1 INTRODUCTION

The costs of falls, both financial and non-financial, are well-recognised¹ and, with the ageing of populations in many countries, research into the factors that influence the risk of falls has assumed an increasing urgency.² A large volume of published research has investigated possible contributory factors covering a multitude of domains, ranging from demographics, physical function, disease, medications, mental and cognitive function to lifestyle.³ A review of the literature by Ambrose et al.⁴ published in 2013 identified impaired balance and gait, polypharmacy and a history of previous falls as major risk factors, together with a host of other less-significant contributory factors.

A systematic review of fall prediction models published in 2021 found that ‘all models exhibited a high risk of bias rendering them unreliable for prediction in clinical practice’.⁵ Deficiencies have been identified in the falls literature around the definition of falls and recording of falls occurrence⁶ and attempts have been made to standardise the reporting of outcomes⁷ and establish protocols for transparent presentation of prediction model results.⁸ Given that proprioception, specifically ankle proprioception, has been found to be directly associated with two of the factors identified by Ambrose et al. as major risks for falls – balance⁹ and gait¹⁰ – a prospective study examining the role of proprioception in falls that complies with these recommendations around data capture, analysis and reporting is warranted.

Proprioception was described by Sherrington,¹¹ a pioneer in this field, as the perception of the position and motion of the body in space. It is a complex mechanism, relying on a network of mechanoreceptors in the muscle spindles, joint tendons and glabrous skin surfaces, communicating through the afferent and efferent neural pathways, with cognitive function playing a key role in integrating these inputs with other sources of information from the visual and vestibular systems.¹² In the context of falls, the evidence suggesting that, relative to young and middle-aged adults, older adults rely more heavily on proprioception rather than visual or vestibular cues in order to maintain postural control¹³ assumes particular importance.

Attempts to measure proprioceptive skill directly have relied on three broad approaches: joint position replication; time to detection of passive motion; and active movement extent discrimination. Each of these approaches has advantages and disadvantages¹⁴ and is thought, to some extent, to be measuring different aspects of proprioception.¹⁵ For the current study, we used an active movement extent discrimination assessment (AMEDA) apparatus designed to measure proprioception of the ankle joint with a testing protocol developed for an older population.¹⁶ As described in the Appendix S1, the apparatus allows a subject's ankle to be placed in different degrees of inversion and tests their ability to discern the direction of movement between successive inversions. The AMEDA has the advantage of testing skills in a weight-bearing unconstrained stance that is more akin to the situations in which proprioception is exercised in daily life. It has not previously been investigated as a possible predictor of fall risk.

The Sensory Organisation Test (SOT) measures proprioception and neuromuscular control in the context of an individual's ability to maintain balance under disruptions to the visual, vestibular and somatosensory systems. It has been used to measure the effectiveness of interventions to mitigate fall risk¹⁷ but not as a predictor of falls.

The aim of this study was to establish if there was a relationship between performance on the AMEDA and SOT tasks and subsequent falls. We hypothesised that poorer performance on these tasks would be associated with a higher risk of falls. A secondary hypothesis was that performance would be worse among individuals who fall in the least challenging circumstances. Where relationships were found, we planned to develop a prediction model complying with the TRIPOD guidelines⁸ that could be further tested in older adults to identify individuals who might be at greater risk of falls, noting that the relatively small number of participants means that the study should be treated as a pilot of the approach and analytic techniques adopted. A completed TRIPOD checklist for the study is included in the Appendix S1.

2 METHODS

2.6 Data analysis

R²² was used for all data analysis.

2.6.1 Data preparation

Previous research has drawn attention to differences between types of falls. For example, Kelsey et al.²³ report on differences between falls occurring inside and outside. As the factors that might lead to a fall are likely to vary with the type of fall, we adopted a method of classifying falls developed and validated in the context of monitoring falls among people suffering from Parkinson's disease.²⁴ The classification system is quite nuanced in that it takes account of the challenges arising from both the environment and the activity being undertaken (see Figure 1). As an example of a Type 1 fall, where both the environment and activity are challenging, one participant fell while walking backwards carrying a roll of fencing wire. By contrast, a Type 3 fall might be one occurring in the participant's home when undertaking a simple activity such as turning to change direction. A typical Type 2 fall might be a trip when walking on an unfamiliar footpath.

Prior to any analysis, the classification paradigm was applied independently by two of the authors, who were both blinded to the initial test results. A reconciliation process was undertaken for those cases where they had arrived at a different classification, again without any knowledge of the test results.

As Figure 1 shows, falls can be considered to lie along a continuum in terms of the challenge of the environment or activity. We defined three binary outcome variables by setting the fall definition at different points along this continuum; the first is whether a fall of any type occurred during the 12 months following initial testing; the second is whether a fall of Type 2 or 3 was experienced and the third limits the definition to just Type 3 falls. The predictor variables were as follows:

• the AMEDA score, derived as described in Antcliff et al.¹⁶ A score of 1 corresponds to the average performance, with a score of less than 1 indicating poorer performance and, conversely, a score of more than 1 indicating superior performance.

• the SOT composite score, calculated as:

  $$\left\{\sum _{i=1}^2\sum _{j=1}^3\frac{{\text{Score}}_{i,j}}{3}+\sum _{i=3}^6\sum _{j=1}^3{\text{Score}}_{i,j}\right\}/14$$

where: i is the test condition as set out in Table S2; and j is the iteration.

A higher score indicates better performance.

• the answers to the yes/no questions in the fall risk questionnaire as shown in Table S3.

2.6.2 Missing data

Partway through the initial testing of participants, we discovered that the force plate in the computerised posturography machine used for the SOT was faulty and that the measures that had been taken for the first 27 participants were impacted. The affected participants were all retested within 4 months. However, SOT scores have been found to improve on retesting²⁵ and this was seen to be the case for those not affected by the equipment malfunction. To adjust for this improvement, we used the second reading for those whose initial reading was affected, adjusted down by the average learning effect observed in the unaffected group.

One participant was unable to undertake the SOT test because the harness could not be fastened and another was unable to maintain their balance in the final segment of the SOT test, which precluded the calculation of their composite score. These two participants were excluded from the data used for the t-tests for the SOT composite score. The logistic regression model fitting process restricts the analysis to complete cases and thus these two participants were also excluded from the models that included the composite score. All participants were included in other analyses.

2.6.3 Model selection and validation

Our model selection process was guided by an examination of the relationships between the predictor variables and the relevant outcome variable using t-tests for the two proprioception measures and chi-square tests for the questions in the fall risk questionnaire. For each of the outcome variables, we determined whether a logistic model using the variables found to be associated with an increased risk of falls showed any power to predict those falls. For the purposes of evaluating model performance, we used a likelihood ratio test, three pseudo-R² measures, balanced accuracy and the area under the curve (AUC). In deciding whether to include an additional predictor variable in the model, we had regard to the Akaike information criterion (AIC).

As we did not have sufficient data to implement a standard cross-validation approach, the models were fit on the entire dataset, but the sensitivity of the results was tested using leave one out cross validation (LOOCV) and bootstrapping, with predictions generated from the fitted model using the ‘predict’ function in R.

3 RESULTS

3.1 Number and classification of falls

A total of 38 (69%) of the 55 participants who completed at least 12 months in the study experienced at least one fall within their first 12 months after entry into the study (see Figure 2 for the breakdown by fall type), with 105 falls being recorded in total.

4 DISCUSSION

The main hypothesis was supported in that proprioceptive function, as measured by the AMEDA, was significantly associated with Type 2/3 falls and the most significant predictor of Type 3 falls. This is not surprising: a lower AMEDA score indicates poorer ankle proprioception which has been found to be associated with impaired balance¹⁰ and Type 3 falls are defined by the absence of any challenges arising from the environment or the activity. That is, they are fundamentally related to poor intrinsic balance. The question from the fall risk questionnaire, eliciting information on balance confirms this; none of those who experienced a Type 3 fall reported being able to put on their socks while standing on one leg. Fear of falling was also clearly important with none of the non-faller having a fear of falling and a noticeably higher percentage of Type 3 faller having a fear of falling relative to the other two fall categories. This is likely to be a two-way relationship, with a fear of falling increasing the risk, and falls themselves leading to a well-founded fear of falling.

Participants reporting Type 3 falls represented only 15% of the cohort and only 20% of those who experienced a fall (see Figure 2). However, they were much more likely to have multiple falls, including Type 1 and Type 2 falls. Of the 30 participants who had either a Type 1 or Type 2 fall, but not a Type 3 fall, 19 had only one fall in total. By contrast, all but one of the participants who had a Type 3 fall had multiple falls. In total, the eight participants with a Type 3 fall accounted for 50 of the 105 falls. Thus, predicting those with an increased risk of Type 3 falls could be expected to identify those who are at greater risk of all types of falls and, given the greater number of falls, an increased likelihood of a fall with adverse consequences.

While we observed lower SOT scores for those with Type 2 and Type 3 falls, the difference was insufficient to achieve significance at the 5% level once an adjustment was made for multiple comparisons, and we found no benefit from including the composite score in the logistic model. The problems with the apparatus which affected the reliability of the initial measurement would also support the exclusion of this variable. A previous study comparing the SOT scores and recalled falls history found an association between a lower composite score and the report of one or more falls in the previous 6 months.²⁷ However, this was a retrospective study, and the SOT measurement was taken after the falls and so may have been affected by the fall rather than being predictive of a fall.

The remaining questions on the fall risk questionnaire operated broadly in the direction that would be expected. For example, those who fell were less likely to be able to walk 1 km without stopping and more likely to have been admitted to hospital than those who did not experience a fall during the 12 months. The differences in absolute numbers were small, however, and not statistically significant.

5 CONCLUSIONS

The falls classification used here provides a more nuanced understanding of falls and how their characteristics and the factors driving them might vary. While all types of falls can be potentially catastrophic, Type 3 falls are likely to be of most concern given the likelihood that a person experiencing a Type 3 fall will have multiple falls. The ability to put on one's socks while standing on one leg appears to protect against a Type 3 fall, with none of those who were able to do this suffering a Type 3 fall during the period they were monitored. However, over half of the participants were unable to meet this benchmark and most did not have a Type 3 fall. The additional information from the AMEDA score and the question regarding fear of falling allowed the fitted logistic model to achieve high sensitivity and specificity in predicting Type 3 falls in the study population. We found that poorer performance on the AMEDA was the most significant predictor of a Type 3 fall in the following 12 months, while a fear of falling, probably justified, was far more common among those experiencing a Type 3 fall than among other participants.

This study points to the potential to use the AMEDA both in future research into falls mitigation and in a clinical setting to diagnose and measure the rehabilitative effects of interventions.

CONFLICT OF INTEREST STATEMENT

No conflicts of interest declared.