Abstract
Background:
Freezing of gait (FOG) is a debilitating feature of Parkinson’s disease (PD) for which treatments are limited. To develop neuroprotective strategies, determining whether disease progression is different in phenotypic variants of PD is essential.
Objective:
To determine if freezers have a faster decline in spatiotemporal gait parameters.
Methods:
Subjects were enrolled in a longitudinal study and assessed every 3– 6 months. Continuous gait in the levodopa ON-state was collected using a gait mat (Protokinetics). The slope of change/year in spatiotemporal gait parameters was calculated.
Results:
26 freezers, 31 non-freezers, and 25 controls completed an average of 6 visits over 28 months. Freezers had a faster decline in mean stride-length, stride-velocity, swing-%, single-support-%, and variability in single-support-% compared to non-freezers (p < 0.05). Gait decline was not correlated with initial levodopa dose, duration of levodopa therapy, change in levodopa dose or change in Montreal Cognitive Assessment scores (p > 0.25). Gait progression parameters were required to obtain 95% accuracy in categorizing freezers and non-freezers groups in a forward step-wise binary regression model. Change in mean stride-length, mean stride-width, and swing-% variability along with initial foot-length variability, mean swing-% and apathy scores were significant variables in the model.
Conclusion:
Freezers had a faster temporal decline in objectively quantified gait, and inclusion of longitudinal gait changes in a binary regression model greatly increased categorization accuracy. Levodopa dosing, cognitive decline and disease severity were not significant in our model. Early detection of this differential decline may help define freezing prone groups for testing putative treatments.
INTRODUCTION
Freezing of gait (FOG) is one of the more debilitating motor complications of Parkinson’s disease (PD) [1], manifested by the feet “sticking to the ground” during active movement. FOG most commonly occurs on gait initiation and turning [2]. FOG leads to increased instability and falls [3, 4], potential morbidity from falls [5], and development of the fear of falling [6]. FOG prevalence ranges from 7.1% in early disease (DATATOP study) [7] to 92% in a cohort with PD confirmed on autopsy [8]. Due to its episodic nature, and the fact that levodopa can also partially treat FOG, the signs may be masked until severe FOG develops, at which point interventions to treat the underlying etiology of FOG are likely too late. Symptom diaries and questionnaires [9, 10] have been used to increase detection and quantify frequency and severity of FOG, but are inherently subjective and made less reliable by concurrent cognitive impairments in PD patients.
Most studies on FOG have been cross-sectional, focusing on continuous gait (gait between freezing episodes), with FOG defined on the basis of retrospective questionnaires. Based on these studies, FOG patients have differences in stride length setting [11–14], foot strike variability [13], asymmetry in steppage [15], and turn dynamics [16], which have been attributed to deficits in visuospatial processing [17] and executive function [18].
The risk factors for developing FOG have been analyzed retrospectively using a few longitudinal cohorts. FOG risk was higher in more advanced patients with initial gait, balance, speech and bradykinetic motor symptoms in the DATATOP cohort [7]. In our autopsy confirmed cohort at the New York Brain bank, earlier onset of FOG was associated with earlier onset of postural instability, dyskinesias, sudden OFF states, hallucinations and cognitive complaints [8]. Faster progression of FOG severity was also associated with earlier onset of hallucinations [8]. In five studies that prospectively focused on convertors from a non-freezing (noFOG) to a FOG phenotype, finger tapping dysfunction [19], a subset of Unified Parkinson’s disease rating scale (UPDRS) Part 1 and 2 scores (including subjective speech, turning and getting out of bed, walking and balance, cognitive and autonomic function) [19], ON vs. OFF levodopa motor UPDRS scores [20], worse mini-BESTest scores [21], higher anxiety [22], and depression scores [20, 23], higher freezing of gait questionnaire (FOG-Q) scores [22], older age [23], lower education level [23], akinetic rigid phenotypes [23], slower gait speed (measured during a timed walk) [20], and earlier levodopa use [23] were associated with FOG. Only two studies monitored gait objectively and found no change [21], or a change on univariate (primarily in dual-task asymmetry) but not multivariate analysis [19].
Gait freezing results in a significant worsening of quality of life for patients [24]. Currently, no good alternatives to levodopa exist for FOG and although deep brain stimulation of the pedunculopontine nucleus (PPN) is potentially beneficial for gait disorders [25], it is invasive and not routinely performed. Defining objective measures of FOG that are time dependent, would provide an invaluable tool for developing and testing new symptom modifying therapies (as gait could potentially decline over the course of a study negating any potential benefits). Additionally an accurate predictive model is necessary to design neuroprotective strategies and greater knowledge regarding temporal changes in gait, motor and non-motor disease features are needed.
Given the differences in spatiotemporal gait parameters in FOG and noFOG patients, we postulated that a more rapid decline in objectively measured gait should be present in PD FOG patients and that inclusion of progressive changes would help refine predictive models of FOG in the future. To test this hypothesis, we enrolled a cohort of PD patients and controls and objectively monitored their gait repeatedly over a minimum period of 1 year, in order to quantify the slope of progression in these disease markers and use these progression markers in binary regression models to categorize PD subjects into FOG and noFOG groups.
MATERIALS AND METHODS
Standard protocol approvals, registrations, and patient consents
Patients presenting to the Movement Disorders Clinic at the University of Arkansas for Medical Sciences (UAMS) between 2014-2018 for clinical care (by T.V.) were asked to participate if they met enrollment criteria below. The study protocol was approved by the UAMS Institutional Review Board (UAMS IRB# 203234), and written informed consent was obtained from each participant before study procedures were performed. The study was conducted in accordance with the guidelines of the Declaration of Helsinki.
Study population
Subjects between the ages of 50– 90 with PD based on UK brain bank criteria [26] were enrolled. Family members of enrolled PD subjects, who did not report any known neurologic disease affecting gait (other than possibly neuropathy) were asked to participate as controls. Exclusion criteria for all subjects included >1 fall/day, Montreal Cognitive Assessment (MoCA) [27] score <10 (which was set low as FOG subjects have been reported to have lower cognitive performance on testing and we did not want to exclude a large portion of this population), and use of dopamine antagonist medications in the prior year (to exclude patients with possible drug-induced parkinsonism; however, the three antipsychotics safe in PD psychosis, i.e., low dose quetiapine <100 mg/day, clozapine or pimavanserin were allowed). Additional exclusion criteria for controls were a known diagnosis of a neurologic disease, other than stable neuropathy, that could affect gait.
Control subjects were included for comparison, as spatiotemporal gait parameters in healthy elderly subjects have been shown to change with age [28], including decrease in gait speed [29–31], increased variability in step-length and step-width [32], stride-width [33], stride-time and swing-time, stance-time and single-support-time [34] with advancing age. Control subjects were required to complete a questionnaire documenting their medical history. Three control subjects reported neurologic symptoms, 1 subject had fibromyalgia and migraines, 1 subject had neuropathy from B12 deficiency and migraines, and 1 subject migraines alone, and as these subjects were not outliers they were included in the analysis.
Freezers were classified at enrollment based on self-report of “feet sticking to the ground” (item 14 score ≥1 on Unified Parkinson’s Disease Rating scale, UPDRS) [35], which is the equivalent to a score >0 on item 3 of the FOG-Q. Additionally, subjects with visualized freezing by a movement disorders neurologist (T.V.) were also included in the FOG group. Seven of the 26 PD FOG subjects reported no clear levodopa responsiveness in their gait (and therefore by definition would be ON-state freezers) and are included in the FOG group.
Assessments were performed every 3– 6 months, after scheduled clinic visits to facilitate patient participation and retention. All subjects were taking their levodopa doses as per their regular schedule and were thus examined in the levodopa ON-state. Subjects who completed at least 12-months of follow-up with at least 3 visits were included in the analysis. Of 119 subjects enrolled by December 2018, 82 met criteria (range 3– 12 visits over 12– 40 months) for inclusion in our analysis.
Gait assessment
Subjects were instructed to walk at a “comfortable” pace, 8 lengths of a 20’×4’ pressure sensor impregnated mat (Zeno Walkway, Protokinetics, Havertown, PA). Data was collected and analyzed using PKMAS software (Protokinetics). Freezing episodes were uncommon, but when they occurred (shorter steps leading to a pause in forward momentum), that section of the footprints were excluded from calculations. As subjects performed 8 lengths, there were an adequate number of footsteps in those cases to still provide an accurate assessment of continuous gait and on average FOG subjects, due to shorter stride length, had a greater number of footsteps analyzed (data not shown).
The mean and stride-to-stride percent coefficient of variability (% CV) for steady state gait for each subject at each visit was calculated for 8 gait characteristics; stride length (SL), stride width (SW), stride velocity (SV), stride time (ST), swing phase percent (Sw%), single support percent (SS%), integrated pressure applied (IP), and foot strike (FS), using the combined left and right footprints. Foot strike was defined as the long axis of an ellipse surrounding the footprint on the gait mat by an inbuilt algorithm in the PKMAS software, and was included as we previously showed increased foot strike variability in FOG subjects in a cross-sectional analysis [13]. Integrated pressure provides a measure of the pressure applied, over the area of foot contact, for the time the foot was in contact with the gait mat and is calculated by an inbuilt algorithm in the PKMAS software. Integrated pressure was included to determine if variability in pressure applied during footsteps changed with time, especially if foot strike variability changed.
Progression slope calculations
In order to determine the impact of each individual’s course of progression in gait dysfunction, we calculated the percent change in each spatiotemporal gait parameter for each subject. To accomplish this, the value of the mean and % CV at each visit was normalized to the value for that individual’s initial visit. For example:
This normalization allowed us to focus on the comparison of the rate of change in a particular spatiotemporal parameter across subjects and groups, irrespective of their initial gait deficit. We then defined the annual percent change from each subject’s individual baseline (delta), by determining the slope of change in each parameter for each individual (delta). Sample calculations are shown in Fig. 1. Figure 1A shows one control, one FOG and one noFOG subjects’ normalized mean stride length (y-axis) plotted over time (x-axis), and their respective best fit lines from which the slope is calculated. Figure 1B shows the best fit lines for all individual subjects in the noFOG (light gray) and FOG (dark gray) groups; controls were excluded in Fig. 1B only to enhance clarity. Excel’s slope function was used to calculate the slope of these best fit lines. The group average and standard error in mean (SEM) of the slopes was then calculated to define the annual percent change in mean or % CV for the control, noFOG, and FOG groups.
Spatiotemporal gait parameters progression slope (delta) calculations. A) Normalized stride length measurements from one control, one noFOG and one FOG subject plotted over time. B) Best fit lines for normalized stride length of all FOG and noFOG subjects. The slopes of the best fit lines for all subjects in a group were averaged to define the delta bar graphs shown in subsequent figures.
Other assessments
Non-gait assessments performed at every visit included a complete UPDRS [35], a Hoehn and Yahr staging score (H&Y) [36], the Freezing of Gait Questionnaire (FOG-Q) [9], and the Montreal Cognitive Assessment (MoCA) [27], and the slope of change in these scales was calculated as described for the gait parameters above. Assessments performed at the initial visit included a Frontal Assessment Battery (FAB) [37], the Scales for Outcome in Parkinson’s disease – Cognition (SCOPA-Cog) [38], the Hamilton depression (HAM-D) [39], and anxiety (HAM-A) [40] rating scales, Apathy Evaluation Scale (AES) [41], Parkinson’s disease questionnaire for quality of life (PDQ-39) [42], Epworth Sleepiness scale [43], and REM Sleep Behavior Disorder Screening Questionnaire (RBD-Q) [44].
Medications taken by PD patients were recorded at each visit, and total daily levodopa doses at each visit was calculated based on 100% and 70% bioavailability for immediate and extended release formulations, respectively [45]. The Levodopa Equivalent daily dose (LEDD) at the initial visit was calculated based on prior reported estimations levodopa equivalency of dopamine agonist and MAO-Inhibitor dosages [46].
Statistical analysis
Statistical analysis was performed with SPSS 24 (IBM). Normality of data was assessed using the Shapiro-Wilk test. Statistical significance was individually calculated for each of the subject demographics, baseline gait and delta gait parameters was calculated using a one-way ANOVA for normally distributed data and the Kruskal-Wallis test otherwise. A post-hoc Bonferonni correction was applied for multiple group comparisons to determine significance between the subgroups (example noFOG vs. FOG) when the initial ANOVA and Kruskal-Wallis test showed significance. The parameters undergoing each specific test are indicated in the tables.
As subjects underwent research visits at the time of their regular scheduled clinic visits, not all subjects presented at the same time interval. We therefore also used a linear mixed model to fit the FOG and noFOG group data as this allowed us to include all subjects who participated in the analysis. Using SPSS, the mean and % CV of each spatiotemporal gait parameter was modeled using time as a repeated measure, and group and sex as fixed variables. Covariates included in the models were age, disease duration, motor UPDRS scores and levodopa doses. For all linear mixed models, a first order autoregressive variance pattern was assumed, with random effects for intercept, age, motor UPDRS, disease duration and levodopa dose modeled with first order variance structure. Maximum likelihood method of estimation was applied. Subject variability was included for both random and repeated effects. We conducted this analysis using 1) the actual values for the different variables, and 2) the square-root transformation of the data to account for the non-parametric nature of the data for some variables. Differences between the FOG and noFOG groups in MoCA, motor UPDRS, total UPDRS, FOG-Q scores, and differential changes in levodopa dose were also determined using the linear mixed models approach using age and disease duration as covariates with random effects as described above both for non-transformed and square-root transformed variables.
As the majority of measured and calculated features did not show normal distribution, Spearman’s Correlation coefficients for non-parametric data were calculated, to determine associations between motor and non-motor features and the delta spatiotemporal parameters, separately for FOG and noFOG groups. A forward, stepwise binary logistic regression model was applied to the FOG and noFOG groups, to determine which measured parameters (independent variables) best categorized FOG and noFOG status (dependent variable). Multicollinearity was assessed using the linear regression tool in SPSS.
Data availability statement
As this is an ongoing longitudinal study anonymized data sets can be shared at the request of a qualified investigator.
RESULTS
Demographics of the 82 subjects (26 FOG, 31 noFOG, and 25 controls) analyzed are shown in Table 1. The age of subjects at enrollment and duration of follow-up (6.3±2.2 visits over 27.6±2.2 months), was similar between the groups (Table 1). There was no significant difference in the mean motor UPDRS at enrollment between the FOG and noFOG groups, although total UPDRS scores were higher and disease duration was longer in the FOG group. There was no significant difference between PD groups on the MoCA and SCOPA-Cog, but mean FAB scores were lower in the FOG group. While mood and anxiety scores were similar in the PD groups, FOG subjects reported greater apathy (Table 1). Although RBD-Q significantly differed between PD and controls there was no difference between the PD groups (Table 1). Spatiotemporal gait parameters at the initial visit (baseline) are shown in Table 2. FS and SL variability were significantly different between the FOG and noFOG groups at initial visit but other parameters did not reach statistical significance (Table 2).
Baseline subject characteristics
+ANOVA; #chi-square, ∧Mann-Whitney, otherwise KW (values with post-Hoc Bonferroni correction if indicated).
Spatiotemporal Gait Parameters at initial visit
+ANOVA; otherwise KW with post-hoc Bonferroni correction applied for individual gait parameters.
Progression in spatiotemporal gait parameters
The progression in mean normalized spatiotemporal gait parameters in the three groups is shown in Fig. 2 using two different methods. The calculated slope of progression (described in detail in the methods section) for each normalized mean spatiotemporal gait parameter is shown on the right of each section labeled “delta”, while the average normalized mean spatiotemporal gait parameter at each time point (for subjects with data at that time point) are shown on the left of each section labeled “visits”. Controls are shown as a comparison as spatiotemporal gait parameters have been shown to change with age in otherwise healthy non-falling adults [28].
Progression of mean spatiotemporal gait parameters. The mean normalized spatiotemporal gait parameters for controls (black circles), non-freezers (white squares) and freezers (gray triangles) are plotted both as function of time (visits; left graph) and slope of progression (delta; right graph) for (A) stride length, (B) stride width, (C) stride velocity, (D) stride time, (E) swing %, (F) single support %, (G) integrated pressure, and (H) foot strike. Best fit trendlines are shown for the graphs plotting parameters as a function of time. All values are shown as mean±sem.
Utilizing slopes of each individual subjects progression (delta) FOG subjects had a faster annual percent decline in mean SL (Fig. 2A, delta; p = 0.003), SV (Fig. 2C, delta; p = 0.016), Sw% (Fig. 2E, delta; p = 0.001) and SS% (Fig. 2F, delta; p < 0.001) compared to noFOG. Using the slope method, the rate of mean FS shortening was not statistically faster than noFOG once post-hoc Bonferroni corrections were applied (Fig. 2H, delta; p = 0.037 assuming 2 groups, p = 0.110 after Bonferroni correction). The linear mixed model (Fig. 2A, visits), including age, disease duration, motor UPDRS scores and daily equivalent levodopa dose as covariates with random effects, showed significant difference of time×group for SL (Fig. 2A, visits; p = 0.002), Sw% (Fig. 2E, visits; p = 0.004), SS% (Fig. 2F, visits; p = 0.006), and FS (Fig. 2H, visits; p = 0.041). As some of these variables did not show normal distribution to ensure this was not interfering with the results, we also took the square-root transformation of the data set and used the transformed data for our linear mixed model with similar results except decline in mean SV was significantly faster applying the transformation: SL (Fig. 2A, visits; p = 0.001), SV (Fig. 2C, visits; p = 0.017), Sw% (Fig. 2E, visits; p = 0.001), SS% (Fig. 2F, visits; p = 0.001), and FS (Fig. 2H, visits; p = 0.024).
The progression in % CV spatiotemporal gait parameters is shown in Fig. 3. With the linear mixed model analysis, variability in foot strike (Fig. 3H, visits, p < 0.001, p = 0.001 with square root transformation) showed a significant time×group interaction between FOG and noFOG groups, with greater variability over time in the FOG group, but not using the slope of change method (Fig. 3H, delta). Delta SS% (Fig. 3F, delta, p = 0.041) was increasing significantly faster using individual slopes but not on linear mixed model analysis (Fig. 3F, visits). Stride time variability was increasing significantly faster using both the linear mixed models analysis (Fig. 3D, visits, p = 0.003, p = 0.001 with square root transformation) and the slope of change methods (Fig. 3D, delta, p < 0.001).
Progression of stride-to-stride variability (% CV) in spatiotemporal gait parameters. The mean normalized spatiotemporal gait parameters for controls (black circles), non-freezers (white squares) and freezers (gray triangles) are plotted both as function of time (visits; left graph) and slope of progression (delta; right graph) for (A) stride length, (B) stride width, (C) stride velocity, (D) stride time, (E) swing %, (F) single support %, (G) integrated pressure, and (H) foot strike. Best fit trendlines are shown for the graphs plotting parameters as a function of time. All values are shown as mean±sem.
Disease severity and gait progression
Spearman’s correlation coefficients were calculated to determine if there were any correlations between disease severity at enrollment, and gait decline in FOG subjects (Table 3). Only two gait parameters that were differentially affected in FOG vs noFOG subjects over time, delta mean SS% and delta % CV FS, were inversely correlated with initial motor UPDRS scores. None of the delta gait parameters were correlated with disease duration or FOG duration in the FOG group. Together this suggests that gait decline was not primarily related to disease severity at enrollment. Delta mean SL, SV, Sw% and SS% did show inverse correlation to delta motor UPDRS in the FOG group, while delta mean FS, delta % CV FS, delta CV ST and delta CV SS% did not, suggesting subjects with worsening gait over time also had some degree of overall motor decline over time (Table 4). The delta motor UPDRS however was not significantly different between the FOG and noFOG groups (Fig. 4A).
Levodopa and gait progression
In our FOG group, delta gait parameters were not correlated with initial levodopa dose or duration on levodopa (Table 3). In noFOG subjects there was an inverse correlation between initial levodopa dose and delta mean SL, SV, Sw%, SS% and a direct correlation with delta % CV SS% (Table 3). Levodopa dose changes during the study period were also not correlated with delta gait parameters in either PD group (Table 4). The delta levodopa dose (Fig 4D) was not significantly different between the noFOG and FOG groups (Fig. 4D)
Cognition and freezing of gait
As changes in cognition have been related to FOG, we studied cognitive function in our cohort in more detail. Initial MoCA, using Spearman’s correlation coefficients, was correlated only with delta mean FS, FAB was correlated with delta mean SL, SV, Sw%, SS%, and delta % CV ST, and Scopa-Cog with delta mean SL, Sw% and SS% (Table 3). There was also a faster decline in MoCA scores in the FOG vs noFOG groups using both a linear mixed model approach (Fig. 4C, visits; p = 0.019, p = 0.008 with square root transformation) including age and disease duration as covariates with random effects, and the slope of change in MoCA (Fig. 4C, delta; p = 0.018). The delta MoCA was inversely correlated with delta mean SV, Sw% and SS% in the FOG but not noFOG groups (Table 4).
Spearman’s correlation coefficients between individual baseline subject characteristics and significant delta spatiotemporal gait parameters
Legend: Spearman’s correlation coefficient (upper row), p value (lower row).
Progression of other repeatedly measured parameters. The mean parameters for controls (black circles), non-freezers (white squares) and freezers (gray triangles) are plotted both as function of time (visits; left graph) and slope of progression (delta; right graph) for (A) motor UPDRS, (B) total UPDRS, (C) Montreal Cognitive Assessment (MoCA), (D) daily equivalent levodopa dose, and (E) freezing of gait questionnaire (FOG-Q) score. All values are shown as mean±sem.
Spearman’s correlation coefficients between individual delta cognitive-motor features and significant delta spatiotemporal gait parameters
Legend: Spearman’s correlation coefficient (upper row), p value (lower row).
Factor discrimination for freezing of gait
A forward step-wise binary logistic regression analysis was performed to determine the effects of the studied parameters (independent variables) on the likelihood of having FOG vs noFOG (dependent variable), in order to develop a core feature set that best differentiated the two groups. Control subjects were not included in this analysis. In a step-wise iteration shown in Table 5, we eventually reached a model that included all variables (initial and delta gait parameters, initial and delta cognitive scales, initial sleep, mood and quality of life scales). The most accurate categorization of subjects required inclusion of gait progression parameters (delta gait parameters) along with initial mood scores. The combined final model resulted in 95.9% of subjects being correctly binned into their respective groups (96.4% noFOG, 95.2% FOG), with significant variables being delta mean SL (p = 0.004), delta mean SW (p = 0.023), delta % CV Sw% (p = 0.021), base % CV FS (p = 0.040), Apathy (p = 0.029) and base mean Sw% (p = 0.081). The sensitivity and specificity of this model were 95.2% and 95.4% respectively. Multicollinearity between these 6 variables was assessed and all combinations had VIF scores <2 suggesting independence of parameters. In an earlier iteration of the model, without initial mean Sw%, the model was only 87.8% correct (92.9% noFOG, 81% FOG). The other iterative combinations of parameters without either initial mood scores or delta gait parameters resulted in lower discriminatory power in the range of 53.8% – 81.8% (Table 5).
Iterative FOG vs noFOG categorization using a step-wise binary logistic regression model
Key: Initial gait parameters (initial mean and % CV: SL, SW, SV, ST, Sw%, SS%, FL, IP); Initial cognitive scores (MoCA, SCOPA-Cog, FAB); Initial mood scores (HAM-A, HAM-D, AES); Initial sleep scores (Epworth, RBD); Initial levodopa (total daily levodopa dose, duration on levodopa); Delta gait parameters (delta mean and % CV: SL, SW, SV, ST, Sw%, SS%, FL, IP); Delta levodopa (mean dose change per year); PPV, positive predictive value.
DISCUSSION
In this study, we report for the first time to our knowledge, a differential, more rapid decline in continuous gait in PD FOG subjects. Using longitudinally measured objective spatiotemporal parameters of gait in PD patients and controls, we show using both a linear mixed model approach and normalized slopes of gait parameter progression (which we have defined as delta variables), that the rate of decline in mean SL (– 6.9% /y), Sw% (– 3.9% /y), SS% (– 3.8% /y) and increase in % CV ST (20.6% /y) is faster in FOG compared to noFOG subjects. The rate of change calculations also showed that the rate of decline in mean SV (– 7.8% /y) and increase in % CV SS% (20.4% /y) was also faster in PD FOG while the linear mixed model suggested that mean FS (– 0.5% /y) declined and % CV FS (70% /y) increased faster in FOG subjects. When we added spatiotemporal gait progression to a binary regression model with parameter values from only the initial visit, the accuracy of the model to correctly discriminate between FOG and noFOG subjects improved from 75% to 95%, higher than any previously published model that we know of.
In a cross-sectional cohort, we previously reported that variability in foot strike was significantly different between FOG and noFOG groups [13]. In our longitudinal cohort, there was a trend towards shorter foot strike over time in the FOG group (Fig. 2H, delta; 0.6±2.0 noFOG vs. – 0.5±2.0 FOG % change/year, p = 0.032), which was not statistically significant after correcting for multiple comparisons (p = 0.11). Using a linear mixed model, mean FS (Fig. 2H, visits) was significantly shorter and % CV FS (Fig. 3H, visits) significantly higher. Initial foot strike variability was also a significant feature in our final factor model. When asked, FOG patients commonly report to us in clinic visits (T.V.) that in the OFF levodopa state, or prior to a freeze they find themselves walking more on their toes. As we also previously suggested [13], increased foot strike variability in FOG subjects means not only longer foot strikes (shuffles) but shorter contact length of the foot (that could be more on toes or heels), both of which could precede a freeze. Future studies using instrumented gait mats, which allow determination of foot strike, are needed to correlate these changes to actual freezing events.
Many groups have shown stride length is shorter in FOG subjects, and successive decline in stride length (sequence effect) predisposes to freezing episodes [12–14]. It was therefore not surprising that stride length declined at a faster rate in our FOG cohort compared to noFOG (Fig. 1A) and was a significant variable in our final factor model. Stride time variability has also been reported as a measure of impaired gait control in FOG [11], and in our study stride time variability increased faster in the FOG group compared to the noFOG group (Fig. 3D). Herman et al. [20] noted gait speed (using a timed walk) to also be a predictor of future noFOG to FOG conversion. Additionally, more rapidly decreasing Sw% and SS% along with decreased SL and SV is consistent with the more severe clinical phenotype of shorter, shuffling steps seen in FOG subjects.
A prior study by Vervoort and colleagues found no differences in gait parameters between FOG and noFOG subjects over 12 months [21]. This could be due to a number of differences between our studies. Vervoort et al. used proportional change in 2 visits over 12 months, with assessments performed in the unmedicated state, while we used multiple visits (average of 6 visits over 26 months) in the medicated state, which possibly reduced day-to-day performance variability and allowed a longer period to monitor for change. In the Vervoort study, noFOG subjects were much younger (58.6 y) than their FOG subjects (67.4 y), and were also younger than the noFOG group (65.8 y) in our study, although disease duration was similar. Given these age differences, one would expect to see an even greater difference between Vervoort’s PD cohorts as gait parameters decline with age [28].
A more recent study by D’Cruz et al. [19] looked at gait parameters annually over a two year course in noFOG subjects and found that those that converted to a FOG phenotype, on univariate analysis, had greater initial gait asymmetry in both single and dual-task gait (utilizing an auditory stroop interference task) followed by swing time variability. While the study design is different from ours (noFOG converters compared to known noFOG and FOG subjects in our study), it is interesting that rate of change of variability in swing percent was also present as a factor with the best discriminatory power in our binary regression model. We have not looked at asymmetry variables in our cohort and would be important to do so in the future. However, in the D’Cruz study [19], the gait features did not reach significance in their final multivariate model, although objective finger tapping measures were predictive of conversion; consistent with previous studies showing that noFOG subjects also show freezing on upper limb tasks [47, 48]. Addition of objective upper limb tasks into longitudinal studies is also likely important for future FOG prediction studies.
In an early PD cohort (average 6.3 months disease), tremor dominant (TD) subjects had a shorter more variable step length after 18 months, while gait parameters were not significantly different in postural instability gait disorder (PIGD) subjects (approximately 16% with FOG) [49]. In our more advanced cohort, (average 7.9 y disease), we find significant objective gait decline in FOG (85% PIGD) compared to noFOG (45% PIGD). PIGD and TD subtypes also shift over time, especially due to tremor responsiveness to dopaminergic medications [50]. Two studies exploring the progression of PIGD versus tremor subtypes of PD have also not shown differential gait decline in spatiotemporal parameters [49, 51]. This could suggest that PIGD and FOG while interrelated are two independently regulated sub-groups of patients and supports our results as FOG, not PIGD related.
While subjects in the FOG group in our study did have a longer disease course, and higher motor UPDRS scores at presentation, the motor UPDRS scores were not statistically significant between the FOG and noFOG groups. Additionally, the initial motor UPDRS scores were only correlated with delta mean Sw% and delta % CV FS and no other gait parameters in the FOG group, but also gait parameters were not correlated with disease duration or FOG duration, arguing against a general trend of gait declining faster in more severely affected FOG individuals or FOG individuals with longer disease course. The presence of freezing in and of itself portends a more severe disease and as it leads to imbalance and falls, correctly characterized FOG subjects will as a result always have higher Hoehn and Yahr scores. When analyzing our data using a linear mixed model, accounting for age, disease duration, initial motor UPDRS scores and levodopa equivalent doses, mean SL, Sw%, SS% still showed significant difference, while SV was approaching significance (p = 0.074), and SV was significant if levodopa dose was not included as a co-variate (p = 0.04). While we cannot completely exclude a component of disease severity accounting for faster gait progression in the FOG group, as delta motor UPDRS was correlated with some of the spatiotemporal gait parameters, the delta motor UPDRS was not significantly different between the noFOG and FOG groups (Fig. 4A), and neither the initial nor the delta motor UPDRS was a factor in our final discriminatory binary regression model classifying FOG subjects with 95% accuracy. Herman et al. [20] showed in their cohort that noFOG to FOG phenotype conversion over 5 years was independently predicted by motor UPDRS ON vs OFF levodopa scores at the initial visit, which taken together with our data could suggest that motor response to levodopa, and not general disease severity, may be a key component for conversion. As all our subjects underwent evaluation in the ON levodopa state, we could not further add to this hypothesis.
It has been suggested that levodopa may contribute to freezing, especially in the case of levodopa ON-state freezing [52]. Two reports used the lack of documented FOG in historical records prior to levodopa introduction to suggest a causal link [53, 54]. Phenotypic conversion to FOG was also associated with earlier levodopa use and higher daily dose [23]. While our results cannot directly speak to the issue of FOG development or conversion, in our cohort, levodopa did not appear to play a role in the more rapid gait decline in FOG subjects. The levodopa equivalent dose, duration of levodopa therapy at enrollment, and the change in levodopa dose over the study course, were not significant variables in our binary categorization of FOG vs. noFOG phenotypes with any combination of tested parameters (Table 5), nor were they correlated with declining gait parameters in FOG subjects (Tables 3 and 4). In noFOG subjects higher levodopa dose led to less decline in mean SV, Sw%, SS%, and % CV SS%, as these were inversely correlated with levodopa dose, suggesting that higher levodopa dose in noFOG subjects might lead to a slower gait decline.
Cognition has also been shown to be worse in patients with FOG, with cross-sectional studies showing greater deficits in visuospatial processing [17] and executive function [18]. Our results are somewhat mixed in this regard. There were significantly lower FAB scores in the FOG compared to the noFOG group, although MoCA scores and SCOPA-Cog scores were not significantly different. Initial cognitive scores were correlated with a number of delta gait variables; however, none of the cognitive assessments featured in the final FOG categorization model. It is possible that more thorough, domain specific, neurocognitive testing could provide a different outcome. Cognitive decline in our cohort (based on the MoCA) occurred at a faster rate in FOG subjects, similar to one previous study [55]. This decline was correlated with a decline in mean SV, Sw%, and SS% in FOG subjects. The cognitive model of freezing suggests that baseline cognitive executive dysfunction in PD FOG along with greater impairment in set-shifting tasks and conflict resolution could lead to freezing when situations require more rapid decision making [56, 57]. It is possible that the worsening cognition over time, leads to slower gait and more time with the feet on the ground (decreased Sw%) thereby increasing shuffling and possibly freezing. However, it remains unclear whether cognition drives gait deficits and freezing, or the opposite where activity level from gait dysfunction drives cognitive decline and longitudinal studies in large cohorts of noFOG subjects who convert to FOG during the study period would be needed to tease out this temporal relationship.
Anxiety [22] and depression [20, 23] have previously been suggested to predict the development of FOG in longitudinal cohorts. In our cohort, Hamilton depression and anxiety scores were not significantly different between the PD groups but apathy (AES scores) was significantly higher in the FOG group. Apathy (but not mood) scores were also a significant factor in our final FOG categorization model (Table 5). Based on our results, future studies looking for conversion from a noFOG to FOG phenotype should also include a measure of apathy.
The strengths of this work include a prospective and longitudinal objective measurement of steady state gait and cognition in PD patients and controls along with initial assessments of mood, apathy, sleep and quality of life, with 518 subject visits included in our analysis. However, to limit patient attrition, subjects were only assessed in a levodopa ON-state as in prior studies [49, 58]. While this does not provide un-medicated disease progression, it better mimics a population seen during routine clinic follow-up. Despite the medicated state, which would mask disease progression leading to our findings underestimating the rate of decline, we still see differential objective changes in gait between FOG and noFOG subjects. While our study population is too small to apply more complex machine learning modeling algorithms, it is on the order of many published cross-sectional studies on FOG. To enhance sensitivity, we only used subjects who had completed at least three visits, allowing us to calculate a slope of change for each individual, compared to other published studies that looked at the absolute change that occurred at two different time points [49]. While subjects with longer follow-up and more visits would have more accurate slope calculations leading to possible random bias, the number of visits (6.4±2.2 vs. 6.7±2.2 visits, noFOG vs. FOG) and duration of follow-up (26.1±8.5 vs. 26.9±8.0 months, noFOG vs. FOG) was similar in the noFOG and FOG groups, which should limit any such bias.
In summary we find that objectively measured gait declines faster in PD patients with FOG and inclusion of the rate of decline in gait greatly enhanced the ability of a binary logistic model to find factors that correctly identified freezers (95% correctly). Gait decline in freezers did not appear to be related to levodopa dose or duration of therapy, levodopa changes during the study, or initial disease severity. Addition of longitudinally assessed non-motor features and measures of turning dynamics, balance and upper limb function, may help further refine the model in the future. Future longitudinal assessment of spatiotemporal gait progression dynamics in subjects converting from a noFOG to a FOG phenotype, with comparison to both groups will also be needed. In the future, building upon the results from our study, multicenter collaborative studies, or databases allowing researchers to pool datasets to increase population size, are likely to be necessary to develop an accurate predictive model for freezing of gait.
CONFLICT OF INTEREST
The authors have no conflict of interest to report.
Financial disclosures
Dr. Virmani, Ms. Glover, and Ms. Pillai, received salary support from the University of Arkansas Clinician Scientist Program. Ms. Glover also received salary support from the NIGMS pilot award to Tuhin Virmani. Dr. Virmani and Ms. Doerhoff received salary support from the University of Arkansas for Medical Sciences. None of the other authors have any financial disclosures or conflicts of interest related to the research covered in this manuscript.
Footnotes
ACKNOWLEDGMENTS
This study was funded in part by the UAMS Clinician Scientist Program, NIH NIGMS (GM110702) and the Parkinson’s Foundation (PF-JFA-1935). We appreciate the mentorship of Drs. Garcia-Rill and Larson-Prior. We thank Dr. Misty Virmani for critical review of the manuscript and Dr. Arvind Virmani for guidance on statistical methods. We also greatly appreciate the commitment and dedication of our participants.