This study is embedded in the Korean Longitudinal Study on Cognitive Aging and Dementia (KLOSCAD), a population-based prospective multicenter cohort study of Koreans aged 60 years and older. The KLOSCAD was launched in 2009 and was followed up biennially until 2020 [20]. We included 207 of the 232 individuals who simultaneously completed gait evaluation and brain MRI in the KLOSCAD cohort in the final analysis after excluding the following conditions: (1) dementia or major psychiatric disorders according to the Diagnostic and Statistical Manual of Mental Disorders (4th ed., text revision) criteria; (2) major neurologic disorders including Parkinson’s disease, brain tumor, or stroke; (3) history of traumatic brain injury; (4) Tinetti Performance Oriented Mobility Assessment—Gait subscale (POMA-G) score of ≤ 10; (5) one or more cardinal signs (bradykinesia, tremor, rigidity) or two or more non-cardinal signs of parkinsonism according to the Unified Parkinson’s Disease Rating Scale Part III (UPDRS).

All participants provided written informed consent themselves or via their legal guardians. The present study was approved by the Institutional Review Board of the Seoul National University Bundang Hospital.

Geriatric psychiatrists performed a standardized diagnostic interview that included a detailed medical history, physical and neurological examinations, and laboratory tests on each subject using the Korean version of the Consortium to Establish a Registry for Alzheimer’s Disease Assessment Packet Clinical Assessment Battery (CERAD-K-C) and Mini International Neuropsychiatric Interview [21, 22]. They evaluated the comorbidity and vascular burdens using the Cumulative Illness Rating Scale (CIRS) and Modified Hachinski Ischemic Score, respectively, and determined whether degenerative arthritis of the spine and/or lower extremities was present using the musculoskeletal category of the CIRS. They evaluated Parkinsonian symptoms and gait disturbances using the UPDRS and POMA-G. The maximum UPDRS score is 108, and higher scores indicate more severe Parkinsonian motor symptoms. The maximum POMA-G score is 12 and higher scores indicate improved gait performance.

Trained neuropsychologists or research nurses performed neuropsychological assessments, including the Korean version of the Consortium to Establish a Registry for Alzheimer’s Disease Neuropsychological Assessment Battery (CERAD-K-N) [21], Korean version of the Frontal Assessment Battery [23], and Digit Span Test. The CERAD-K-N consists of nine neuropsychological tests, including the Categorical Fluency Test (CFT), Modified Boston Naming Test (mBNT), Mini Mental Status Examination (MMSE), Word List Memory Test (WLMT), Constructional Praxis Test (CPT), Word List Recall Test (WLRT), Word List Recognition Test (WLRcT), Constructional Recall Test (CRT), and Trail Making Test A. We calculated the CERAD-K total scores (CERAD-TS) by summing the CFT, mBNT, WLMT, WLRT, WLRcT, and CRT scores. We defined the Verbal Memory Score (VMS) as the weighted average of the WLMT, WLRT, and WLRcT scores. The CERAD-TS and VMS range from 0 to 100 and 0 to 30, respectively, and higher scores indicate better cognitive function.

Research nurses asked the participants to self-perform the Korean version of Geriatric Depression Scale (GDS) to evaluate the severity of depressive symptoms [24].

We measured the temporal gait variability because temporal parameters were more affected by dementia-related gait parameters than spatial parameters. Additionally, temporal parameters were associated with AD pathology; however, spatial parameters were not [25]. We used steps instead of strides to measure temporal gait variability because the gait variability from the left and right steps combined was more reliable than using strides [26].

We previously measured variations in the step time of each participant using a TAA (FITMETER® [FitLife Inc., Suwon, Korea] or ActiGraph® [SMD solution, Seoul, Korea]) placed over the center of body mass (CoM) [19]. The inertial measurement units (IMU) were hexahedrons (35 × 35 × 13 mm [14 g]/30 × 40 × 10 mm [17 g]) with smooth edges and a digital TAA (BMA255, BOSCH, Germany) and gyroscope (BMX055, BOSCH, Germany). They could measure tri-axial acceleration and velocity up to ± 8 g (with a resolution of 0.004 g/0.00024 g) and ± 1000°/s (with a resolution of 0.03°/s) at 250 Hz, respectively. We attached an IMU to each participant at the 3rd–4th lumbar vertebrae using Hypafix. We asked each participant to walk back and forth three times on a 14-m flat straight walkway at a comfortable self-selected pace and start turning after passing the 14 m line. To measure steady-state walking, we analyzed the data of the central 10 m of the 14 m-walk after the 2 m-walks prior to the start and each turn were eliminated. We calculated step time variability from the vertical acceleration data using the method described by Zijlstra and Hof [i.e., % coefficient of variation (% CV) of step time = (standard deviation of step time/mean step time) × 100] [27]. In the present study, we used the natural log transformation of %CV of step time as the gait variability as the %CV of step time was not normally distributed. The detailed methods of signal processing and gait variability calculation are described elsewhere [19].

We also measured the leg length, i.e., the distance between the anterior superior iliac spine and lateral malleolus, as a covariate because leg length is associated with spatiotemporal gait parameters.

We obtained three-dimensional structural T1-weighted spoiled gradient echo magnetic resonance (MR) images of the participants within a year after their clinical and neuropsychological assessments using a 3.0 Tesla GE SIGNA Scanner (GE Healthcare; Milwaukee, WI) in Digital Imaging and Communications in Medicine format with the following parameters: acquired voxel size, 1.0 × 0.5 × 0.5 mm3; 1.0 mm sagittal slices with no inter-slice gap; echo time, 3.68 ms; repetition time, 25.0 ms; number of excitations, 1; flip angle, 90°; field of view, 240 × 240 mm; 175 × 240 × 240 matrix in the x-, y-, and z- dimensions. We bias-corrected the T1 images to remove intensity inhomogeneity artifacts using Statistical Parametric Mapping software (version 8, SPM8; Wellcome Trust Centre for Neuroimaging, London; http://www.fil.ion.ucl. ac.uk/spm). We then resliced the bias-corrected T1 images into isotropic voxels (1.0 × 1.0 × 1.0 mm3).

We performed cortical reconstruction and volumetric segmentation using FreeSurfer v6.0 (http://surfer.nmr.mgh.harvard.edu/). We smoothed thickness maps with a 10 mm full-width half-maximum (FWHM) Gaussian kernel before performing statistical analysis. Based on gyral and sulcal anatomy, we segmented the cortex into 34 gyral regions per hemisphere (13 frontal, 9 temporal, 4 occipital, 7 parietal, and the insula), using the Desikan–Killiany Atlas [28].

To examine the association of gait variability with cognitive function measures (CERAD-TS and VMS), we created a multivariate general linear model (GLM) adjusted for age, sex, education, GDS, CIRS, leg length, and presence of arthritis using the linear model function of the Stats package in R version 3.3.2 software (R Foundation for Statistical Computing).

To determine the association between gait variability and cortical thickness, we performed vertex-wise analyses using the FreeSurfer QDEC module (Query, Design, Estimate, Contrast (http://surfer.nmr.mgh.harvard.edu)), which allows users to perform inter-subject/group averaging and inference using the general linear model on morphometric data. We applied corrections for multiple comparisons using the built-in Monte Carlo simulation at a threshold of p = 0.05, a cluster-wise correction that controls for the rate of false positive clusters. In QDEC, we used a GLM with each gait parameter as the continuous predictor. Age and estimated total intracranial volume (eTIV) were set as nuisance variables within the different offset and slope design matrix. As the number of covariates in QDEC is limited, we exported each participant’s cortical thickness in the identified clusters to R to assess whether the associations withstood correction for confounding factors. To do so, we created a ROI for each cluster that was significantly associated with gait variability. We mapped this normalized ROI to each participant to generate a mean thickness value for that ROI for each participant. We performed additional linear model analyses using the mean cortical thickness of the ROIs as dependent variables and gait variability as an independent variable. We corrected for age, sex, education level, GDS, CIRS, leg length, presence of arthritis, and eTIV.

To determine the association between gait variability and the volumes of the subcortical grey matter structures (caudate, putamen, globus pallidus, thalamus, and nucleus accumbens), amygdala, hippocampus, and cerebellum, we created a multivariate GLM adjusted for age, sex, education, GDS, CIRS, leg length, presence of arthritis, and eTIV. False discovery rate correction was applied to correct for multiple comparisons. Eight ROIs were selected a priori from each hemisphere based on their known associations with gait control.

To determine the association between cognitive function measures and the cortical thickness and subcortical volume of the structures associated with the gait variability, we created a multivariate GLM that adjusted for age, sex, education, GDS and CIRS scores, and eTIV.

Lastly, we analyzed the mediation effect of the cortical thickness and subcortical volume of clusters that were significantly associated with both gait variability and cognitive function on the association between these factors (VMS, CERAD-TS) using the PROCESS macro developed for SPSS [29]. We performed parallel mediation analyses separately for each cognitive assessment using 5000 bootstrapped samples. In these analyses, we adjusted for sex, age, education, GDS, CIRS, and eTIV. Path a represents the effect of gait variability on the neuroimaging measures, whereas path b represents the effect of neuroimaging measures on cognition. Paths c and c’ indicate the total and direct effects of gait variability on cognition, respectively. The indirect effect (path a × b) measures the effect of gait variability on cognition via the cluster cortical thickness or subcortical volume. A significant indirect effect is indicated by 95% confidence intervals that do not include the value of 0.