Fusing GPS, Activity Monitors, and Self-Report to Improve Assessment of Walking Activity and Community Participation After Stroke

Abstract

Background and Purpose:

Walking and participation in the community are important goals for people post-stroke (PPS). These constructs are challenging to measure given limitations in current data collection methodologies. The purpose of this study was to (1) develop a data fusion approach that combined data from global positioning system (GPS), activity monitor (AM), and daily trip log to identify walking activity and participation in the community, and (2) to examine the construct validity of the data fusion method.

Methods:

At 60 days post-stroke, PPS wore a GPS and AM and completed a daily trip log for 7 days. Using a combination of a density-based spatial clustering algorithm and geocoding GPS, AM, and daily trip log data were time synched and fused to identify total trips taken outside the home; locations visited per trip; number of steps taken in the home, in the community, at each location visited, and in total. Associations between stroke outcomes and the data fusion metrics were determined to support the construct validity of the data fusion method.

Results:

Forty-four PPS took a mean of 2,541 steps/day, of which 56% were in the community, and took a mean of 0.39 trips/day outside the home and visited a mean of 0.42 locations. A social visit was the most common reason for going into the community. There were fair associations between number of trips outside the home and gait speed (GS), r = 0.49, Berg Balance Scale (BBS), r = 0.48, modified Rankin Scale (mRS), r = −0.47, and Stroke Impact Scale participation subscale (SIS-P) (0.45). There were moderate associations between steps taken in the community and GS, r = 0.63, BBS, r = 0.51, mRS, r = −0.61, and SIS-P, r = 0.43.

Discussion and Conclusions:

Participants did not often access their community. Fusing GPS, AM, and trip log data may provide a comprehensive method to identify walking activity and community participation in PPS.

INTRODUCTION

Improving walking activity in the home and community and enhancing participation in the community are important goals for many people who have experienced a stroke.¹ The ability to walk independently in the community enables greater participation and reduces burden of care.² However, many people post-stroke (PPS) do not frequently access their community and are not able to walk independently outside their home.³ Although walking in and being able to participate in the community is a major focus of rehabilitation efforts, our ability to measure these constructs is limited. There is a need to develop comprehensive methods to measure these constructs to better understand the barriers that PPS face.

Self-report measures, clinic-based measures of capacity, activity monitors (AM), and more recently global positioning systems devices (GPS) are used to assess walking performance in the home and community, as well as PPS’s ability to participate in their community. While each method provides valuable insights into activity and participation, when used singularly, each is subject to limitations that may compromise data accuracy and validity. Self-report diaries offer unique personal perspectives from PPS but are subject to self-report bias in which the respondent reports information that does not reflect their true behaviors, or due to memory deficits the respondent may not be able to accurately recall.

Clinical measures, such as gait speed and gait endurance, are often used to infer walking activity in the home and community.⁴ However, capacity measured in a clinical setting may not reflect real-world walking.⁵ Activity monitors can track walking activity accurately in the home and community,^6–8 but they lack the ability to distinguish where the activity took place (eg, home vs community) and fail to provide context regarding the purpose and goals of the walking activity. More recently, combining GPS devices with AMs has enhanced the ability to measure community walking activity in PPS.^9–11 When paired with AM, GPS can provide environmental context, indicating where the walking occurred.¹² This combination yields more detailed information about the link between walking activity and participation. However, GPS devices suffer from limitations in accuracy and completeness. For instance, users may not always carry the device with them when they leave the home, or the signal may be lost in certain environments.¹³

Combining self-report diaries with AMs and GPS may offer accurate and comprehensive insights into walking activity and participation in the community.¹⁴ When one data source (GPS, self-reported diaries, or AM) is missing, incomplete, or not accurate; fusing these data sources together may improve our ability to measure these constructs. To the best of our knowledge only one other study has combined AM, GPS, and self-report diaries to attempt to understand community mobility and participation after stroke.⁹ However, these authors did not provide details on how the data from the different sources were combined or how missing data were managed. The purpose of this study was (1) to develop a novel data fusion approach, which fuses granular individual data on location and the reason for the community activity from self-report diaries, AM, and GPS to provide insight into walking activity and participation in community after stroke, and (2) determine the construct validity of this data fusion approach to identify walking activity in the home and community and participation in the community. We explored the construct validity in 2 ways: (1) the cross-sectional association between clinical measures and self-report with walking activity and community participation identified by the data fusion approach and (2) the longitudinal association of clinical measures taken early after stroke with walking activity and community participation identified by the data fusion approach at 60 days post-stroke.

METHODS

Recruitment

Participants were recruited from 2 inpatient rehabilitation hospitals, 1 in the Northeast and 1 in the Midwest of the United States, and are part of a larger, ongoing cohort study to determine the prevalence of non-obstructive sleep apnea (OSA) sleep disorders and their impact on recovery after stroke.¹⁵ Inclusion criteria were diagnosis of stroke, ≥18 years old, National Institutes of Health Stroke Scale (NIHSS) item 1a score < 2 (alert or not alert but arousable by minor stimulation to obey, answer, or respond), provide informed consent or assent, with the participant’s legal guardian providing consent. Exclusion criteria were pre-stroke diagnosis of OSA, oxygen desaturation index (ODI) ≥15, living in nursing home or assisted living prior to stroke, unable to ambulate at least 45.7 meters (150 feet) without physical assistance prior to stroke, other neurologic health condition that could impact recovery, pregnant, recent hemicraniectomy or suboccipital craniectomy, planned discharge location greater than 241 kilometers (150) miles from recruiting hospital, and global aphasia as determined by a score of 3 on NIHSS item 9. See the protocol for the parent study for more detailed information.¹⁵ The funders played no role in the design, conduct, or reporting of this study. All participants provided consent or assent (with participant’s legal guardian providing consent) and the study was approved by all local Institutional Review Boards (IRB) and a central IRB.

Data Collection

During inpatient rehabilitation and at 60 days post-stroke, data were collected on living situation, marital status, National Institutes of Health Stroke Scale¹⁶ (NIHSS, only collected during inpatient rehabilitation), modified Rankin Scale¹⁷ (mRS), Stroke Impact Scale (SIS) 16, SIS mobility (SIS-M) subscale, and SIS participation (SIS-P) subscale (SIS only collected at 60 days post-stroke),^18,19 Patient Health Questionnaire 9 (PHQ-9),²⁰ Montreal Cognitive Assessment (MOCA),²¹ Berg Balance Scale (BBS),^22,23 and gait speed (GS) measured over the middle 6 meters of a 10 meter walk at self-selected speed.^23,24 At 60 days post-stroke data were collected where participants resided.

To assess home and community walking activity and participation in the community at 60 days post-stroke an ActivPal AM, a QStarz BT-Q1000XT GPS device, and a daily trip log were used. These were used for 7 days. The AM was a small thigh worn accelerometer (23.5 mm wide, 43 mm long, 5 mm thick and weighs 9.5 grams [0.93 in × 1.7 in × 0.2 in and weighs 0.34 ounces]) that was attached by one of the researchers to the participant’s least affected thigh using Tegaderm. The AM stayed on the participant 24 hours/day for the full 7 days. It tracked time in different postures (stepping, standing, sitting, lying, seated transport) and steps. The GPS device was approximately the size of a pager (7 cm long × 2 cm thick, and 4.5 cm wide). Participants were instructed to keep it turned on all the time and to charge it every night. They were instructed to carry it with them whenever they left their home in their pocket or purse. The participants were provided a daily trip log on which they were instructed to record the date, departure and return times, mode of travel (car, walk, wheelchair, public transportation, other) and trip destination type (work, social, recreation, medical, shopping, religious, other) for every time they left the home. The ActivPal and GPS are valid and reliable measure of these metrics.^8,10

Data Analysis

Descriptive statistics were used to describe the sample characteristics.

Data Fusion Methods

We developed a 2-stage data fusion approach to obtain robust and granular information of walking patterns summarized in Figure 1. In the first stage, we time linked individual-level information from both AM and GPS data sources and apply a density-based spatial clustering algorithm of applications with noise (DBSCAN) algorithm^25,26 to identify geographic locations with high degrees of residency based on GPS reported locations. In the second stage, we use the self-reported trip logs to identify date-times in which the GPS location data does not correspond to a self-reported trip from the daily trip log or where GPS location data is missing.

Figure 1.

Overview of data fusion methodology. AM, activity monitor; GPS, global positioning system; DBSCAN, density-based spatial clustering algorithm of applications with noise.

Stage I:

Only participants who had 4 or more complete days (continuous 24 hours) of AM data and observations from GPS and or self-reported daily trip logs were in the analyses. Individual AM and GPS data were linked based on 15-second intervals of exact date-time values providing granular temporal information on walking activity matched to GPS coordinates when available. DBSCAN^25,26 was implemented on the GPS data to identify areas of high density of locations separated by areas of low density based on Euclidean distance from one point to another. The DBSCAN algorithm makes an implicit assumption that for each point of a cluster the neighborhood of a given radius must contain at least a minimum number of points, ie, the density in the neighborhood has to exceed some threshold. The shape of a neighborhood is determined by the choice of a distance function for 2 points p and q, denoted dist(p,q).²⁶ We defined a cluster to be a collection of points within a 20-meter radius of neighborhood around a given core point with a minimum number of 20 points within the radius to create a cluster. Sensitivity analyses were performed for 20, 40, and 50 meters which did not change identification of clusters. The “home” location (cluster) was identified as the geographic coordinate pair reported between 12am-6am.

Stage II:

In this stage, self-reported trip logs were integrated with the AM and GPS data processed in Stage I. This was done by matching the start and end times of the self-reported trips with the combined AM and GPS data. In cases where GPS data were incomplete or showing a home cluster during a self-reported trip, we assumed the GPS was left at home or was not functioning during the time of the trip. In such cases GPS coordinates were set to “unknown.” In cases where no GPS data matched a self-reported trip, we adjusted the individual movements by incorporating “clusters” form the self-reported data to account for these trips. This ensured that even trips without GPS data were included in the overall movement and location analysis. To summarize each day’s activities, the granular 15-second data points were aggregated to generate daily location specific summaries. These summaries were included in the following details: (1) if available from Stage I, GPS coordinates were mapped, and the corresponding trip destination type was identified using the R package tidygeocoder,²⁷ (2) self-reported trip destination type (if available from the daily trip log), which could include social, shopping, work, recreational, medical, religious, and other, and (3) the total number of steps recorded at each location and during each trip by the AM was included. For walking activity identified by the AM that was not assigned to either a geographic location or a self-reported trip (ie, unknown locations or trip type destination), we applied a multiplier to the steps not assigned to a specific destination type defined as the ratio of the total number of community steps to the total number of known home steps, thereby assigning a proportion of unknown steps to home and community based on observed individual walking patterns incorporated into final summary measures. If the proportion multiplier is 0, all unknown steps are classified as in-home steps; conversely, if the multiplier is 1, all unknown steps are attributed entirely to community steps. Using a proportion multiplier to handle unknown locations ensures that missing data is accounted for in a systematic and consistent way, allowing for the distribution of unassigned steps based on the individual’s observed patterns, thus maintaining the integrity of the overall activity analysis.

Summary Metrics Derived From Data Fusion Approach

From the data fusion of the GPS, AM, and self-report daily log we extracted the following summary variables related to home and community walking activity and participation: total trips taken outside the home; locations visited on each trip; number of steps taken in the home, in the community, at each location visited, and total.

Construct Validity of Fusion Approach Through Correlation and Regression Analysis

Direct examination of the validity of the data fusion approach in identifying home and community walking activity and participation in the community is challenging due to a lack of a gold standard or ground-truth data. We indirectly assessed construct validity using 2 approaches. First, we used Pearson r correlation analyses to examine the association between walking activity in the home and community (total steps, steps in the home, and steps in the community) and participation variables (total trips taken, locations visited during trips) identified by the data fusion approach and clinical and self-report outcomes taken at 60 days post-stroke (GS, BBS, SIS-16, SIS-M, SIS-P, mRS, MOCA, and PHQ-9). We used guidelines from Portney²⁸ to interpret the associations: ≤.25 little or no relationship; .25 to .50 low to fair, .50 to .75 moderate to good, ≥.75 strong relationship. Associations greater than 0.25 would be supportive of the construct validity of the data fusion methods.

The second approach consisted of performing 4, separate multiple linear regressions across 4 dependent variables determined from the data fusion methods at 60 days post-stroke (mean total steps, mean steps in the community, mean number of trips outside the home, and mean number of locations visited outside the home) using GS at discharge from inpatient rehabilitation, MOCA at admission to inpatient rehabilitation, PHQ-9 at admission, and NIHSS at admission to inpatient rehabilitation as the independent variables. The data were reviewed to ensure that all the assumptions (multicollinearity, linearity, normality of residuals, homoscedasticity of residuals, and outliers) for performing multiple regression analyses were met. These key assumptions were met. Because of the complex nature of participation and community walking we considered that a R² greater than 0.10 would support the construct validity of the data fusion methods.

RESULTS

Data from 79 participants were collected during inpatient rehabilitation. At 60 days post-stroke 68 of the 79 (86%) participants had at least 4 days of complete days of AM data. Of these, 44 (56%) participants had observations from GPS and or self-reported daily trip logs. Data from the 44 participants with at least 4 complete days of AM, and GPS or self-reported daily trip logs observations were included in the analyses. Participants exhibited mild to moderate stroke¹⁶ with a mean age (standard deviation) of 66.2 (13.7) years old, 20 (45%) were males and 24 (55%) were females, 35 (79.5%) were white and 9 (20.5%) were Black, and 40 (91%) lived at home with family/caregiver, 3 (7%) lived alone, and 1 (2%) resided in a skilled nursing facility (see Table 1). Four participants were working at 60 days post-stroke.

Table 1.

Participants’ Characteristics at 60 and 10 Days Post-Stroke

Characteristic	10 Days Post-Stroke	60 Days Post-Stroke
Days since stroke, mean (standard deviation)	9.0 (7.3)	60.3 (4.7)
Living situation	N/Aa
Home with family/caregiver		40/44 (91%)
Home alone		3/44 (6.8%)
Subacute rehab/skilled nursing facility		1/44 (2.3%)
Gait speed (m/s), mean (standard deviation)	0.50 (0.30)b	0.64 (0.40)
Amount of assistance while walking
Total assistance (subject performs 0–24% of task, score of 0 should be documented)	2 (5%)	2 (4.7%)
Moderate assistance (subject performs 50–74% of task)	2 (5%)	0
Minimum assistance (subject performs 75–99% of task)	11 (27%)	4 (9.3%)
Supervision (subject requires stand-by or set-up assistance, no physical contact is provided)	16 (39%)	5 (12%)
Modified independent (subject requires use of assistive devices or bracing, needs extra time, mild safety issues)	7 (17%)	10 (23%)
Independent	3 (7%)	22 (51%)
Assistive device
None	10 (24%)	24 (56%)
Single point cane	0	2 (4.7%)
Small based quad cane	1 (2%)	4 (9.3%)
Large based quad cane	0	1 (2.3%)
Rolling walker	22 (54%)	8 (19%)
Standard walker	0	1 (2.3%)
Platform walker	5 (12%)	3 (7.0%)
Berg Balance Scale, mean (standard deviation)	32.3 (12.0)b	39.8 (15.7)
PHQ-9c, mean (standard deviation)	4.1 (4.8)	4.00 (4.3)
MOCAd, mean (standard deviation)	21.1 (5.3)	22.9 (4.3)
Modified Rankin Scale	N/Te
1		8 (18%)
2		15 (34%)
3		13 (30%)
4		8 (18%)
SISf mobility subscale, mean (standard deviation)	N/Te	0.71 (0.21)
SISf participation subscale, mean (standard deviation)	N/Te	0.65 (0.23)
SISf 16, mean (standard deviation)	N/Te	0.74 (0.18)
NIHSSg, mean (standard deviation)	5.2 (3.5)	N/Te

^aN/A, not applicable.

^bGait speed and Berg Balance Scale taken at discharge from inpatient rehab (27.6 [11.2] days after stroke).

^cPHQ-9, Patient Health Questionnaire 9.

^dMOCA, Montreal Cognitive Assessment.

^eN/T, not tested at this time point.

^fSIS, Stroke Impact Scale.

^gNIHSS, National Institutes of Health Stroke Scale.

Community Participation From Data Fusion Approach

Our data fusion approach identified that participants took a mean (standard deviation) of less than 1 trip outside the home per day (0.39 [0.26]), visited less than 1 location per day (0.42 [0.50]), and visited less than 1 location during a trip outside the home (0.77 [0.64]). 34% (15/44) of the participants did not leave their home during the data collection period, while 66% (29/44) left their home at least once. Participants usually visited only 1 location during a trip outside the home and at most 3 locations. A visit for social purposes was the most common location visited, followed by medical and shopping visits. See Table 2 and Figure 2 for more details on trips out of the home into the community. The Supplemental Digital Content, Appendix, available at: http://links.lww.com/JNPT/A530 provides examples of community participation for selected individual participants.

Table 2.

Trips Outside the Home and Types of Locations Visited During Trips Outside the Home

Characteristic	Mean/Day (SD)	Range	Percentage of Participants Who Had at Least One Visit to the Location
Trips outside the home	0.39 (0.26)	0–8	70%
Locations visited	0.42 (0.50)	1–16	N/Aa
Locations visited/trip outside the home	0.77 (0.64)	1–3	N/Aa
Visits to work	0.02 (0.07)	0–2	9%
Visits to a medical facility	0.11 (0.11)	0–3	59%
Social visits	0.17 (0.20)	0–6	64%
Recreational visits	0.04 (0.10)	0–3	18%
Shopping visits	0.08 (0.13)	0–4	41%

^aN/A, not applicable.

Figure 2.

Mean number of visits across 60-day study period broken down by location type.

Home and Community Walking Activity From Data Fusion Approach

Participants took 2541 (2413) mean (standard deviation) steps/day across the days they were monitored, 44% (1123 steps/day) were in the home and 56% (1418 steps/day) were in the community, this includes 17 participants who did not take any steps outside the home (Figure 3). Participants accumulated the most steps while on a visit to a work location, 12,590 mean steps/work visit. Twenty-eight participants accumulated 2090 (4179.7) mean (standard deviation) steps/visit to a medical facility. See Table 3 for further details on where in the community walking activity occurred. See Supplemental Digital Content, Appendix, available at: http://links.lww.com/JNPT/A530 for examples of home and community walking activity for selected individual participants.

Figure 3.

Walking activity in the home and community.

Table 3.

Walking Activity in the Home and Community

Characteristic	Mean (Standard Deviation)	Number of Participants
Steps/day	2541 (2413)	44
Steps/day in the home	1123 (1537)	44
Steps/day outside the home	1418 (2310)	44
Steps while on a visit to work	12 590 (13 428)	4
Steps while on a visit to a medical facility	2090 (4179)	26
Steps while on a social visit	1120 (1153)	28
Steps while on a recreational visit	1327 (854)	8
Steps while on a shopping visit	948 (846)	18

The 44 participants had a mean of 7.1 complete AM days, 4.6 complete GPS days, and 2.8 complete daily trip log days. Both GPS and daily trip log data were available for 30 participants, only GPS data were available for 12 participants, and only daily trip log data were available for 2 participants. On average, 71% of steps/participant were re-allocated applying the proportion multiplier. The number of unknown steps were, on average, higher for those participants who did not have any recorded community trips versus those that had community trips, 82.3% to 64.9%, respectively.

Validity of the Data Fusion Approach

There were statistically significant fair associations between trips out of the home to visit different locations, and clinical and self-report outcomes. There was a fair association between GS and mean trips taken out of the home (r = 0.49, P < 0.01), mean locations visited (r = 0.43, P < 0.01), mean locations visited/trip (r = 0.35, P < 0.05), mean visits to work (r = 0.35, P < 0.05), mean visits for social activities (r = 0.46, P < 0.01), and mean visits for shopping (r = 0.42, P < 0.01) (see Figure 4A).

Figure 4.

Correlation matrices. (A) Trips, trips taken into the community; locations, locations visited during trips into the community; loctrip, locations visited per trip into the community; rec, recreational trip; shop, shopping trip; gaits, gait speed; bbs, Berg Balance Scale; phq, Patient Health Questionnaire 9; moca, Montreal Cognitive Assessment; mrs, modified Rankin Scale; sis_mob, Stroke Impact Scale Mobility subsection; sis-part, Stroke Impact Scale Participation subscale; sis_16, Stroke Impact Scale 16 subscale. (B)Steps, total steps taken; home, steps taken in the home; comm, steps taken in the community; rec, recreational trips; shop, shopping trip; gaits, gait speed; bbs, Berg Balance Scale; phq, Patient Health Questionnaire 9; moca, Montreal Cognitive Assessment; mrs, modified Rankin Scale; sis_mob, Stroke Impact Scale Mobility subsection; sis-part, Stroke Impact Scale Participation subscale; sis_16, Stroke Impact Scale 16 subscale.

There were statistically significant fair to good associations between walking activity in the home and community and clinical and self-report outcomes. There were good associations between GS and mean total steps (r = 0.74, P < 0.001) and mean steps in the community (r = 0.63, P < 0.001), and a fair association with mean steps during a social visit (r = 0.47, P < 0.05) (see Figure 4B).

We summarized regression findings in Table 4. GS at discharge from inpatient rehabilitation was the only statistically significant predictor associated with total mean steps/day (P = 0.03). The adjusted R² illustrates that the model accounted for approximately 37% of the variability of the mean total steps/day taken at 60 days post-stroke. MOCA score at admission to inpatient rehabilitation was the only significant predictor associated with mean steps/day taken in the community (P = 0.02). The adjusted R² illustrates that the model accounted for approximately 18% of variability in the mean steps/day taken in the community at 60 days post-stroke. The overall fit for the model for determining mean trips/day taken into the community and mean visits/day to different locations at 60 days post-stroke was not statistically significant (P > 0.05).

Table 4.

Regression Analysis Results for Validity Assessment of the Data Fusion Approach

Regression Model
Y = mean total steps/day
	Estimate	Standard Error 95% CI	t value	P
Intercept	−98.63	−3,472.7 to 3275.5	−0.06	0.95
GS	2865.00	262.08 to 5467.92	2.23	0.03*
MOCA	87.17	−19.04 to 193.37	1.66	0.10
PHQ-9	−68.17	−183.06 to 46.72	−1.20	0.24
NIHSS	−112.97	−332.68 to 106.73	−10.4	0.30
F(4,36) = 6.85, P = 0.00	R2 = 0.43 Adj. R2 = 0.37
Regression Model
Y = mean steps/day in the community
	Estimate	Estimate 95% CI	t value	P
Intercept	−2262.94	−6151.36 to 1625.47	−1.18	0.25
GS	1593.26	−1406.42 to 4592.94	1.08	0.29
MOCA	151.46	29.06 to 273.85	2.51	0.02*
PHQ-9	−23.88	−156.29 to 108.52	−0.37	0.72
NIHSS	−67.01	−320.20 to 186.18	−0.54	0.59
F(4,36) = 3.22, P = 0.02	R2 = 0.26 Adj. R2 = 0.18
Regression Model
Y = mean trips taken into the community/day
	Estimate	Estimate 95% CI	t value	P
Intercept	−0.06	−0.55 to 0.44	−0.23	0.82
GS	0.22	−0.16 to 0.60	1.17	0.25
MOCA	0.02	0.00 to 0.03	2.05	0.05
PHQ-9	0.00	−0.02 to 0.01	−0.45	0.66
NIHSS	0.00	−0.03 to 0.04	0.20	0.84
F(4,36) = 2.02, P = 0.11	R2 = 0.18 Adj. R2 = 0.09
Regression Model
Y = mean visits to location into the community/day
	Estimate	Estimate 95% CI	t value	P
Intercept	−0.48	−1.45 to 0.49	−1.01	0.32
GS	0.45	−0.30 to 1.19	1.21	0.23
MOCA	0.03	0.00 to 0.06	2.10	0.04*
PHQ-9	−0.01	−0.04 to 0.02	−0.63	0.53
NIHSS	0.01	−0.05 to 0.07	0.35	0.73
F(4,36) = 2.12, P = 0.10	R2 = 0.19 Adj. R2 = 0.10

Regression analyses are categorized by dependent variable (mean total steps/day, mean steps/day in community, mean trips/day in community, and mean visits to location/day). Summary measures include coefficient estimate, 95% confidence intervals (CIs), t value, and P value. P values <0.05 are denoted with *. GS, gait speed; MOCA, Montreal Cognitive Assessment; PHQ-9, Patient Health Questionnaire 9; NIHSS, National Institutes of Health Stroke Scale.

DISCUSSION

Our data fusion approach that combines data from GPS, AMs, and self-report trip logs provides a comprehensive estimate of home and community walking activity and community participation in PPS. Based on the fusion of the 3 data sources, we obtained information on walking and participation trends when one data source alone may not provide sufficient or accurate information. Overall participant trends in walking activity and participation could be used by researchers to better understand the impact of interventions on performance in the real world. Individual level trends (see Supplemental Digital Content, Appendix, available at: http://links.lww.com/JNPT/A530) could be used by clinicians to guide treatment decisions.

Combining data from 3 sources allowed us to gain unique insights in to walking activity and participation in the community. We were able to uniquely identify where steps were taken (in the home, community, and specific community locations), number of trips taken into the community, and types of locations visited in the community. Other researchers have used GPS alone to identify trips taken into the community,²⁹ a pedometer combined with a self-report trip log to identify total steps and number of community trips taken,³⁰ and a GPS combined with a self-reported trip log and AM to identify trips and steps taken in the community.⁹ Although similar to our data fusion approach in terms of using 3 data sources, these authors did not provide detail on how the data from the 3 sources were combined. Our findings that there were less complete days of GPS and daily trip log data compared to AM data illustrate the benefits of fusing these data together. No one source of data may be sufficient to understand community walking activity and participation after stroke. The fact that many of the steps taken by the participants were not assigned to a location using either GPS or the daily trip log further supports the need to use more than 1 data source. Our data fusion approach provides a reproducible method to identify locations visited in the community and where walking activity occurred.

Due to the limitations associated with AM, GPS, and self-report daily log data we cannot determine the accuracy of our fusion method with 100% certainty. There were fair to good associations between the data fusion metrics (trips into the community, number of visits to different locations, total steps, and steps in the community) and measures of mobility, balance, cognition, disability, and participation supporting the construct validity of the data fusion method. Participants with greater mobility and balance, better cognition, lower disability, and higher levels of self-reported participation were more likely to go into the community, visit more locations when in the community, and take more steps. These findings are similar to those of Jersey and colleagues,³¹ who found fair associations between mental health and self-reported participation, and Fulk and colleagues,³² who found that walking endurance, balance, disability, and participation were associated with stepping activity.

The findings that factors early after stroke were associated with some of the data fusion metrics at 60 days post-stroke provides additional support for the construct validity of this approach. Gait speed at discharge from inpatient rehabilitation accounted 37% of the variance in total walking activity. Handlery and colleagues³³ found that total steps taken/day and balance (BBS) at 2 months post-stroke in community dwelling individuals explained 38% of the variance in total steps taken at 1 year post-stroke. Cognition at discharge from inpatient rehabilitation was significantly associated with walking activity in the community, trips taken into the community, and visits to locations in the community at 60 days post-stroke. Although the R² value was relatively small. Cognition early after stroke may be an important factor in community participation and should be explored in future research.³⁴

Similar to other studies,^9,29,30 participants did not often venture out of their home into the community and were limited in their walking activity. Across 7 days of monitoring, they took less than 1 trip outside the home/day on average. Approximately 35% of the participants did not leave their home at all and when participants did leave their home, they most often went to only 1 location. Walking activity was also limited both in the community and overall. Participants took a mean of approximately 2500 steps/day of which 56% was in the community.

Our approach to identify where walking activity occurred in the community and visits to different types of community locations that fuses GPS, AM, and a daily trip log provides unique insights, however there are limitations in our findings. This is a first step and has a relatively small sample size during which walking activity and community access was measured over only approximately 1 week. Although we were able to identify the type of location visited using the daily trip log and geocoding methods, we are not able to capture the full context of what occurred during the visit. We are not able to understand the importance of the trip and its impact on quality of life to the individual. Qualitative methods combined with our data fusion approach may be valuable to provide a deeper understanding of the importance of different types of community trips.

Using GPS to monitor movement and activity raises ethical considerations for privacy and confidentiality.^35,36 We also acknowledge a limitation of the simplicity in our multiplier approach to account for observed patters in steps when location and trip data were not available. However, the approach we used ensures that missing data is accounted for in a systematic and consistent manner. Using a 6MWT instead of GS may have provided greater insights into the construct validity of the data fusion approach.⁴ Additionally, our sample did not include people with stroke with moderate to severe OSA. People with stroke with OSA are likely to have lower activities of daily living function and greater disability than those with stroke without OSA.³⁷ However, it is not known how OSA may impact walking activity or trips into the community after stroke.

CONCLUSION

Fusing data from GPS, AMs, and a daily trip log, offers a comprehensive way to measure walking activity in the home and community, and participation in the community after stroke that addresses the limitations of each of these methods when used in isolation. The validity of this approach was supported by the associations between the data fusion metrics and the various clinical measures taken early and at 60 days after stroke.

Supplementary Material

Supplemental digital content is available for this article. Direct URL citations appear in the printed text and are provided in the HTML and PDF versions of this article on the journal’s Web site (www.jnpt.org).

Video Abstract available for more insights from the authors (see the Video, Supplemental Digital Content, available at: http://links.lww.com/JNPT/A529).