Past Gait Speed as an Independent Predictor of Mortality in Older Adults Beyond Current Gait Speed

We investigated whether past values of gait speed in older adults provide additional prognostic information beyond current gait speed alone. We assessed various models to best describe past and current value for prediction. Methods We used data from the rst ve yearly rounds of the National Health and Ageing Trends Study, starting from 2011. The cohort consisted of 4289 community-dwelling participants aged 65 years and older. Gait speed was measured at baseline (Y1) and one year later (Y2). Three-year follow-up for mortality started in year 2. We estimated hazard ratios of various models using combinations of Y1 gait speed, Y2 gait speed, and change in gait speed from Y1 to Y2. This study aimed to investigate the use of past gait speed as a predictor of mortality in community-dwelling, older adults. Our results show that change in gait speed over time, beyond current gait speed, is an independent predictive measure of mortality among this cohort. Although current gait speed had a stronger statistical association with mortality, changes in gait speed in relation to baseline was also found to be a signicant predictor of mortality. Best model ts were achieved by including gait speed alongside changes in gait speed.


Introduction
Decline in mobility is associated with mortality and other adverse outcomes, such as functional disability and falls in older adults (1)(2)(3). Of the various measures of frailty and physical performance, gait speed has been recommended for incorporation into the clinical assessment of older adults for its favorable properties (continuous interval properties, low examiner training, strong interrater and intrarater reliability, low material resources required, and minimal learning effect) and its gross indication for overall health (4)(5)(6). Since survival is in uenced by other factors beyond chronological age or chronic conditions, physical function, as assessed by gait speed, is increasingly used to personalize and improve clinical decision-making by assessing current and predicting future health status of older adults. (7).
As a predictor for morbidity and mortality, gait speed has mostly been studied as a single measure in time, without considering past gait speed values. Previous studies have demonstrated a robust association between current gait speed and survival (3), while others have shown that baseline gait speed is a predictor for adverse outcomes and increased disability in older adults (2,8). Gait speed however, is not a xed measure and varies over time, particularly in older adults (9); and may uctuate as a consequence of acute events, such as falls, hospitalizations, vascular events, or of chronic conditions, such as arthritis, Parkinson's disease, or sarcopenia (10)(11)(12). The association between change in gait speed and time when conditioned on baseline gait speed may provide additional prognostic information beyond baseline gait speed alone to predict adverse outcomes. The use and optimal model for using past gait speed to improve current gait speed assessment has not been well described. This study aims to (i) evaluate the association between past gait speed information and mortality in relation to current gait speed, and (ii) determine how to best model and incorporate past gait speed in predicting mortality via both qualitative (i.e., directional change) and quantitative (i.e., magnitude of change) analyses.

Data Source & Cohort
For this cohort study, we used data from the rst ve yearly rounds of the United States of America National Health and Ageing Trends Study (NHATS) starting from 2011. NHATS is a nation-wide survey of data representative of Medicare bene ciaries aged 65 years and older. Study participants (or proxies if participants were unable to self-report) were interviewed in person and assessed for physical performance measures at baseline and each year of follow-up. For this study, we included 4289 community-dwelling participants with gait speed measures at year 1 (Y1) and year 2 (Y2). Participants were followed until 2015 for mortality (total follow-up of three years from Y2 to Y5). Details about the recruitment strategy and design of NHATS have been described elsewhere (13).

Measures of gait speed and covariates
Participants were assessed for gait speed (GS) each year. The faster value of two timed 3-meter walk trials was recorded. Demographic variables and chronic conditions were obtained by self-report.

Outcomes
The primary outcome was all-cause mortality, which was assessed yearly at follow-up. Interviews were administered with a proxy respondent in place of participants who died between rounds (14). Loss to follow-up for mortality ascertainment was 8.4% at the end of the 3-year follow-up.

Statistical Analysis
We used a series of Cox proportional hazards regression models to estimate the hazard ratios (HR) of mortality by gait speed, as summarized in Table 2. Follow-up started at the year 2 assessment of NHATS. We conducted analyses in four stages. First, the association between gait speed at year 2 with mortality was investigated with gait speed as a continuous variable exposure (Model 1). Nonlinearity was tested using quadratic and cubic terms. Second, the exposure was change in gait speed between year 1 and year 2, modelled as a continuous change in gait speed (GS Y2 -GS Y1 ; Model 2). Third, models including baseline gait speed (either at year 1 or 2) and the change between year 1 and year 2 models were examined. We investigated three models at this stage: gait speed Y2 and change in gait speed (Model 3); gait speed Y1 with change in gait speed (Model 4); gait speed trajectory from Y1 to Y2, using three categories at Y1 and at Y2 (from fast, moderate, slow at Y1 to fast, moderate, slow at Y2; total of 9 categories; the cut-offs for trajectory analysis grouping was determined by sub-grouping gait speed into tertiles as follows: slow 0.00-0.62 m/s for Y1 and 0.00-0.66 m/s for Y2, medium 0.63-0.83 m/s for Y1, 0.67-0.88 m/s for Y2, fast 0.84-2.00 m/s for Y1, 0.89-1.81 m/s for Y2; Model 5). Models 1 through 5 were adjusted for age and sex. Finally, the fourth stage examined the independent effects of gait speed Y1 and change in gait speed by including age, sex, race, and comorbidities as adjustment covariates (Model 6). We examined model t and parsimony using Akaike Information Criterion (AIC). A two-sided p-value < 0.05 was considered statistically signi cant. Statistical analyses were performed using R 3.6.1 (R Foundation for Statistical Computing).

Ethics Approval and Consent to Participate
The secondary data used in this manuscript was acquired through a public database. Ethics approval for the primary study was obtained by the Johns Hopkins University, which was approved by John Hopkins Bloomberg School of Public Health Institutional Review Board. This project was reviewed and found exempt under 45 CFR 46 (not human subject research) by the Institutional Review Board of Columbia University Medical Center.

Results
Among the 4289 participants, 2400 (56%) were females and the median age category was 75-79 years. The median age for the fast group was 70-74 years and increased in the moderate and slow groups. The slow gait speed group contained a greater relative representation of Blacks than the other gait speed groups. All nine chronic health conditions were more prevalent with slow gait. Gait speed, demographic characteristics, and chronic conditions of the cohort are summarized in Table 1 by tertiles of gait speed at year 2. Overall, the mean gait speed at year 1 was 0.72 m/s (SD, 0.25), and 0.77 m/s (0.26) at year 2. Gait speed increased slightly by a mean of 0.04 m/s (0.20) from year 1 to year 2. Over the three-year of followup, 482 (11.2%) participants died. Results and model ts for all models are reported in Table 3.  Gait speed trajectories (trichotomized at year 1 and year 2; Model 5; Figure 1) show that remaining in the fast groups from year 1 to year 2 is most protective (reference). Conversely, remaining in the slow group at both years 1 and 2 was associated with the greatest HR of mortality (HR, 4.43 [3.03, 6.48]). Model 5 also shows that any decline in gait speed systematically worsens prognosis compared to staying in the same category. The greater the decline, the worse the prognosis, as seen in the drop from fast to moderate (HR, 1

Comparison of model ts
The adjusted model had the best model t as measured by AIC (Model 6). Among the unadjusted, continuous models, those including gait speed and change in gait speed (AIC 7337) showed better ts than models considering only one of the two. The gait speed-only model had a superior t than the change-only model (AIC = 7341, 7563 respectively).

Discussion
This study aimed to investigate the use of past gait speed as a predictor of mortality in communitydwelling, older adults. Our results show that change in gait speed over time, beyond current gait speed, is an independent predictive measure of mortality among this cohort. Although current gait speed had a stronger statistical association with mortality, changes in gait speed in relation to baseline was also found to be a signi cant predictor of mortality. Best model ts were achieved by including gait speed alongside changes in gait speed.
The use of past gait speed information adds to baseline values by helping trend patient information related to overall health status. Routine use of a single cut-off value for absolute gait speed would miss fast walking individuals with a downward trending gait speed, and inaccurately prognosticate slower walkers who are currently improving. Further, as physical function and frailty vary greatly amongst older adults, a single, reliable cut-off value for absolute gait speed for clinicians to measure against all older adults is an obstacle in a population whom personalized care is an important therapeutic strategy (30). Thus, clinicians caring for older adults should gather, retain, and compare objective measures of gait speed and their differences over time to maximize the clinical utility of gait speed (31). This would allow for patients to act as their own "controls" and allow for improved interpretation of the prognostic information provided by measures of gait speed. Beyond trending data longitudinally, our categorical trajectory analysis showed that patients whose gait speed improved from a lower baseline retained a greater risk of mortality than those whose speed they matched or even surpassed. This suggests that individuals who recover their gait speed after an acute event, such as illness or hospitalization, fare better than those who do not. But ultimately those that do not experience any adverse events and maintain baseline gait speed fare best of all. (26, [32][33][34][35]. With this, our ndings support efforts to rehabilitate walking speed after acute, decompensatory events (26). Together, we recommend the routine measurement of gait speed among older adults as this can provide healthcare professionals with additional data when caring for this vulnerable patient population (4,19,28,36).
Importantly, the direction of time impacts the interpretation of changes in gait speed: prospective improvement in gait speed (results from Model 4) is associated with greater survival, while retrospective improvement in gait speed is associated with greater mortality (results from Model 3). From this, we nd that the time point at which forward change in gait speed is conditioned on alters outcomes. Model 3 evaluates the effect of change in gait speed between participants with a similar gait speed at year 2.
Here, improvements are "harmful" since greater improvements equate to a greater change, and thus a slower year 1 gait speed. This effect is similarly seen in our categorical Model 5, which shows that for a faster walker at year 2, slower gait speed at year 1 is associated with increased mortality. Conversely, Model 4 suggests that increases in gait speed amongst individuals with a similar year 1 gait speed is associated with improved survival. Again, our categorical model similarly shows the protective effect of antegrade improvement in those with slower gait speed at year 1. The differences in HR between models 3 and 4 highlights that when measuring longitudinal gait speed, the direction of time, in addition to the magnitude of change, is critical for data interpretation.

Limitations
Our ndings must be considered in light of their limitations. Firstly, our study is subject to selection bias as our sample included only those who could complete the gait speed assessment at both years 1 and 2. These conditions may have caused an over-exclusion of older adults who were at higher risk of death, whether from diagnosed or undiagnosed comorbidities, including those with neurologic disease (i.e. movement disorders), as this information is not available (3). Secondly, the oor-effect could have applied to individuals with low gait speed at year 1. These participants were limited to a lower maximum decrease in gait speed than participants with higher gait speed. Together, selection bias and oor effect may have dampened the association between change in gait speed and mortality. Thirdly, our adjusted analysis used baseline covariates, whereas our analysis included dynamic variables, namely gait speed and comorbidities. Time-dependent analyses may have better accounted for this. Finally, a self-reported data limitation were subject to information bias (37).

Conclusion
Our results demonstrate that although absolute gait speed remains a more robust predictor of survival in older adults, the integration of past gait speed information and changes in gait speed over time have useful clinical value and survival prognostication. Improved gait speed over time is associated with decreased mortality in older adults, but elevated past gait speed is independently most protective. When modelling past gait speed information, the direction of time has important implications on the correct interpretation of results.

Declarations
Ethics approval and consent to participate The secondary data used in this manuscript was acquired through a public database. Ethics approval for the primary study was obtained by the Johns Hopkins University, which was approved by John Hopkins 37. Okura Y UL, Mahoney DW, Jacobsen SJ, Rodeheffer RJ. Agreement between self-report questionnaires and medical record data was substantial for diabetes, hypertension, myocardial infarction and stroke but not for heart failure. J Clin Epidemiol. 2004(57):1096-103. Figure 1 Mortality on Gait Speed Exposure Models Trajectories (3x3) in Community-Dwelling Older Adults Aged 65+