruitment sites. Exclusion criteria were living in care facilities, the presence of self-reported acute illness within seven days before recruitment, being unable to stand for one minute without any assistive device or support from another person, unable to understand Danish, or having a dementia diagnosis. The latter was checked through hospital records with informed consent from participants. Data collectors only recorded whether the older adults were eligible for participation or not due to time constraints.

We defined falls as “an unexpected event in which the participants come to rest on the ground floor or lower level” [22]. The primary outcome for prediction was falls rate within 12 months. Fall calendars with a daily recording of falls were returned monthly by post. If calendars had falls recorded, research secretaries contacted participants by telephone to validate and describe the falls.

The test battery has been described previously [23]. In brief, it consisted of physical tests of static balance under dual-tasking conditions, grip- and lower limb strength, reaction time of lower limbs, and habitual gait speed; a self-report questionnaire on demographic characteristics, frailty, nutrition, disability, fear of falling, health-related quality of life, depression, several physical symptoms, and a cognitive test. Static balance, reaction time, grip- and lower limb strength were measured using a Nintendo Wii Balance Board due to its valid and reliable measures and portable setup [24‐27]. Gait speed was measured using the four-meter gait speed test [28]. Furthermore, frailty was assessed using Tilburg Frailty Indicator (TFI) [29], nutrition using the short Mini Nutritional Assessment (MNA) [30], disability using Vulnerable Elders Survey 13, (VES) [31], fear of falling using short Falls Efficacy Scale International [32], health-related quality of life using EuroQol Five Dimensions Three Levels (EQ-5D-3L), depressive symptoms using the Geriatric Depression Scale with four items (GDS) [33], and the cognitive test using the Orientation-Memory-Concentration test [34]. In total, 34 candidate predictors corresponding to 41 degrees of freedom were available for model building. Additional file 1: Appendix 3. eTable 1 provides details of all candidate predictors, how they were defined, along with how and when they were measured.

Municipality staff and the first author performed the baseline data collection. Data collectors in PHVs and SACs were blinded to all predictors except the physical tests. Participants were blinded to all predictors except those in the questionnaire and the number of all prescribed drugs. The first author was blinded to physical tests performed in PHVs and SACs. Research secretaries assessing the outcome were blinded to baseline predictors.

A recommended method for calculating sample size was not available when the study commenced. Thus, we based the sample size on feasible recruitment within the 13 months inclusion period granted by the municipality.

During the study period, 6197 community-dwelling older adults (75+ years) lived in the municipality. Of these, we assessed 912 for eligibility, from which we included 241 older adults (66.4% women) with median (IQR) age 82 (80; 86) years. Figure 1 displays the flow of participants, and Table 1 provides the baseline characteristics of participants. Fifteen (6%) participants had missing data on predictors. Their demographic characteristics did not differ from participants without missing data.

The median (IQR) follow-up time for all participants was 365 (365–365) days. Forty-three (18%) participants had an incomplete follow-up. Median (IQR) age was higher for these, 85 (81–87) years, compared to participants retained in the study, 82 (79–86) years, (p = 0.01). Eighty-seven (36%) participants fell during follow-up, and 44 of these fell more than once. In total, we recorded 178 falls and calculated a fall rate of 0.94 falls per person-year. Table 2 displays univariate falls IRR for all predictors. Risk factors associated with falls were the completion of a short-cycle or medium-cycle higher education (IRR 2.97, 95% CI 1.80–4.74; IRR 2.00, 95% CI 1.36–2.92), drinking more than seven units of alcohol weekly (IRR 1.55, 95% CI 1.12–2.11), using walking aids (IRR 1.37, 95% CI 1.02–1.83), having three points on the Geriatric Depression Scale (IRR 9.5, 95% CI 5.34–15.70), having fallen within the last year (IRR 3.47, 95% CI 2.53–4.85), believing to fall within the next year (IRR 4.10, 95% CI 3.04–5.50), presence of dizziness (IRR 2.17, 95% CI 1.59–2.99), and increased balance sway area when dual-tasking (IRR 1.00, 95% CI 1.00–1.00). An increasing number of prescribed drugs (IRR 1.06, 95% CI 1.02–1.10), TFI score (IRR 1.13, 95% CI 1.06–1.20), Short Falls Efficacy Scale score (IRR 1.06, 95% CI 1.01–1.11), and VES score (IRR 1.13, 95% CI 1.07–1.18) were also associated with falls. Finally, use of multifocal lenses (IRR 0.60, 95% CI 0.44–0.83) along with increasing age (IRR 0.97, 95% CI 0.94–1.00), MNA score (IRR 0.84, 95% CI 0.78–0.91), and EQ-5D-3L score (IRR 0.19, 95% CI 0.09–0.40) were associated with a lower falls rate.

Figure 2 displays which predictors were most frequently selected for the prediction model in each bootstrap. Table 3 gives the fully specified prediction model and its performance measures. The final model included educational level, dizziness, weekly alcohol consumption, prior falls, self-perceived falls risk, disability, and depressive symptoms; all reported in a 22-item self-report questionnaire. For apparent performance, i.e., before internal validation, the MAE was 0.80 falls. After internal validation, MAE (95% CI) increased to 0.88 (0.71–1.16) falls. Additional file 1: Appendix 4. eTable 2 gives an example of how to use the model.

Other studies have developed falls prediction models for community-dwelling older adults with the same setting and falls definition as in the present study [11‐17]. Models were either purely questionnaire-based [13, 15, 17] or a combination of questionnaires and objective measures [11, 12, 15, 16]. Similarly, they found prior falls [12, 13, 15‐17], self-perceived falls risk [17], activities of daily living [12, 16], dizziness [12, 16], educational level [12, 16], alcohol consumption [12, 16], and depressive symptoms [11] to predict falls. In contrast to these studies, we did not find self-reported visual impairment [11, 13, 15], grip strength [12, 16], body mass index [11, 12], medication [13], or global cognition [11] to be predictive of falls. This discrepancy may be due to different participant characteristics along with assessment methods for outcomes and predictors. For participant characteristics, younger participants (> 65 years) were included [11‐13, 15‐17] resulting in a median age around 75 years for the majority of studies [11‐13, 16]. In contrast, our sample was older but smaller. Thus, our composition of predictors may only be representative for older adults (75+ years). For outcome assessments, we applied fall calendars, with a daily recording of falls, returned monthly, and outcome assessors blinded to baseline predictors. For the other studies, falls were recorded either yearly [15], quarterly [13, 16, 17], or weekly [11, 12, 14] which may increase the risk of recall bias and influence predictor selection for models. Also, only one study reported whether outcome assessors were blinded when validating falls according to their definition [17]. Thus, it is unclear whether detection bias influencing the predictive abilities of risk factors was present in the remaining studies. Lastly, differing measurement methods may cause different predictive abilities for the same risk factor across studies. These limitations can potentially lead to the inclusion of risk factors with poor predictive abilities in models. Consequently, model predictors may differ, and predictions may be less reliable when using models in clinical practice.

On a group level, we found 16 risk factors associated with falls through univariate analyses. However, only half of these were able to predict falls on an individual level. This emphasizes the need for clinicians to know when to best apply evidence from association and prediction modeling studies on falls since a statistically significant association does not imply that a risk factor is a strong predictor. For example, lower limb strength is strongly associated with falls [38] but may be less predictive of falls in community-dwelling older adults in the general population [12, 15, 16]. In our study, lower limb strength was not included in our prediction model, indicating no additional predictive value in the presence of the other candidate predictors. This was substantiated during our model development, where lower limb strength was less frequently chosen than the predictors in the final model. This finding is in line with three other prediction modeling studies measuring lower limb strength [12, 15, 16]. Here, only one study included a measure of lower limb strength in a final model but concluded that this was not superior to simpler questionnaire-based models [15]. Reasons for this could be because the test does not differentiate well between fallers and non-fallers as the average lower limb strength compares to the background population [39], and more functional tests of lower limb strength could be more relevant in terms of investigating for predictive performance. This knowledge may be applied when following the current recommendations for falls prevention [6]. Here, initial screening of falls risk is necessary for referring high-risk individuals to further assessment for modifiable risk factors and targeted interventions. Falls prediction models are appropriate for screening but not necessarily explaining falls risk. When presumed high-risk individuals have been identified, evidence on associated modifiable falls risk factors can explain why the older person is at risk and what to intervene on. In this scenario, self-report questionnaire models that predict falls may be both time-efficient and implementable due to their simplicity. Regarding time consumption, our model included the Vulnerable Elders Survey 13 [31], taking five minutes to administer [40], and nine questions giving a completion time of approximately 15 min. With current population prospects, it is highly relevant to identify fallers among older adults before they develop physical impairments that further increase their falls risk. However, our falls prediction model needs external validation before recommendation for clinical practice.

Since falls are multifactorial, they are also complex to predict. Hence, after externally validating our model, updating it with other predictors may further optimize performance. Other modeling studies have found static balance measures [14, 15] and comorbidities [11, 13, 15] to predict falls. We measured the latter but did not include this in the modeling process due to issues with implementability as diagnoses were not available for data collectors within our setting. Measures of executive and global cognitive function have also gained increased attention in falls research [41, 42]. Our study used a simple test for global cognitive function, Orientation-Memory-Concentration test, and executive function using dual tasking during balance. Similar to lower limb strength, none of these predicted falls, which could be due to the tests not differentiating well between fallers and non-fallers. After we commenced the study, recommendations for tests examining the interactions between cognition and motor function were published [43]. Here, the Montréal Cognitive Assessment [44], Trail Making Test [45], and dual-task gait measures were highlighted [43]. These tests may be more challenging for well-functioning community-dwellers, which could potentially increase the predictive performance of a model. Lastly, physical activity measures [11, 16] have been used in models, and accelerometry measures have shown promising results in falls prediction [46]. Future studies may consider updating our model using these predictors.

The study has a number of strengths. Missing data were sparse in the sample, the design ensured blinding within the study, and the follow-up period was long, thereby increasing chances of falls occurring. We used current recommendations on falls definitions and recording methods and reported the study transparently [20, 22]. There were also some limitations present. Half of the participants were not consecutively sampled, and the recruitment rate was low. This may have introduced selection and non-response bias that could influence predictor-outcome relationships and predictive performance when predicting falls in everyday clinical practice. This may explain the increase in MAE when performing internal validation. Furthermore, we collected predictor data within a week rather than at one point in time. This choice was a practical solution to collect data on as many predictors as possible without compromising recruitment. Predictive performance may have been influenced only to a minor degree since none of the participants fell during the first week after enrolment (data not shown), and questionnaire predictors were known to be stable over time [34, 47‐50]. Furthermore, the final model was questionnaire-based, and hence predictors could be available at the time of predicting. Another limitation of the prediction model was the VES not being validated in Danish. Regarding univariate analyses, increasing age and high educational level were associated with a lower and higher falls rate, respectively. These findings are in contrast to other studies, where increasing age is associated with falls, and higher educational level is not [4]. In terms of age, this surprising finding could be due to survival bias. However, we performed a subsequent comparison between the ones dying during follow-up (n = 3, median age (IQR): 81 (82,83) years) and not (n = 238, median age (IQR): 82 (80,86) years) and found no statistically significant age difference (p = 0.89). Therefore, we do not believe survival bias to be of relevance in this study. However, the univariate analysis may explain this association since no adjustments for confounders were made, along with age not being included in the final model. The latter indicates that age was found less predictive of falls in the presence of other candidate predictors in this study. Regarding educational level, a comparison was made across its categories in terms of length of follow-up time, and no significant differences were found (p = 0.85). These findings may be due to selection bias since the same levels of education were included in the final prediction model. Overall, the results need to be confirmed in external validation studies to assess transportability to other populations.