Development of a Gait Feature–Based Model for Classifying Cognitive Disorders Using a Single Wearable Inertial Sensor

Abstract

Background and Objectives

Gait changes are potential markers of cognitive disorders (CDs). We developed a model for classifying older adults with CD from those with normal cognition using gait speed and variability captured from a wearable inertial sensor and compared its diagnostic performance for CD with that of the model using the Mini-Mental State Examination (MMSE).

Methods

We enrolled community-dwelling older adults with normal gait from the Korean Longitudinal Study on Cognitive Aging and Dementia and measured their gait features using a wearable inertial sensor placed at the center of body mass while they walked on a 14-m long walkway thrice at comfortable paces. We randomly split our entire dataset into the development (80%) and validation (20%) datasets. We developed a model for classifying CD using logistic regression analysis from the development dataset and validated it in the validation dataset. In both datasets, we compared the diagnostic performance of the model with that using the MMSE. We estimated optimal cutoff score of our model using receiver operator characteristic analysis.

Results

In total, 595 participants were enrolled, of which 101 of them experienced CD. Our model included both gait speed and temporal gait variability and exhibited good diagnostic performance for classifying CD from normal cognition in both the development (area under the receiver operator characteristic curve [AUC] = 0.788, 95% CI 0.748–0.823, p < 0.001) and validation datasets (AUC = 0.811, 95% CI 0.729–0.877, p < 0.001). Our model showed comparable diagnostic performance for CD with that of the model using the MMSE in both the development (difference in AUC = 0.026, standard error [SE] = 0.043, z statistic = 0.610, p = 0.542) and validation datasets (difference in AUC = 0.070, SE = 0.073, z statistic = 0.956, p = 0.330). The optimal cutoff score of the gait-based model was >−1.56.

Discussion

Our gait-based model using a wearable inertial sensor may be a promising diagnostic marker of CD in older adults.

Classification of Evidence

This study provides Class III evidence that gait analysis can accurately distinguish older adults with CDs from healthy controls.

Human gait involves both the peripheral and CNS and changes in speed, cadence, step length, step time, variability, and asymmetry with advancing age.¹ Human gait also changes in cognitive disorders (CDs) such as mild cognitive impairment (MCI) and dementia.^1-3 Slower gait and higher gait variability are associated with both prevalent and incident CDs.^2-6 Slower gait was associated with brain regions involved in processing speed,⁷ while more temporally variable gait was associated with those involved in cortical sensorimotor control.⁸ In line with these results, slower gait was associated with the risk of vascular dementia (VaD) or Parkinson disease (PD) dementia, while more temporally variable gait was associated with the risks of cognitive decline without cerebrovascular disease and/or PD.^2-5 It has been proposed that the motoric cognitive risk syndrome, which has both cognitive complaints and slow gait speed, is a novel high-risk condition for dementia, and it was found to be associated with the risks of VaD or PD dementia but not with the risk of Alzheimer disease (AD).^2,4

Although temporally variable gait was also associated with cognitive decline, no previous studies have defined a high-risk condition for dementia using temporal gait variability instead of gait speed or using both simultaneously, which may be, at least partly, attributable to the lack of reliable and accessible measurements of temporal gait variability. Although gait can be analyzed in many ways, laboratory-based 3-dimensional motion capture systems and instrumented walkway systems show low reliability in the measurement of gait variability because they are spatially limited and can thus analyze a limited number of steps.^9,10 Furthermore, such systems are too expensive to be used as a screening tool for CDs.

By contrast, gait analysis using a wearable inertial sensor is as reliable and valid as other gait analysis systems, but more cost-efficient and less spatially and temporally limited than other gait analysis systems.^9,11 In addition, a wearable inertial sensor is more suitable for ambulatory monitoring of subtle gait changes in real-life than other gait analysis systems.¹¹ Therefore, we developed and validated the models for estimating gait speed and variability using a wearable inertial sensor placed at the center of body mass (CoM) in our previous works.^9,12

In this study, we developed a model for classifying older adults with CD from those with normal cognition using gait speed and temporal variability captured from a wearable inertial sensor placed at the CoM, and we investigated its criterion and concurrent validities. The primary research question being addressed in this study was as follows: can a model using gait features captured by a single wearable inertial sensor accurately classify older adults with CDs from healthy controls?

Methods

Participants

We enrolled community-dwelling adults from the Korean Longitudinal Study on Cognitive Aging and Dementia (KLOSCAD),¹³ a nationwide population-based prospective cohort study of older Koreans. In the KLOSCAD, 6,818 community-dwelling Koreans aged 60 years or older were randomly sampled from 30 villages and towns across South Korea using residential rosters. The baseline evaluation of the KLOSCAD was conducted in 2010–2012, and follow-up evaluations were conducted every 2 years until 2020.

Among the 6,818 participants, 837 who were enrolled from Yong-In, the largest satellite city of metropolitan Seoul, were invited to this KLOSCAD-GAIT study. Among them, 595 were eligible for inclusion in this study after excluding the participants with the following conditions that could affect the gait features: a history of stroke; PD and other movement disorders; a history of unilateral knee, bilateral knee, or hip arthroplasty; knee, hip, or ankle joint surgery; and spinal diseases such as spinal stenosis, scoliosis, spondylolysis, and spondylolisthesis.^14,15

Standard Protocol Approvals, Registrations, and Patient Consents

All participants were fully informed of the study protocol and provided written informed consent signed by themselves. The study protocol was approved by the Institutional Review Board of the Seoul National University Bundang Hospital (IRB No. B-2107-696-115).

Clinical Assessment

Trained research nurses collected demographic and anthropometric data including age (years), sex, education (years), height (centimeters), weight (kilograms), and the presence of lower limb arthritis, such as osteoarthritis, rheumatoid arthritis, or posttraumatic arthritis. Research neuropsychologists or trained research nurses administered the Korean version of the Consortium to Establish a Registry for Alzheimer Disease (CERAD) Assessment Packet Neuropsychological Assessment Battery,¹⁶ Digit Span Test,¹⁷ Frontal Assessment Battery,¹⁸ and Mini-Mental State Examination (MMSE)¹⁹ to each participant. We defined objective cognitive impairment as performing −1.5 SD or below in any neuropsychological test except MMSE compared with age-stratified, sex-stratified, and education-stratified normative data of older Korean adults.^18,20

Geriatric psychiatrists or geriatrists conducted face-to-face study-specific interviews to obtain detailed medical histories, laboratory tests, and physical and neurologic examinations using the Korean version of the CERAD Clinical Assessment Battery¹⁶ and the Korean version of the Mini International Neuropsychiatric Interview.²¹ Then, a panel of geriatric psychiatrists confirmed the final diagnosis for each participant. We diagnosed dementia according to the Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition criteria,²²and MCI according to the Consensus Criteria from the International Working Group on MCI.²³ We determined AD according to the National Institute of Neurological and Communicative Disorders and Stroke and the Alzheimer Disease and Related Disorders Association criteria²⁴ and dementia with Lewy bodies (DLB) according to the consensus criteria proposed by McKeith et al.²⁵ We defined CD as having MCI or dementia and cognitively normal (CN) as functioning independently in the community and showing no evidence of cognitive impairment on objective neuropsychological tests.

Gait Analysis

As described in detail in our previous work,²⁶ we analyzed the gait of each participant within a month after the clinical assessment using a wearable inertial sensor, FITMETER (FitLife Inc., Suwon, Korea), or ActiGraph (SMD Solution, Seoul, Korea). The sensors were hexahedrons (FITMETER = 3.5 × 3.5 × 1.3 cm, weighting 14 g; ActiGraph = 3.0 × 4.0 × 1.0 cm, weighting 17 g) with smooth edges and have a digital triaxial accelerometer (BMA255, Bosch, Stuttgart, Germany) and a gyroscope (BMX055, Bosch). They measured triaxial acceleration up to ±8 g (resolution = 0.004 g/0.00024 g) and triaxial angular velocity up to ±1,000°/s (resolution = 0.03°/s) at a sample rate of 250 Hz. We fixed an inertial sensor to each participant with an elastic band at the level of third–fourth lumbar vertebrae, which is the approximate center of the body mass. All participants walked back and forth 3 times on a 14-m flat straight walkway at a comfortable self-selected pace and started turning after passing the 14-m line. We used the averaged data of the 3 walking trials in the analyses.

To measure steady-state walking by minimizing acceleration effects, we analyzed the data of the central 10-m mark of the 14-m flat straight walkways after eliminating the 2-m long walks before the start and each turn. We preprocessed the inertial sensor signals, identified each step, and estimated 6 gait features, including cadence (steps per minute), step time (seconds), gait speed (centimeters per second), step length (centimeters), step time variability (%), and step time asymmetry (milliseconds) using the methods described in our previous works.^12,26 We defined gait speed as the distance of body movement on level ground in centimeters per second and estimated the gait speed using the method described in detail in our previous work.²⁶ We estimated gait variability using step time variability. We defined step time as the duration of each step(s) from the initial contact of 1 foot to the initial contact of the opposite foot and estimated the step time by multiplying 60 with the inversed cadence. We defined the step time variability as the coefficient of variance (CoV, % = SD/mean × 100). We calculated the scores from the left and right steps separately and then averaged them (CoV_{Step time} = [CoV_{Left step time} + CoV_{Right step time}]/2).

Statistical Analysis

We compared the continuous variables using Student t tests and categorical variables using χ² tests between groups. To develop and validate a model for classifying CD, we randomly split our entire participants into the development (80%) and validation (20%) datasets. The proportion of CD was approximately 17% in both datasets. In the development dataset, we developed a regression-based model for classifying CD using a binary logistic regression analysis with forward selection. In the logistic regression model, we computed step time variability, gait speed, age, sex, education, height, and the presence of lower limb arthritis as independent variables. In addition, we developed a regression-based model for classifying CD using the MMSE score instead of step time variability and gait speed. We evaluated the goodness-of-fit of the models using the Hosmer-Lemeshow test with p > 0.05 and the Nagelkerke R² index. We interpreted the regression coefficient (β) as the relative weight of each independent variable on the dependent variable. We then calculated a weighted composite score using logit scores weighted by their coefficients from each logistic regression model.

We examined the diagnostic performance of the models using the receiver operating characteristic analysis and compared the area under the curve (AUC) between the models using the method proposed by Hanley and McNeil.²⁷ We estimated optimal cutoff of the weighted composite score and their sensitivity and specificity for CD using a bootstrap sampling estimation (100 resamples). In each run, 80% of the development dataset was randomly selected. We obtained bootstrap estimations of AUC, optimal cutoff, sensitivity, and specificity by aggregating the results of the 100 runs. We then compared the bootstrap-estimated mean AUC between the gait-based and MMSE-based models. We also validated these models by applying them to an independent validation dataset. We compared the AUC of the gait-based model obtained from the validation dataset with that obtained from the development dataset to check the possibility of overfitting. We also compared the AUC of the gait-based and MMSE-based models obtained from the validation dataset.

In the entire dataset, we also examined the correlations of the weighted composite score of the gait-based model, MMSE-based model, and MMSE score using the Pearson correlation analysis. We performed statistical analyses using R version 4.2.0 (Foundation for Statistical Computing, Vienna, Austria), the Statistical Package for the Social Sciences for Windows version 20 (IBM Corp., Armonk, NY), and MedCalc for Windows version 18.11.3 (MedCalc Software, Mariakerke, Belgium). Statistical significance was set at a 2-tailed p < 0.05.

Data Availability

Datasets used and/or analyzed in this study are available from the corresponding author on reasonable request.

Results

Among the 595 participants, 101 (17%) experienced CDs (94 MCI, 6 dementia due to AD, and 1 DLB). Compared with the CN group, the CD group was older, less educated, shorter, and had worse MMSE scores. Compared with the CN group, the CD group also showed lower cadence, longer step time, higher step time variability, higher step time asymmetry, slower gait speed, and shorter step length (Table 1).

Table 1.

Characteristics of the Participants

As summarized in Table 2, we developed a gait-based model and an MMSE-based model for classifying CD from CN from the development dataset using binary logistic regression analyses. In the gait-based model, gait speed, step-time variability, and education were selected as significant predictors of CD (2.286 − 0.040 × gait speed + 0.535 × gait variability − 0.115 × education). The gait-based model showed adequate goodness of fit (Hosmer–Lemeshow test χ² = 4.16, df = 8, p = 0.843, Nagelkerke R² = 0.243) and good classification accuracy (83.8%), with an AUC of 0.788 (standard error [SE] = 0.027, 95% CI 0.748–0.823, p < 0.001) in the development dataset. In the MMSE-based model, the MMSE score and age were selected as significant predictors of CD (5.977 − 0.429 × MMSE score + 0.055 × age). The MMSE-based model also showed adequate goodness of fit (Hosmer-Lemeshow test χ² = 12.99, df = 8, p = 0.112, Nagelkerke R² = 0.271) and good classification accuracy (86.1%) with an AUC of 0.762 (SE = 0.033, 95% CI 0.721–0.799, p < 0.001). Although the AUC of the gait-based model was larger than that of the MMSE-based model, the difference was not apparent (difference in AUC = 0.026, SE = 0.043, z statistic = 0.610, p = 0.542; Figure, A).

Table 2.

Logistic Regression Models for Classifying the Participants With Cognitive Disorder From Cognitively Normal Controls^a

Figure. Comparisons of the Diagnostic Performance for Cognitive Disorder Between Models and Datasets*.

(A) Comparison of the diagnostic performance of the gait-based model with that of the MMSE-based model and the MMSE score in the development dataset. (B) Comparison of the diagnostic performance of the gait-based model with that of the MMSE-based model and the MMSE score in the validation dataset. (C) Comparison of the diagnostic performance of the gait-based model in the development dataset with that in the validation dataset. AUC = area under the receiver operator characteristic curve; MMSE = Mini-Mental State Examination. *Hanley and McNeil method.²⁷

In the 100 bootstrapped resamples of the development dataset, the bootstrap-estimated mean AUC was 0.787 ± 0.015 for the gait-based model and 0.760 ± 0.016 for the MMSE-based model. The difference in the AUC between them was not statistically significant (p = 0.565). The optimal cutoff of the gait-based model for classifying CD was estimated to be −1.56 or below.

When we applied the models to the independent validation dataset, the diagnostic performance of the gait-based model was good (AUC = 0.811, SE = 0.046, 95% CI 0.729–0.877, p < 0.001). Although the AUC of the gait-based model was larger than that that of the MMSE-based model (AUC = 0.741, SE = 0.057, 95% CI 0.653–0.817, p < 0.001) as was in the development dataset, the difference was not apparent (difference in AUC = 0.070, SE = 0.073, z statistic = 0.956, p = 0.339, Figure, B). The diagnostic performance of the gait-based model in the validation dataset was comparable with that in the development dataset (difference in AUC = 0.023, SE = 0.053, z statistic = 0.431, p = 0.666, Figure, C). In addition, the correlations of the weighted composite score of the gait-based model with that of the MMSE-based model and MMSE score were significant in both the CD and CN groups (Table 3). When we performed all these analyses after excluding the participants with dementia from the CD group, the results were not changed (data are not shown).

Table 3.

Correlations of the Weighted Composite Score of the Gait-Based Model With That of the MMSE-Based Model and the MMSE Score^a

Classification of Evidence

This study provides Class III evidence that gait analysis can accurately distinguish older adults with CDs from healthy controls.

Discussion

We developed and validated a model for classifying CD using gait features captured by a single wearable inertial sensor placed at the CoM. Our model showed good diagnostic performance for CD, similar to the model using the MMSE, which is most widely used for screening CD in both clinical and research settings.¹⁹

Both gait speed and temporal gait variability were selected in our gait-based model, which is consistent with previous studies.^2-6 Because gait speed and temporal gait variability are associated with different types of CDs, both might have been selected in the model for classifying all-cause CDs. In previous neuroimaging studies, gait speed was associated with the volumes of the brain stem, cerebellum, primary and supplementary motor cortices, and prefrontal cortex regions that are involved in processing speed and executive function.⁷ By contrast, temporal gait variability was associated with the volumes of the primary sensorimotor cortex, hippocampus, anterior cingulate cortex, basal ganglia, and posterior thalamic radiation, which are involved in gait integration, memory, and executive function.⁸

Our gait-based model showed comparable diagnostic performance for CD with the MMSE-based model, and the weighted composite score of our gait-based model correlated well with that of the MMSE-based model. Compared with the MMSE-based model, our gait-based model showed a higher sensitivity but showed a lower specificity. Although the MMSE is most widely used for screening CDs in both clinical and research settings, it has some shortcomings as a screening and/or monitoring instrument for CDs. It cannot be self-administered,²⁸ and it is subject to socioeducational influence and the learning effect when administered repeatedly.²⁸ Furthermore, it is insensitive to early stage CDs.²⁹ These shortcomings can be overcome by using our gait-based model. Gait analysis using a wearable inertial sensor can be easily and repeatedly self-administered at home and is robust to socioeducational influences and learning effects. Therefore, our gait-based model is advantageous for both screening and monitoring CDs and for capturing real-life functions. Currently, most smartphone models are equipped with triaxial acceleration sensors.¹¹ Therefore, our gait-based model can be easily packaged into a mobile application in the future because the smartphone itself can act as both an inertial sensor and an inertial signal processing device without a separate external inertial sensor.³⁰ This makes it possible to use our model more cheaply and widely available. In addition, older adults with CD are at a higher risk of falling and frequently comorbid with other geriatric diseases that may also influence their gait.³¹ By adding additional gait-based models for detecting other geriatric conditions, we may develop smartphones into a comprehensive health platform based on gait analysis in the future.

Wearable inertial sensors are known to be reliable and valid in estimating spatiotemporal gait features of older adults.^9,11 In addition, the CoM was reported to be an optimal location for placing an inertial sensor in capturing gait features.¹¹ The accuracy and reliability of our method for estimating gait features of older adults using a single inertia sensor placed at the CoM was found to be excellent in our previous work (intraclass correlation coefficient with an instrumented walkway system = 0.91–0.96).⁹ In addition, we used the step time variability for estimating gait variability in this study because temporal gait features were more reliable³² and more sensitive to early abnormal gait variability than the spatial gait variability.⁸

This study has several limitations. First, we confirmed the performance of our model in a separate validation dataset and found that the performance was comparable between the development and validation sets, indicating little possibility of overfitting. However, we did not conduct an external validation and may need to be validated in other populations to verify its generalizability. Second, we developed a model to detect all-cause CD. If the gait features were associated with specific types of CD, the sensitivity and specificity for CD of our model may be different in the populations with different proportions of CD types.³³ To overcome this limitation, models for detecting cause-specific CD need to be developed using a larger sample. Third, a single inertial sensor may not be able to provide reliable estimate of gait features in certain geriatric conditions. For example, multiple inertial sensors located across paretic and nonparetic limbs significantly improved the accuracy of gait features in poststroke patients.³⁴ Finally, our study used a cross-sectional design. Its predictive validity for incident CD should be examined in a prospective study.

In conclusion, we developed a gait-based model for classifying CD using a wearable inertial sensor. If its validity is proved in future studies on independent samples, it may help to make the screening and monitoring of CD easier, wider, and cheaper for older adults.

Glossary

AD

Alzheimer disease

AUC

area under the curve

CD

cognitive disorder

CERAD

Consortium to Establish a Registry for Alzheimer Disease

CN

cognitively normal

CoM

center of body mass

CoV

coefficient of variance

DLB

dementia with Lewy bodies

KLOSCAD

Korean Longitudinal Study on Cognitive Aging and Dementia

MCI

mild cognitive impairment

MMSE

Mini-Mental State Examination

PD

Parkinson disease

SE

standard error

VaD

vascular dementia

Appendix. Authors

Study Funding

This work was supported by a grant from the Korean Health Technology R&D Project, Ministry of Health and Welfare, Republic of Korea (grant no. HI09C1379 [A092077]), and PlanB4U Co., Ltd. (grant no. 06-2021-0273).

Disclosure

The authors report no disclosures relevant to the manuscript. Go to Neurology.org/N for full disclosures.