Gray matter volume covariance networks associated with dual‐task cost during walking‐while‐talking

Abstract

We studied gray matter volume covariance networks associated with normal pace walking (NPW) speed and dual‐task costs (DTCs) during walking‐while‐talking (WWT)—a mobility stress test that involves walking while reciting alternate letters of the alphabet. Using a multivariate covariance‐based analytic approach, we identified gray matter networks associated with NPW speed (mean 102.1 cm/s ±22.5 cm/s) and DTC (percent difference in gait speed between NPW and WWT, mean 25.9% ± 18.8%) in 139 older adults without dementia (M = 75.3 ± 6.1 years). The gray matter network associated with NPW was primarily composed of supplementary motor area, precuneus cortex, and the middle frontal gyrus. Greater expression of this NPW network was associated with better processing speed (trail‐making test A [r = −0.30, p = 0.005]) and executive function (trail‐making test B − A [r = −0.43, p < 0.0001]). The gray matter network associated with DTC was primarily composed of medial prefrontal, cingulate, and thalamic regions. Greater expression of this DTC network was associated with better episodic memory performance on the free and cued selective reminding test (r = 0.30, p = 0.007). These results suggest that NPW speed and DTC are supported by different networks, and are associated with different cognitive domains.

1. INTRODUCTION

Walking in a straight line across an empty pathway is fairly simple—but everyday walking is often more complex. In fact, it often requires simultaneous performance of two or more motor and cognitive tasks such as walking while carrying on a conversation with a friend and making real‐time adjustments for sudden changes in the environment. Age‐related decline in gait performance is particularly pronounced during divided‐attention or dual‐task walking (Li, Lindenberger, Freund, & Baltes, 2001; Lindenberger, Marsiske, & Baltes, 2000), as the simultaneous performance of a secondary cognitive task while walking may more accurately model everyday gait. Even healthy older adults who perform well on activities of daily living (ADLs) walk slower under such dual‐task walking conditions (MacAulay, Wagner, Szeles, & Milano, 2017). Like instrumental ADLs, dual‐task walking performance in healthy older adults without cognitive difficulties is linked to executive function and attention (Yogev‐Seligmann, Hausdorff, & Giladi, 2008), and performance on executive function tasks accounts for a large portion (up to 27.4%) of the variance in dual‐task gait speed (MacAulay et al., 2017). Dual‐task costs (DTCs) are exceptionally high in older adults with mild cognitive impairment and dementia (Odden et al., 2017) but they also affect relatively healthy older adults. Among these older adults without dementia, the walking‐while‐talking test (walking while reciting alternate letters of the alphabet; WWT), a mobility stress test, is a reliable predictor of falls, frailty, disability, incident dementia, and mortality (Ayers, Tow, Holtzer, & Verghese, 2014; Ceïde, Ayers, Lipton, & Verghese, 2018; Verghese et al., 2002; Verghese, Holtzer, Lipton, & Wang, 2012).

We are only beginning to uncover the functional neural substrates of dual‐task walking in older adults. Studies utilizing resting‐state fMRI and functional near‐infrared spectroscopy (fNIRS) have provided early evidence linking both distinct and shared functional activation and connectivity to single and dual‐task walking performance. In a resting‐state fMRI study of older adults without dementia, for example, both WWT and normal pace walking (NPW) speed were associated with functional connectivity in sensorimotor, visual, vestibular, and left fronto‐parietal resting‐state networks. Gait speed in the WWT condition was additionally associated with greater connectivity in the supplementary motor and prefrontal components of the left fronto‐parietal network (Yuan, Blumen, Verghese, & Holtzer, 2015). In two other studies of older adults without dementia, fNIRS technology was used to show increased activation of the prefrontal cortex (PFC) during dual‐task walking relative to NPW (Doi et al., 2013; Holtzer et al., 2015). Though functional studies further our understanding of the vasodynamic response during NPW and dual‐task gait, structural analyses are needed for investigating the neural substrates supporting gait across different regions of the brain.

The structural neural substrates of gait are frequently studied through total and focal gray matter volumes (GMVs) and white matter (WM) hyperintensities (Holtzer, Epstein, Mahoney, Izzetoglu, & Blumen, 2014; Taki et al., 2011). NPW speed has been linked to GMV in cortical and subcortical regions by multiple studies (Blumen et al., 2018; Callisaya et al., 2013; Callisaya, Beare, Phan, Chen, & Srikanth, 2014; Dumurgier et al., 2012; Nadkarni et al., 2014; Rosano, Aizenstein, Studenski, & Newman, 2007; Rosano et al., 2012), although results are sometimes conflicting. While several of them report significant associations between gait speed and GMV in single brain structures (Callisaya et al., 2013; Dumurgier et al., 2012; Nadkarni et al., 2014; Rosano, Studenski, et al., 2012), there is also evidence to suggest that a more distributed network supports NPW (Blumen et al., 2018; Callisaya et al., 2014). Among studies focusing on single regions‐of‐interest, lower GMV in the PFC (Rosano, Studenski, et al., 2012) and cerebellum (Nadkarni et al., 2014) was associated with slow NPW speed in two nondemented, elderly cohorts. In a separate, but similar cohort composed of nondemented older adults, gait speed was associated with GMV in the caudate nucleus (Dumurgier et al., 2012). In contrast, results from a nondemented Australian cohort of community‐dwelling older adults suggested that gait speed was associated with a distributed GMV network, which included cerebellar, occipital, basal ganglia, parietal, temporal, frontal and prefrontal regions (Callisaya et al., 2013). Another study from our group reported on GMV patterns associated with gait speed in three separate cohorts (Blumen et al., 2018). Here, our group reported that GMV networks associated with NPW gait speed in two cohorts, both composed of community‐dwelling older adults without dementia, were topographically similar and correlated with each other, but not with the third cohort, which was composed of nondemented adults drawn from a memory clinic. While all three patterns were distinct, they shared key structural features, including precuneus, fusiform, motor, supplementary motor, and PFC regions. In addition, the extent to which community‐dwelling older adults displayed these patterns was associated with processing speed in all cohorts, and executive function in one of the cohorts. Methodological approaches likely inform the discrepancies in these results. Studies linking gait speed with single ROIs (Rosano et al., 2012; Rosano, Studenski, et al., 2012) analyzed associations between gait speed and GMV calculated in particular brain regions. In contrast, studies reporting distributed patterns (Blumen et al., 2018; Callisaya et al., 2014) correlated GMV and gait across the whole brain using either voxel‐by‐voxel (Callisaya et al., 2014) or multivariate covariance‐based analyses (Blumen et al., 2018).

Unlike with NPW speed, the structural correlates of DTC and dual‐task performance are not well understood. The pairing of a cognitive or attentional task with walking is predicated upon the well‐studied notion that increasing age is associated with worse dual‐task gait performance (Li et al., 2001; Lindenberger et al., 2000). This relationship is mediated by older adults compensating for declining postural control by applying more attention resources (Brown, Shumway‐Cook, & Woollacott, 1999; Chen et al., 1996; Teasdale, Bard, LaRue, & Fleury, 1993). So far, one study has looked at the structural correlates of walking while performing an interference task in older adults without dementia (Allali et al., 2018). Here, the GMV in the left middle frontal gyrus was significantly associated with gait speed during walking while counting backwards. Furthermore, this ROI covaried with a larger cluster involving various parts of the prefrontal and medial frontal gyrus. We are interested in the GMV covariance patterns associated with DTCs during walking while reciting alternate letters of the alphabet (WWT). In our study, we utilized a multivariate covariance‐based approach to study GMV networks correlated with DTC in older adults without dementia. We studied the GMV covariance pattern associated with this compensation because its efficiency provides new insight into the preclinical decline occurring in older adults.

We utilized a multivariate covariance‐based statistical approach, which allowed us to circumvent the multiple comparisons problem of traditional univariate neuroimaging methods (Ashby, 2011; Habeck et al., 2005; Habeck, Stern, & Alzheimer'’s Disease Neuroimaging Initiative, 2010). Between subject variability and collinearity are important considerations when assessing neuroimaging data from older adults, and our multivariate approach is resistant to these factors at the same time remaining sensitive to effects (Ashby, 2011; Habeck et al., 2005, 2010). With this approach, we separately identified GMV covariance patterns or “networks” associated with NPW speed and DTC. We hypothesized a broadly distributed GM network for NPW speed (similar to that reported in previously mentioned studies (Blumen et al., 2018; Callisaya et al., 2014)), and we expected that the GM network associated with DTC to show greater involvement of frontal regions. Additionally, we tested the association between expression of GMV covariance networks associated with single and dual‐task gait performance and certain cognitive functions, including executive function, attention, and episodic memory.

2. MATERIALS AND METHODS

2.1. Participants

We examined GMV covariance patterns linked to normal walking and DTC during a walking‐while‐talking task in a sample (n = 139 older adults; mean age = 75.3 years) of nondemented, community‐dwelling older adults identified from the Central Control of Mobility in Aging Study (CCMA) based in Westchester county in the United States (see previous descriptions for additional details; Blumen, Holtzer, Brown, Gazes, & Verghese, 2014; Holtzer, Mahoney, & Verghese, 2014). Briefly, exclusion criteria were: “inability to speak English, inability to ambulate independently, dementia, significant loss of vision and/or hearing, current or history of neurological or psychiatric disorders, recent or anticipated medical procedures that may affect mobility, and receiving hemodialysis.”

2.2. Gait tasks and measurement

Gait speed (cm/s) was measured quantitatively using an instrumented walkway (GAITRite System® Clifton, NJ) over a fixed distance (609.60 cm/20 ft). Participants walked for one trial per task condition: (a) normal pace walk (NPW) and (b) walking while reciting alternate letters of the English alphabet (WWT). During NPW, participants were asked to walk at their usual pace over the instrumented walkway (GAITRite) wearing comfortable footwear and without any attached monitors. The GAITRite system is a widely used instrumented walkway with demonstrated reliability (Kressig, Herrmann, Grandjean, Michel, & Beauchet, 2008; Menz, Latt, Tiedemann, Mun San Kwan, & Lord, 2004; Verghese, Holtzer, Lipton, & Wang, 2009). During WWT, participants walked on the computerized walkway while reciting alternate letters of the English alphabet aloud for one trial. Prior to the single trial, participants were instructed to direct equal attention to walking and talking and to avoid task prioritization (Verghese et al., 2007). Learning effects were reduced by assigning participants randomly to start with either “A” or “B” for each WWT trial (Verghese et al., 2007). WWT was utilized as the interference task given its clinical accessibility and retest reliability (Montero‐Odasso et al., 2009). We tested DTC rather the WWT gait speed because DTC models the efficiency of gait performance with the addition of an interference task. We derived DTC from the percent difference in gait speed between the NPW and WWT conditions, rather than from the raw difference; this approach has been widely used in other studies (Hall & Heusel‐Gillig, 2010; Howell, Buckley, Lynall, & Meehan, 2018; Lindenberger et al., 2000; Venema, Hansen, High, Goetsch, & Siu, 2018). Furthermore, a 10 cm/s decrease in gait speed is, perhaps, not as noticeable for a participant walking >100 cm/s, but is meaningful for a participant walking <100 cm/s, therefore treating them equally (as with raw difference) is likely incorrect. Additionally, the variance of NPW speed (our baseline measure) is lower than that of WWT gait speed (504 vs. 598), and this also supports using percent difference over the raw difference.

2.3. MRI data acquisition

Images were acquired at the Gruss Magnetic Resonance Research Center at Albert Einstein College of Medicine (Bronx, NY) with a Philips 3T MRI scanner (Achieva Quasar TX; Philips Medical Systems, Best, the Netherlands). Standard three‐dimensional T1‐weighted images were analyzed (TR/TE of 9.9/4.6 ms, 240 mm² field of view (size of the displayed image), 240 × 240 × 220 matrix and 1 mm voxel size; for additional acquisition details see (Blumen et al., 2014).

2.4. MRI preprocessing

T1‐weighted images were first manually re‐oriented to the anterior commissure—posterior commissure line and then preprocessed in the same manner using SPM12 (Wellcome Department of Cognitive Neurology) implemented with MATLAB R2015a (Mathworks, Natick, MA). Each structural MRI image was analyzed using voxel‐based morphometry, using a unified segmentation procedure that included Diffeomorphic Anatomical Registration Through Exponentiated Line Algebra (DARTEL; Ashburner, 2007). First, each structural image was segmented into gray matter (GM), and WM and cerebrospinal fluid. Then DARTEL was used to determine the nonlinear transformation for warping these GM and WM maps to match each other. DARTEL enhance intersubject alignment by modeling the shape of the brain using three parameters for each voxel, and simultaneously align GM and WM to produce a study‐specific and increasingly crisp template to which the data are iteratively aligned. Finally, the enhanced GM and WM maps were generated, spatially normalized (Wright et al., 1995) into Montreal Neurologic Institute space, and smoothed with an isotropic Gaussian kernel, full‐width‐at half‐maximum = 8 mm. Only GM maps were used in the upcoming analyses.

2.5. Group‐level covariance analyses

Multivariate analyses were performed to identify GM covariance patterns or “networks” associated with NPW and DTC, separately. All analyses were adjusted by age, sex, education, and total intracranial volume. Analyses were implemented with the principal components analysis (PCA) suite, http://www.nitrc.org/projects/gcva_pca (Brickman, Habeck, Zarahn, Flynn, & Stern, 2007; Habeck et al., 2005). GM probability maps were first masked with a GM mask supplied by SPM12 to only include voxels with >20% probability of being GM. A PCA was then performed after participant means were subtracted from each voxel, in order to generate a set of PCs and their associated participant‐specific (or pattern) expression scores. Participant‐specific expression scores reflect the degree to which a participant displays a particular component or pattern. The GMV covariance patterns associated with gait speed or DTCs were then computed by regressing the participant‐specific factor scores from the best linear combination of PCs—selected using the Akaike information criteria (Burnham & Anderson, 2004)—against gait speed or DTCs. The stability of the voxels in each GM volume covariance pattern associated with gait speed were then tested using 1,000 bootstrap resamples (Efron & Tibshirani, 1993). This process led to a distribution of sample means that could then be subjected to inferential statistics. Voxels with bootstrap samples of [Z] > 1.96 and p < 0.05 were considered significant.

These group‐level covariance analyses allowed us to identify key “nodes” in the GMV covariance “networks” associated with NPW speed and DTC (Brickman et al., 2007; Habeck et al., 2005, 2010; Steffener, Brickman, Habeck, Salthouse, & Stern, 2013). Note that covariance patterns obtained from any multivariate analysis assigns positive and negative weightings (or loadings) to each voxel (or variable) included in the analysis (Habeck et al., 2008). In the current study, positively weighted regions will be interpreted as regions that have relatively more volume with increasing NPW speed and DTC, while negatively weighted regions will be interpreted as regions that have relatively less volume with increasing NPW speed and DTC. The negatively weighted network associated with DTC, therefore, has relatively lower GMV with higher DTC (higher percent‐difference between NPW and WWT speeds). It is important to note, however, that both positively and negatively weighted regions contribute to the derived GM covariance patterns that are associated with gait speed and DTC (Habeck et al., 2008; Spetsieris & Eidelberg, 2011; Steffener et al., 2013).

2.6. Localization of clusters contributing to GMV covariance patterns

Localization of clusters contributing to the GMV covariance patterns associated with gait speed were identified through mricron using a threshold of |z| > 1.96 (p < 0.05), https://www.nitrc.org/projects/mricron. The brain regions of peak voxels in clusters with |z| > 1.96 and a size >50 voxels are listed in Tables 2 and 3.

Table 2.

Significant clusters comprising the positively and negatively weighted gray matter volume covariance patterns for normal pace walking, corrected for age, sex, education, and total intracranial volume

X	Y	Z	k	z value	Most significant focus
−43	−22	19	2095	2.29	Central opercular cortex
49	−18	8	2,117	2.27	Heschl's gyrus (includes H1 and H2)
13	−57	1	1,077	2.24	Lingual gyrus
53	−8	−33	144	2.12	Inferior temporal gyrus, anterior division
−6	−74	5	1,049	2.12	Lingual gyrus
30	−76	−11	66	2.10	Occipital fusiform gyrus
47	−63	−13	50	2.07	Lateral occipital cortex, inferior division
−22	−49	−9	50	2.05	Lingual gyrus
29	−54	−9	61	2.03	Temporal occipital fusiform cortex
11	−90	−14	79	2.00	Occipital pole
14	−2	52	624	−2.22	Supplementary motor area
17	−50	49	301	−2.20	Precuneous cortex
−16	0	53	378	−2.19	Superior frontal gyrus
28	25	30	142	−2.17	Middle frontal gyrus
−16	−55	47	379	−2.15	Superior parietal lobule
46	−17	−21	129	−2.14	Inferior temporal gyrus, posterior division
−34	−66	20	98	−2.10	Lateral occipital cortex, superior division
44	−40	−37	80	−2.08	Cerebellum (VI)
−35	−63	−3	51	−2.08	Occipital fusiform gyrus

All clusters with |z| > 1.96 and k > 50 voxels are presented

Table 3.

Significant clusters comprising the positively and negatively weighted gray matter volume covariance patterns for normal pace walking, corrected for age, sex, education, and total intracranial volume

X	Y	Z	k	z value	Most significant focus
−1	−36	35	278	2.11	Cingulate gyrus, posterior division
8	39	23	338	2.10	Paracingulate gyrus
−6	28	30	744	2.07	Paracingulate gyrus
−3	−12	10	443	2.07	Thalamus
37	−36	52	208	2.05	Postcentral gyrus
5	46	12	104	2.05	Paracingulate gyrus
−68	−14	1	321	−2.18	Superior temporal gyrus, posterior division
−62	−62	0	233	−2.17	Middle temporal gyrus, temporooccipital part
−24	3	−51	514	−2.16	Temporal pole
−57	−60	42	204	−2.16	Lateral occipital cortex, superior division
37	−14	−45	690	−2.15	Temporal fusiform cortex, posterior division
−56	1	−39	724	−2.14	Inferior temporal gyrus, anterior division
−51	−24	61	409	−2.14	Postcentral gyrus
66	−9	−21	348	−2.14	Middle temporal gyrus, posterior division
−68	−38	−16	677	−2.13	Middle temporal gyrus, posterior division
−33	−75	−57	399	−2.11	Cerebellum (crus II and VIIb)
−15	42	−28	193	−2.11	Frontal pole
51	−20	62	182	−2.08	Postcentral gyrus
70	−15	4	174	−2.08	Superior temporal gyrus, posterior division
15	46	−28	149	−2.08	Frontal pole
59	−33	−30	278	−2.07	Inferior temporal gyrus, posterior division
70	−37	−8	184	−2.07	Middle temporal gyrus, posterior division
53	32	32	142	−2.07	Middle frontal gyrus
6	62	−25	87	−2.06	Frontal pole
10	64	29	66	−2.06	Frontal pole
−56	−39	−30	161	−2.05	Inferior temporal gyrus, posterior division
70	−31	11	55	−2.04	Superior temporal gyrus, posterior division
30	−72	−58	164	−2.03	Cerebellum (crus II and VIIb)
−51	−81	−11	167	−2.02	Lateral occipital cortex, inferior division
−36	−15	−43	62	−2.02	Temporal fusiform cortex, posterior division
9	−88	41	61	−2.02	Occipital pole
58	6	−36	60	−2.00	Temporal pole

All clusters with |z| > 1.96 and k > 50 voxels are presented.

2.7. Cognitive measures and covariates

Participants completed various cognitive, psychological, and mobility assessments, as well as comprehensive neuropsychological assessments and a structured neurological exam. Measures of global cognition (Repeatable Battery for the Assessment of Neuropsychological Status; RBANS; Randolph, Tierney, Mohr, & Chase, 1998), executive function (The Trail Making Test: Time to complete Part B minus Part A; TMT B − A; Reitan, 1979), processing speed (Trail Making Test: Time to completed Part A; TMT:A), and episodic memory (free recall on the Free and Cued Selective Reminding Test; FCSRT; Buschke, 1973) were also obtained during this process. FCSRT scoring is divided in to “free” and “cued,” in which participants try to recall 16 items without cues (“free”), then the remaining items with categorical clues (“cued”); after three trials their score is calculated separately for “free,” “cued,” and “Total.” Performance in the free and cued portions tends to be negatively correlated since higher score in the first section necessitates a lower score in the next. We identified medical diagnoses through structured clinical interviews. As in our previous studies, a dichotomous rating scale (presence or absence) of physician diagnosed diabetes, chronic heart failure, arthritis, hypertension, depression, stroke, Parkinson's disease, chronic obstructive lung disease, angina, and myocardial infarction was used to calculate a global health score (GHS) between 0 and 10. Mild cognitive impairment was determined based on a consensus procedure as previously described (Petersen, 2004; Petersen et al., 1999, 2009) and presence of strokes (clinical and silent) were determined via self‐report, study clinician's examination findings as well as MRI inspection.

3. RESULTS

3.1. Baseline characteristics

Demographic, gait, and MRI data were analyzed for 139 participants, drawn from a subset of the CCMA cohort with available WWT and structural MRI data (Table 1). The average age of the participants was 75.3 years (SD 6.1 years) and 49.6% participants were female. The sample was mostly Caucasian (70%), with some Black (20%) participants. The mean GHS was 1.3 (SD 1.0), signifying overall good health and few comorbidities. Arthritis (52%), Hypertension (43.9%), and Diabetes (17.3%) were the most frequently recorded GHS components in this group. Participants' global cognition was assessed using the RBANS total index (M = 92.9 ± 12.3). Attention and executive processing were assessed using Trail Making Test A (M = 46.5 s ± 24.7 s). Executive function was assessed using the difference in time taken to complete Trail Making Test B minus Test A (M = 75.5 s ± 52.1 s). Episodic memory was assessed using the Free and Cued Selective Reminding Test, specifically using the total score in the free recall task (M = 31.1 ± 7.2) to avoid the ceiling effect observed in cumulative scores. Mild cognitive impairment was identified in 11.3% of the participants (see Section 2.7). On average, NPW gait speed (mean 102.0 cm/s ± 22.5 cm/s) was 25.9% faster than WWT gait speed (mean 75.3 cm/s ± 24.5 cm/s).

Table 1.

Summary of demographics, comorbidities, and measures of motor and cognitive performance for 139 CCMA participants included in this study

Variable	CCMA participants (n = 139)
Age, years, mean (SD)	75.3 (6.1)
Female % (N)	49.6 (69)
Education years, mean (SD)	15.7 (3.3)
Total intracranial volume, liters, mean (SD)	1.34 (0.14)
Mild cognitive impairment % (N)	11.3 (14)
Diabetes % (N)	17.3 (24)
Hypertension % (N)	43.9 (61)
Myocardial infarction % (N)	0.7 (1)
Congestive heart failure % (N)	1.4 (2)
Arthritis % (N)	52 (70)
Depression % (N)	6.5 (9)
Stroke % (N)	2.9 (4)
Parkinson's disease % (N)	0 (0)
COPD % (N)	2.9 (4)
Angina % (N)	2.3 (3)
Global health score (range 1–10). median (IQR)	1 (0–2)
Medication number, median (IQR)	5 (3–8)
Falls, % (N)
Last 12 months	22.6 (31)
Ever	62.8 (86)
GDS score (range 0–15), median (IQR)	3 (2–6)
Visual acuity <1/200 (%)	0 (0)
Repeated chair stand, s, mean (SD)	13.2 (3.8)
4ADLs total score, median (IQR)	0 (0–1)
RBANS total index: 40–160, mean (SD)	92.9 (12.3)
Normal walking velocity, cm/s, mean (SD)	102.0 (22.5)
Dual‐task walking velocity, cm/s, mean (SD)	75.3 (24.5)
Dual‐task cost, % mean (SD)	25.9 (18.8)

3.2. GMV covariance pattern associated with normal pace walking gait speed

The GMV covariance pattern associated with NPW gait speed was composed of three PCs with an R ² = 0.14. The GMV covariance network associated with NPW speed and individual factor scores were derived after adjusting for age, sex, education, and total intracranial volume and included several significant clusters (|z| ≥ 1.96, p < 0.05, k ≥ 50) shown in Table 2. Greater expression of this GMV network, represented by a higher individual factor score, is associated with relatively higher GMV in the positively weighted regions; meanwhile, negatively weighted regions have relatively lower GMV with greater expression of this network. More simply, a higher factor score indicates both (a) relatively greater GMV in positively weighted regions and (b) relatively lower GMV in negatively weighted regions. The largest positively and negatively‐weighted clusters (k ≥ 100) localized in a distributive pattern (Figure 1a). The peaks of these larger positive clusters were localized in central opercular cortex (extending into rolandic operculum and superior temporal gyrus), heschl's gyrus (extending into rolandic operculum and superior temporal gyrus), lingual gyrus (extending into calcarine cortex), and the inferior temporal gyrus. Peaks of the largest negatively‐weighted regions included supplementary motor area (extending into frontal superior cortex), precuneus cortex (extending into superior parietal lobule), superior frontal gyrus (extending into supplementary motor area), middle frontal gyrus, and inferior temporal gyrus (posterior division). Pairwise correlation with Bonferroni correction (Table 4) showed that expression of the GMV pattern of NPW gait speed was associated with processing speed (TMT A: r = −0.30, p = 0.005) and executive function (TMT B − A: r = −0.43, p < 0.0001). Expression of this GMV pattern was not significantly associated with episodic memory (FCSRT Total Free: r = 0.086, p > 0.99).

Figure 1.

Gray matter volume covariance patterns associated with normal pace walking gait speed (a) and dual‐task cost during walking‐while‐talking (b) were adjusted for age, sex, education years, and total intracranial volume; analyses were completed using T1‐weighted MRI for 139 participants. Positively (red; z value ≥1.96) and negatively (blue; z value ≤−1.96) weighted areas correspond with relatively higher and lower GMV as a function of the indicated measures of interest

Table 4.

Pairwise correlations of normal pace walking gait speed and dual‐task cost with cognitive measures of processing speed (trail making test A), executive function (trail making test B − A), and episodic memory (free and cued selective reminding test, FCSRT)

Test	Normal‐pace walking speed	Dual‐task cost
Trail making test A	r = −0.30, p = 0.005	r = 0.17, p = 0.76
Trail making test B − A	r = −0.43, p < 0.0001	r = 0.08, p > 0.99
Free and cued selective reminding test	r = 0.086, p > 0.99	r = 0.30, p = 0.007

All results presented with Bonferroni correction.

3.3. GMV covariance pattern associated with DTC

The GMV covariance pattern associated with DTC was composed of three PCs and had an R ² of 0.05. The GMV covariance network associated with DTC and individual factor scores were derived after adjusting for age, sex, education, and total intracranial volume (Table 3). Increased DTC is associated with greater expression of this network. Additionally, greater expression of the network associated with DTC, represented by a higher individual factor score, is associated with relatively higher GMV in the positively weighted regions, while negatively weighted regions have relatively lower GMV with greater expression of this network. A participant with a higher factor score would therefore demonstrate relatively higher GMV in positively weighted regions and relatively lower GMV in negatively weighted regions. The largest positively and negatively‐weighted clusters (≥100 voxels) localized in a distributive pattern (Figure 1b). The peaks of these larger positive clusters localized in the cingulate gyrus, paracingulate gyrus (extending into middle frontal gyrus; medial part), and postcentral gyrus. Peaks of the largest negatively‐weighted clusters localized in the temporal gyrus (superior, middle, and inferior), frontal and temporal poles, lateral occipital cortex, temporal fusiform cortex, postcentral gyrus, and middle frontal gyrus. Pairwise correlation adjusted for multiple comparisons (Table 4) showed significant association between expression of the GMV pattern of DTC and episodic memory (FCSRT total free: r = 0.30, p = 0.007). This GMV pattern was not significantly associated with measures for processing speed (TMT A: r = 0.17, p = 0.76) and executive function (TMT B − A: r = 0.08, p > 0.99).

4. DISCUSSION

Our study examined GM networks associated with NPW and DTC in a sample of community‐dwelling older adults with few comorbidities. GM networks associated with both conditions were composed of distinct brain structures and regions after adjusting for confounders, including age, sex, education years, and total intracranial volume. The extent to which older adults expressed the GM networks associated with NPW speed correlated with executive function and processing speed. The expression of the GM network linked to DTC, however, was associated with episodic memory performance.

4.1. GMV covariance pattern associated with single‐task gait speed and dual‐task cost

Structures comprising the GM networks associated with NPW speed in older adults without dementia reported in this study are consistent with aspects of existing literature (Allali et al., 2018; Callisaya et al., 2014; Doi et al., 2017; Ezzati, Katz, Lipton, Lipton, & Verghese, 2015; Nadkarni et al., 2014; Rosano et al., 2007). Broadly, these similarities indicate that single‐task gait speed is associated with GM networks composed of structures in both the motor and the control pathways of locomotion (Leisman, Moustafa, & Shafir, 2016). Structures and connections in the motoric pathway primarily support gait initiation and maintenance. This pathway is representative of the automaticity of gait; structures in this pathway draw from brain stem and relay with motor cortex and dorsal basal ganglia. The control pathway directs motor planning; it localizes in the supplementary motor and prefrontal cortices, relaying with the brain stem and cerebellum through the ventral basal ganglia.

DTCs were modeled as the percent difference in gait speed between the single task and dual‐task (walking‐while‐talking) conditions. The GMV covariance pattern with relatively higher volume with increasing DTC included significant clusters in the PFC, cingulate, and paracingulate regions. These areas have an established role in directing executive function and attention (Yogev‐Seligmann et al., 2008) and are involved in the control pathway of locomotion (Leisman et al., 2016). This association may represent a shift from the motor to the control pathway during dual‐tasking, especially with the increased need for motor planning and coordination during WWT. The negatively weighted GMV covariance pattern, which has relatively lower GMV with increasing DTC, included regions in a distributive pattern across the brain. These areas have relatively lower volume with increasing DTC, which may correspond with global loss of GMV correlating with reduced dual‐tasking ability. Another interpretation for these findings would be that overall cortical surface area or thickness is associated with performance in the WWT condition. Many of the clusters in both GMV covariance patterns colocalized in structures previously associated with healthy dual‐tasking in other functional and region‐specific structural MRI studies (Blumen et al., 2014; Doi et al., 2013; Hartley, Jonides, & Sylvester, 2011; Lin & Lin, 2016; Yuan et al., 2015).

One reason for increasing DTC with aging is age‐related atrophy in the brain, which follows a variable course. In addition to interindividual variability in the overall patterns of GMV atrophy, population‐wide studies have shown that some areas, such as the PFC, atrophy more than others (Doi et al., 2017; Dumurgier et al., 2012; Storsve et al., 2014). One other study (Allali et al., 2018) has looked at the GMV networks associated with walking speed during dual‐task performance. In that study, prefrontal regions supported gait speed in every walking condition, including a divided‐attention task (NPW, fast‐paced walking, and walking while counting back from 50). Similarly, our results point to the importance of the PFC in the network associated with DTC, which may be explained by the PFC's role in executive function and attention (Elderkin‐Thompson, Ballmaier, Hellemann, Pham, & Kumar, 2008; Raz et al., 1997; West, 1996). Among older adults, performing complex tasks relies upon increasing levels of brain activation; this is also true of dual‐task performance and walking‐while‐talking. Atrophy and loss of GMV in structures and pathways supporting dual‐task performance imposes a ceiling upon resources that can be activated toward performing complex tasks. This hinders effective resource recruitment, manifesting, in part, as increasing DTC. We can understand our findings in the context of this variable atrophy, consequently impaired neural compensation, and its relationship with alternate pathways supporting dual‐task performance.

In healthy, young adults the PFC supports dual‐task performance through its established role in coordinating attention resources to competing task demands (Stelzel, Brandt, & Schubert, 2009). Among older adults, however, functional studies have shown evidence both for and against a PFC‐mediated pathway for supporting dual‐task performance (Beurskens, Helmich, Rein, & Bock, 2014; Holtzer et al., 2011, 2015). The positively weighted GMV covariance network for DTC represents a fairly focal, albeit significant pattern. It implies that relatively higher GMV in paracingulate (including the medial PFC), cingulate, thalamus, and postcentral regions is associated with worse dual‐task performance. This shows the increasing role of the control pathway of locomotion with worsening dual‐task performance. The increased volume of these areas with worse overall dual task performance points to the importance of coordinating attention resources when limits to neural compensation are reached. Structural atrophy of pre‐existing pathways supporting dual‐task performance leads to these “control” structures assuming greater responsibility in directing alternate attention resources for the execution of increasingly demanding tasks. This may help to explain the wide distribution of negatively weighted areas in the GMV covariance pattern associated with DTC. Here, many additional pathways may be competing to support the execution of complex tasks. Fittingly, relatively higher GMV in a fairly distributive pattern is associated with lower DTC or better dual task performance. From this pattern, it is likely that common brain structures and pathways sub‐serve complex motor and cognitive functions. Consequently, we considered the correlation between performance in neurocognitive testing and expression of GMV covariance patterns associated with single‐task gait speed and DTC.

4.2. Association of pattern expression with cognitive domains

Greater expression of the GMV pattern associated with single‐task gait speed was associated with faster processing speed and better executive function, but not with episodic memory. This highlights the shared neural substrates supporting ambulation, processing speed, and executive function and is consistent with previous literature (Allali et al., 2018; Blumen et al., 2018; Yogev‐Seligmann et al., 2008). Normal and pathological aging affect these structures and consequently impact performance of shared functions (Blumen et al., 2018; Rosano, Studenski, et al., 2012). This association is also not surprising, as multiple studies have linked poor performance in tests measuring executive function and processing speed to decreased gait speed and increased falls (Atkinson et al., 2010; Hajjar et al., 2009; Herman, Mirelman, Giladi, Schweiger, & Hausdorff, 2010; Holtzer, Verghese, Xue, & Lipton, 2006; Inzitari et al., 2007; Liu‐Ambrose, Katarynych, Ashe, Nagamatsu, & Chun, 2009). Conflictingly, one study reported that single‐task gait speed itself is associated with executive function, processing speed, and episodic memory (Verghese, Mahoney, Ambrose, Wang, & Holtzer, 2010). However, our study shows that greater expression of the pattern associated with single‐task gait speed is only associated with executive function and processing speed. Meanwhile, greater expression of the GMV pattern associated with DTC had significant association with better episodic memory, but no significant association with processing speed and executive function. This implies that neural pathways sub‐serving dual‐task performance are also involved in episodic memory. Indeed, cognitive and motor functions are evolutionarily related, as attention and working memory evolved largely to support more complex motor control (Carruthers, 2013; Kentros, 2006). The medial PFC (MPFC), which was included in the positively‐weighted GM network associated with DTC, plays an important role in the encoding and retrieval of remote (Bontempi, Laurent‐Demir, Destrade, & Jaffard, 1999; Frankland, Bontempi, Talton, Kaczmarek, & Silva, 2004; Takashima et al., 2006; items learned weeks earlier), recent (Corcoran & Quirk, 2007; items learned 1–2 days earlier), and short‐term (Corcoran & Quirk, 2007) verbal memory for a variety of tasks. Furthermore, this finding is consistent with significant evidence (Goldman‐Rakic, 2011) that MPFC coordinates working memory processes; this role is further exhibited by greater left prefrontal activation with increasing working memory demands in verbal tasks (Bunge, 2001; Smith & Jonides, 1998; Thompson‐Schill, D'Esposito, Aguirre, & Farah, 1997). Deeper brain structures also play important roles in the GM networks associated with DTC in our study. Sub‐cortical nuclei controlling motor function, like the basal ganglia and cerebellum, are reciprocally interconnected with motor, premotor, and prefrontal areas (Leisman & Melillo, 2013; Middleton & Strick, 2000; Schmahmann, 2019). From these shared functions and interconnected pathways, it follows that increased expression of the GMV covariance pattern associated with performance in tests of complex motor function would be associated with cognitive domains that are evolutionarily linked to motor control.

4.3. Strengths and limitations

The cross‐sectional study design prevents causal inferences. Our study sample consisted of 139 community dwelling older adults without dementia. As such, it is generalizable to the majority of community dwelling older adults, but potentially not applicable to older adults with significant cognitive and motoric dysfunction. The “relatively higher” and “relatively lower” GMV observed in our cross‐sectional analyses may be influenced by the regions of the brain that undergo structural changes differently relative to others. Though rates of cortical thinning are much more variable, the overall rates of GMV loss are also different in certain structures. While evidence from cross‐sectional and longitudinal studies is sometimes conflicting (Fjell et al., 2014; Raz et al., 1997; Raz, Ghisletta, Rodrigue, Kennedy, & Lindenberger, 2010; Storsve et al., 2014), accelerating changes occur bilaterally with increasing age in certain areas. In relation to our study, accelerating atrophy in the occipital (Fjell et al., 2014), temporal (Fjell et al., 2009, 2014), entorhinal, and prefrontal (Storsve et al., 2014) cortices may have contributed to our results. Variability in interindividual brain atrophy and vascular risk factors specific to our subject population are possible other explanations.

Other studies have used different dual tasks such as errors made when counting backwards, calculation errors while reciting serial 7 subtractions from 100, or diminished speed or errors made while reading a passage (Beurskens et al., 2014). Most such studies involved older adults experiencing cognitive decline, memory impairments or older adults with diagnosed MCI. Comparatively, our population is cognitively and physically healthy. While declining gait speed is a much more sensitive test in our studied population, we could have also modeled this difference with variables other than gait speed, such as step duration stride length, double support time, etc.

Previous studies of the brain substrates of gait speed and other gait measures often show lateralized structural and functional patterns (Ezzati et al., 2015; Rosano et al., 2007, 2008; Rosano, Bennett, et al., 2012; Sakurai, Bartha, & Montero‐Odasso, 2018) that are not as clear in our study. Univariate methods are valid for studying structural correlates, but they are restricted by the problem of repeated measures. Voxel‐wise analysis employed in univariate methods require stringent, conservative corrections (Habeck et al., 2010) and may be underpowered to detect bilateral patterns of GMV change. Our multivariate analysis technique evaluates correlation and covariance in GMV, which better reflects functional and structural connectivity (Habeck et al., 2005, 2008). Additionally, our approach is more strongly powered to detect bilateral patterns of GMV change associated with our measures of interest in a nondemented and cognitively healthy older population, in which interindividual variability and collinearity are major obstacles for other methods.

4.4. GMV networks and potential confounders

In a set of follow‐up analyses, we further tested the associations between expression of derived GMV networks and mild cognitive impairment, vascular comorbidities, and arthritis. MCI and vascular risk factors raised concern for pre‐existing lesions among our group of older adults without dementia, and arthritis poses a potentially non‐neurological etiology for poor performance in either gait task. Out of 139 participants, the fourteen (11.3%) older adults with MCI exhibited slower gait speed (p = 0.0001) and decreased expression of the network supporting NPW speed (p = 0.0035). MCI status was not associated with DTC or the DTC GMV network. While hypertension and diabetes are both vascular comorbidities and risk factors for a wide range of neurological disorders (Gorelick et al., 2011), we did not find either to have significant associations with expression of GMV networks for NPW speed and DTC. Arthritis was significantly associated with slower NPW speed (p = 0.025), but not with expression of the NPW speed GMV network, the DTC network, or the DTC. Taken together these follow‐up analyses suggest that while NPW and the GM network associated with NPW differ as a function of MCI and arthritis, DTC and the GM networks associated with DTC do not.

5. CONCLUSION

Our study determined the pattern of GMV covariance associated with DTC in a sample of cognitively healthy older adults. The pattern showed that with worsening dual task performance, there was relatively higher GMV in PFC, cingulum, and paracingulum and relatively lower volume in a bilateral, widely distributed pattern involving various regions of the brain. As this is a cross‐sectional study, it is unclear whether higher DTC is associated with higher GMV in PFC, or that worsening DTC leads to increased importance of the PFC to continue supporting healthy dual task performance via recruitment of additional attention and episodic memory resources. Our results provide structural evidence for the importance of the PFC, cingulum and paracingulum in the performance of Walking‐while‐talking, a mobility stress test, in cognitively‐healthy older adults. Additionally, our GMV covariance pattern draws attention to the stark differences in brain resources utilized toward the completion of single versus dual‐tasks. Structural MRI evidence coupled with walking‐while‐talking performance provides a composite tool for characterizing the changing neural structure and function in cognitively healthy older adults at risk for future decline.

CONFLICT OF INTEREST

The authors have no conflict of interest to report.

Tripathi S, Verghese J, Blumen HM. Gray matter volume covariance networks associated with dual‐task cost during walking‐while‐talking. Hum Brain Mapp. 2019;40:2229–2240. 10.1002/hbm.24520