External validation of a 3-step falls prediction model in mild Parkinson’s disease

Falls and balance problems are common already early in Parkinson’s disease (PD) [1‐4] and progress over time [5‐7]. Avoiding falls and its consequences is a major goal and challenge in the management of PD [8]. Several predictive factors for future falls and near falls have been identified, e.g. history of falls and near falls, impaired balance, retropulsion, reduced comfortable gait speed, freezing of gait (FOG), cognitive impairments, pain, and fear of falling (FOF) (e.g. [9‐12]). However, prediction of falls is still a clinical challenge. For example, most available studies have identified predictive factors based on logistic regression models and associated odds ratios (ORs) (e.g. [10‐12]). However, ORs do not inform about the ability of predictors to discriminate between future fallers and non-fallers [13, 14] and are therefore not easily implemented in clinical practice [15]. It is therefore unclear exactly what components to consider. For example, history of falls was proposed as the strongest predictor in several prospective studies [16], whereas other observations suggest that history of near falls is a stronger predictor and may be seen as a precursor of falls [12].

Successful implementation of prediction models in clinical practice generally requires three main phases: model development, external validation, and investigations of their clinical impact; it is also recommended to consider whether existing models can be improved by, e.g. additional predictors [17]. Furthermore, in order to be applicable, useful and practical for routine clinical use, prediction models need not only to have sufficient ability to discriminate between future fallers and non-fallers, but also be easy and quick to implement. To this end, Paul et al. [18] proposed a 3-step falls prediction model (3-step model) consisting of three variables: history of falls during the past 12 months, history of FOG, and comfortable gait speed <1.1 m/s based on the mean of two trials of walking a standardized distance. Receiver operating characteristics (ROC) curve analyses found the 3-step model to discriminate between fallers and non-fallers over a prospective 6-month period with an area-under-the-curve (AUC) of 0.80 (95 % CI 0.73–0.86) [18]. These results were later replicated in a different PD sample (AUC, 0.83; 95 % CI 0.76–0.89) [19].

In contrast to other suggested fall prediction models, the 3-step model avoids reliance on relatively lengthy and time consuming clinical tests and assessments. For example, the model proposed by Kerr et al. [9] involves total scores of the Unified PD Rating Scale (UPDRS), the Tinetti scale and the FOG Questionnaire (FOGQ). The 3-step model therefore appears clinically promising and has been recommended in the European Physiotherapy Guidelines for PD [20]. However, it may be argued that measurement of gait speed according to the 3-step model may be considered cumbersome or difficult to achieve in clinical practice since it requires a standardized distance free of narrow passages, timing of two walking trials, and the calculation of the corresponding mean velocity as m/s. Meanwhile, common and easily conducted clinical PD tests such as pull-tests [21‐24] and Tandem Gait (TG) [23, 25] were not considered in the development of the 3-step model [18], although previously recommended for the prediction of falls [23, 24].

This study aimed to externally validate the 3-step model in an independent sample of people with relatively mild PD. In addition, we explored the ability of additional historical information and clinical tests to predict falls as well as near falls, and compared those with the proposed 3-step model.

The final sample (n = 138) is summarized in Table 1. During the 6 months of follow-up, 33 % (45/138) reported ≥1 fall and 14 % (19/138) reported only near falls (Table 2).

Testing the ability of the 3-step model to distinguish between individuals with and without future falls yielded an AUC (95 % CI) of 0.74 (0.65–0.84). Further details are presented in Table 3. Considering the suggested risk categories [18], 55, 48 and 32 people scored in the low, moderate and high risk categories, respectively. Of these, there were 7 (13 %), 18 (38 %) and 20 (63 %) who actually fell during the subsequent 6 months.

Simple logistic regression analyses of all available predictors and with falls as the dependent variable (Table 4) showed that near falls (AUC, 0.69) had the highest ability to distinguish between individuals with and without future falls. Rerunning these simple logistic regression analyses but with future falls and/or near falls as the dependent variable showed that a history of near falls and TG had the highest discriminant ability, both with an AUC (95 % CI) of 0.71 (0.62–0.80). Corresponding ORs (95 % CI) were 6.21 (2.91–13.25) for TG and 7.45 (3.32–16.70) for history of near falls (details available on request).

Multiple logistic regression analysis (backward method) using all available predictors as independent variables resulted in three significant predictors for the occurrence of future falls (Table 5): history of near falls, retropulsion (NRT), and comfortable gait speed <1.1 m/s, with an AUC (95 % CI) of 0.78 (0.70–0.86). Rerunning this analysis using falls and/or near falls as the dependent variable identified three predictors with an AUC (95 % CI) of 0.82 (0.75–0.89). The three predictors were (OR; 95 % CI): history of near falls (5.08; 2.04–12.66), retropulsion (NRT) (3.40; 1.26–9.14), and TG (4.41; 1.91–10.19) (sensitivity/specificity, 0.57/0.86, tolerance, ≥0.87 details available on request).

In this prospective study of individuals with relatively mild PD we externally validated the accuracy of a previously suggested 3-step model for prediction of falls [18, 19]. We found the discriminant ability of this model to be lower but acceptable and overlapping (given the 95 % CIs of AUCs) compared to previous studies [18, 19]. Importantly, discriminant abilities of each single predictor were lower and below acceptable values. This supports the value of the 3-step model over reliance on single predictors.

Different study samples have revealed some differences regarding the contribution of each predictor in the 3-step model. For example, in the development study [18], FOG was significant and associated with more than a two-fold increased odds of falling, while it was not significant in the subsequent [19] or in our study despite similar percentages of individuals reporting FOG (41–46 %) in all three samples. These discrepancies may be due to methodological aspects, as FOG was not assessed uniformly across the studies; Paul et al. [18] specified a retrospective time frame of 1 month, whereas both Duncan et al. [19] and we used a dichotomized version of item 3 of the FOGQ, which does not specify the recall period. Furthermore, information on history of FOG does not take FOG severity into account [38]. Similarly, differences in observed ORs for history of falls may relate to different modes of data collection. Specifically, whereas we and Paul et al. [18] inquired about the presence or absence of falls during the past year, Duncan et al. [19] combined data from two time points 6 months apart, where a question with five response categories was used.

Regardless of the cause(s) for the observed discrepancies in ORs of individual predictors, this has implications for the suggested scoring weights and risk categories of the 3-step model. That is, the weights (scores) suggested by Paul et al. [18] were based on the observed ORs in that study, which have not been replicated either here or by Duncan et al. [19]. It can be noted that the percentages of individuals who actually fell in our study was 13, 38 and 63 % in the low, moderate and high risk categories, respectively. Corresponding values in the study by Duncan et al. were 9, 28 and 66 % [19]. This is in contrast to the expected probabilities suggested by Paul et al. (17, 51 and 85 %, respectively) [18]. This calls for caution regarding the use of the suggested weighted total score. Further studies are needed to firmly establish a scoring system and risk categories.

The contribution of gait speed was relatively similar here as compared to the study by Paul et al. [18], despite differences in motor status according to the UPDRS part III (12 vs. about 24). Thus, comfortable gait speed <1.1 m/s is associated with approximately a two-fold increase in odds of falling regardless of whether a 4- [18] or 10-meter walking distance was used. This suggests the possibility to adjust the walking distance according to practical circumstances. However, the need to calculate the mean value for two trials should be evaluated in order to explore the possibility to simplify the test.

According to current recommendations regarding improvement of prediction models [17, 19] we explored the addition of history of near falls, retropulsion, and TG to the 3-step model. This generated a new model including history of near falls, retropulsion (NRT) and gait speed. The discriminate ability of this new model as well as its sensitivity of prediction was somewhat better compared to the proposed 3-step model but the AUC 95 % CIs overlapped. Similarly, using falls and/or near falls as the dependent variable generated a model including history of near falls, retropulsion (NRT) and TG. These observations have important clinical implications. Near falls are more frequent than falls in PD [39, 40] and may occur also among those who do not experience falls [39, 41]. We previously found, in the same project, that history of near falls but not falls was a risk factor for future falls [12]. This is further supported here and suggests that information about near falls may be a useful predictor of future falls. Furthermore, since near falls may be seen as an early precursor of increased fall risk [42, 43], it is argued that prediction of falls and/or near falls has greater clinical value than prediction of falls alone. This is also in line with previous studies highlighting the importance of fall risk identification before the first fall has occurred, in order to optimize planning of interventions [9, 16]. From this perspective, our new model (history of near falls, TG and retropulsion according to the NRT) may be considered a promising alternative to the suggested 3-step model, at least among people with milder PD. Indeed, the use of TG and NRT has been recommended in the prediction of falls before [23, 24, 44]. However, this suggested new model needs further confirmation in additional studies.

NRT, but not UPDRS item 30 was identified as a predictor in both new models. UPDRS item 30 involves prior instructions, which does not mimic daily life circumstances where perturbations per definition are unexpected [44]. Accordingly, the unexpected pull test according to the NRT has been considered more relevant in the context of fall prediction [24], which is supported by our findings.

This study confirms the value of the 3-step model as a clinical fall prediction tool and illustrates that it outperforms the use of single predictors. However, further studies are needed to firmly establish a scoring system and risk categories based on this model, and to better understand the influence of methodological aspects of data collection regarding gait speed and history of falls and FOG. A new model for prediction of falls and near falls, including history of near falls, TG and retropulsion according to the NRT is considered a promising alternative to the 3-step model in milder PD, but needs to be tested in additional samples. Taken together, our observations provide important additions to the evidence base for clinical fall prediction in PD.