Introduction
Active travel, predominantly by walking and cycling, is an accessible form of physical activity that is associated with positive health outcomes.152–154 Quantifying the proportion of time spent in different active travel modes is therefore important to understand how these contribute to overall physical activity and health, and to assess the effectiveness of interventions that aim to increase active travel.
Travel modes have previously been assessed using detailed travel diaries;155 however, self-reported data have limitations, including that they may be subject to social desirability and recall bias and will often relate to only a single day of travel.156 Activity and movement patterns are now increasingly objectively assessed using accelerometers and GPS receivers;157–159 combining accelerometer and GPS data allows for the identification of both the intensity and the location of physical activity. This combination is potentially valuable in describing travel behaviour, and particularly active travel behaviour. In recent years, supervised machine learning has shown the potential to identify active travel from physical activity data. Supervised machine learning algorithms are trained on an example data set and are then used for prediction to other data sets. Until recently, the most promising algorithms were random forests: ensemble supervised learning algorithms where predictions are taken from a consensus across a large number of decision trees.160–162 A related algorithm, gradient boosted trees, has recently replaced random forests as a leading algorithm for data science tasks, with many machine learning approaches using the XGBoost (University of Washington, Seattle, WA, USA) implementation of gradient boosting instead of random forests.163,164
Using data collected in the ENABLE London study, an algorithm was developed to distinguish five travel modes (walk, cycle, motorised vehicle, train and stationary) using accelerometer and GPS data and the supervised machine learning tool XGBoost. We then applied this algorithm to the complete baseline and follow-up ENABLE study data sets to characterise the different modes of travel throughout the day and then examined changes between baseline and follow-up in the different travel modes. All code necessary to replicate our findings and apply our predictive model to other data sets is made available as a package of the open-source statistical software environment R (The R Foundation for Statistical Computing, Vienna, Austria).165 Full details of the methodology and a full usage example have been published.165,166
Development of a method to identify travel modes from accelerometry and GPS data
Methods
The method to identify travel modes from accelerometry and GPS data was developed using data collected at baseline in the ENABLE London study. Full details of this are published166 and are summarised below. Parts of this section are adapted or reproduced from Procter et al.166 This is an Open Access article distributed in accordance with the terms of the Creative Commons Attribution (CC BY 4.0) license, which permits others to distribute, remix, adapt and build upon this work, for commercial use, provided the original work is properly cited. See: http://creativecommons.org/licenses/by/4.0/. The text below includes minor additions and formatting changes to the original text.
Accelerometry has been widely used to measure physical activity, with many different devices deployed. The most common devices in the literature are the ActiGraph, Actiheart (CamNtech, Cambridge, UK), Actical (Philips Respironics, Amsterdam, the Netherlands), activPAL (PAL Technologies Ltd, Glasgow, UK) and GeneActiv (Activinsights Ltd, Huntingdon, UK), with over half of published studies up to 2015 using the ActiGraph.167 Many of these different devices convert raw acceleration (measured in g) into a form of activity count variable, which has then been used to classify physical activity intensity and energy expenditure.67,168–170 However, the methods used by instrument manufacturers to convert raw acceleration into counts are often unclear. For this reason, and because the raw acceleration data contain much more information to train an algorithm than derived count variables, we have made use of the raw data in this study.
Previous work on travel mode identification has been developed from the transport perspective rather than that of physical activity, where segmentation into journeys is important to assess travel behaviour.160,171–173 A focus on journeys often results in short periods of physically active transit behaviour, such as walking between bus stops, being identified as part of a non-active travel mode. However, for physical activity researchers, quantifying the volume and intensity of physical activity when actively travelling is an essential component of the overall purpose of a journey. Consequently, it is important to identify all data points that denote active travel, so that all time in active travel modes can be quantified. As a result, we identify the travel mode of each GPS data point (recorded every 10 seconds) without prior segmentation into journeys.
Participants in the ENABLE London study were asked to wear an accelerometer (ActiGraph GT3X+) and a GPS receiver on an elasticated belt around their waist for 7 consecutive days. Participants also completed a questionnaire to describe their travel patterns to work/place of study. They reported the specific days on which they would be travelling to work or study during the ActiGraph and GPS wear period and whether or not they commuted at the same time on each day. Reported travel modes for these journeys to and from work were underground (tube), train (over ground), bus/minibus/coach, taxi, motorcycle/moped, driving a car or a van, passenger in a car or a van, bicycle, walk, jog and other. These travel modes were recategorised into walk, cycle, vehicle (taxi, motorcycle, car/van driver, car/van passenger and bus/minibus/coach) and train (underground and overground rail). Insufficient participants consistently jogged to work for us to be able to separate ‘jog’ as an additional mode. In addition, time of leaving and arriving for each journey to and from work were collected.
Data preparation and cleaning
The raw accelerometer data were extracted as .csv files using ActiLife 6 software (ActiGraph, Florida, USA) summarised in 10-second epochs and merged by time stamp to GPS data. For the GPS data, measures of satellite signal quality were derived using the sum of the signal to noise ratio from each satellite the GPS device was connected to at each epoch. This takes into account possible obstructions to the GPS signal caused by the participant being indoors and gives a single measure of signal quality. Variations in speed and accelerometer data across a 10-second epoch were smoothed out using a 4-minute moving window calculating mean, SD, 10th and 90th percentiles of data from each accelerometer axis and speed from the GPS device. Distance from GPS time points to train lines was calculated using a combination of Meridian 2 (Ordnance Survey, Southampton, UK) rail network data for the UK and OS OpenMAP (Ordnance Survey, Southampton, UK) data for central London.174 Distance travelled over the previous and next minute were also calculated. All variables were chosen as they were likely to differ between travel modes: vehicles and trains have higher speed than walking and cycling; walking shows greater accelerometer activity than other modes; both vehicles and trains have metal structures that may obstruct GPS signal; distance over the previous or next minute would be low if the participant is stationary.
Development of the algorithm
A training data set was created using a subset of participants (n = 326) identified from the baseline data who commuted to and from work using the same mode of transport every day and specified the time that they usually travelled for both journeys. All combined accelerometer and GPS data were extracted and exported to a Geographical Information System [ArcGIS 10.4 (ESRI, CA, USA)], and all points during the commute of each participant were manually identified to confirm the stated mode of commute. Vehicle and train journeys were confirmed by location of appropriate features (e.g. roads or rail tracks). Points in the commute clustered around a single location were classified as ‘stationary’. It should be noted that ‘stationary’ does not imply inactive, as walking within a building will appear stationary in terms of GPS signal and be classified as such using our method. Any sequences of points for which a mode of travel could not be clearly identified were excluded, for example repeated GPS signal loss. If the participant appeared to be travelling by a method other than that stated, for example they usually commute by train but on one day they do not move along a train line, then these points were not assigned a travel mode and were excluded from the training data set.
The training data set was then used to develop a model. We tested different moving window sizes (i.e. 1-, 2-, 3-, 4-, and 5-minute moving windows) and assessed how accurate each was for predicting travel mode on a subset of the training data. To test the generalisability of our fitted model to other data sets, we compared manually identified data from 10 participants from the Sedentary Time and Metabolic Health in People with type-2 diabetes mellitus study (STAMP-2) (Professor Ashley R Cooper, University of Bristol, 2018, personal communication) data set with those predicted using our model. The STAMP-2 study was a cross-sectional observational study of sedentary behaviour in adults with newly diagnosed type 2 diabetes mellitus, conducted in two English NHS foundation trusts in south-west England. Participants in one centre wore an accelerometer and GPS monitor concurrently for 7 days, providing similar data to the ENABLE London study. The advantages of testing our model on the STAMP-2 participants are that they were independent of the ENABLE London study and recruited from a city with different travel options from that of central London. They were also recently diagnosed with type 2 diabetes mellitus, were generally older and had a higher prevalence of obesity than the ENABLE London participants. A good predictive performance on this data set would demonstrate the generalisability of our algorithm to other populations.
Results
Full results are available elsewhere;166 a summary is provided here. Parts of this section are adapted or reproduced from Procter et al.166 This is an Open Access article distributed in accordance with the terms of the Creative Commons Attribution (CC BY 4.0) license, which permits others to distribute, remix, adapt and build upon this work, for commercial use, provided the original work is properly cited. See: http://creativecommons.org/licenses/by/4.0/. The text below includes minor additions and formatting changes to the original text.
Training data
The training data set consisted of 326 participants providing 131,573 data points (365.5 hours): 66 (20%) who walked to work (12,791 points, 35.5 hours), 34 (10%) cyclists (11,607 points, 32.2 hours), 94 (29%) vehicle users (29,407 points, 81.7 hours), 132 (40%) train users (18,269 points, 50.7 hours) plus an additional 59,499 ‘stationary’ data points (165.3 hours). Approximately one-quarter of the training data (33,529 points) were used to test moving window sizes of 1, 2, 3, 4 and 5 minutes and the 4-minute moving window resulted in the highest predictive accuracy for active travel modes. The remaining 98,387 points were used to build a cross-validated model.
Model prediction
In the model cross-validation, we correctly predicted 97.3% of points and all five travel modes were predicted with high accuracy. In comparison with manually identified data (n = 21), overall accuracy was still high, with 96% of predictions correct, although this was driven by the fact that most people spent most of their time stationary (83.7% of time stationary). Cycling was least well predicted. When our model was tested on the STAMP-2 data set, we correctly predicted 96.5% of points compared with manual identification of travel modes, demonstrating that our predictions performed well considering the different participants and context. However, this was again driven by our high accuracy on stationary points, which was the dominant mode (86.8% of time stationary), and the poorest-performing predictive mode was, again, cycling.
Applying the algorithm to the ENABLE London study: baseline and 2-year follow-up findings
Methods
The algorithm developed on a subset of the ENABLE London baseline data set was applied to the full baseline and 2-year follow-up accelerometry and GPS data.
Outcomes
The primary outcome was minutes of ‘active’ travel, that is, daily minutes of walking and cycling. Secondary outcomes were vehicle, train, stationary and total minutes.
Statistical methods
The data collection and processing procedures were identical at baseline and at follow-up. Daily GPS data were processed to classify each 10-second epoch into one of five travel modes: walking, cycling, vehicle, train and stationary. Owing to the loss of GPS signal when travelling underground by tube, thus leading to gaps in the data, further work was carried out to classify some of these gaps as ‘underground’, when all of the following conditions were true:
- GPS signal was lost within 200 m of an underground station or had the predicted travel mode ‘train’.
- GPS signal was regained within 200 m of an underground station or had the predicted travel mode ‘train’.
- The time gap between loss and reacquisition of the signal was between 2 minutes and 2 hours.
- The loss and regaining of GPS signal did not take place within 200 m of the same underground station.
It should be noted that underground trains in the London transport system also run above ground, so there is potential for overlap between ‘underground’ and ‘train’ as methods of travel in our data.
The daily 10-second epoch data were summed to provide daily minutes for each travel mode and total daily GPS minutes. Walking and cycling minutes were also combined to provide a measure of ‘active travel’. This provided a data set of up to 7 days of data for each participant, with daily time spent in each travel mode. Daily data were processed to provide an average summary daily figure for each participant, using identical methods to those for the ENABLE London accelerometer data, as described in Chapter 3. Participants were asked to wear the accelerometer at the same time as the GPS monitor, and in line with accelerometry analyses, days where accelerometer wear time was less than 540 minutes (9 hours) were dropped from further analyses. For each mode of travel, average daily minutes at baseline were derived using multilevel linear regression models (level 1 was day within individual and level 2 was individual), regressing daily time on day order of wear, day of week and month of wear. This was repeated for the follow-up data. The change in average daily minutes from baseline to follow-up for each travel mode was then examined using multilevel linear regression models, where level 1 was individual and level 2 was household. For each travel mode, average daily minutes at follow-up was regressed on average daily minutes at baseline, East Village/control group, sex, age group and ethnic group as fixed effects and household as a random effect. The model coefficient for East Village/control group is thus the additional change in the East Village group adjusted for any change in the control group. This method of analysis measures change while controlling for regression to the mean, and has been shown to both minimise bias and maintain power. Checks were carried out to confirm that the distribution of residuals from the models were normally distributed. Sensitivity analyses included limiting to those who were working or studying at baseline. Imputation analyses for the main outcome of interest, ‘active travel’ (walking and cycling minutes), were carried out using multiple imputation methods using Stata Special Edition version 15 for Windows.
Results
The ENABLE London baseline cohort
At baseline, 1063 out of 1278 (83%) participants provided GPS data, of whom 991 provided ≥ 1 corresponding day with ≥ 540 minutes accelerometer wear time. Baseline characteristics for these 991 participants are shown in Table 15, overall and by housing sector. Across all the housing sectors, two-thirds were aged < 35 years, 56% were female and 52% were of white ethnic origin. Participants from the social housing sector were more likely to be female (74%), older and of black or Asian ethnic origin. Intermediate and market-rent participants were younger, equally likely to be male or female and more likely to be of white ethnic origin. These patterns are similar to those seen in the full ENABLE London cohort (see Chapter 2). Overall, the majority of the cohort (85%) were either working or studying at baseline, although this fell to two-thirds in the social housing sector and was higher (95%) in the intermediate and market-rent sectors. This is reflected in the proportions that were classified as economically inactive: 47% in the social sector and 7–9% in the intermediate and market-rent sectors. Three-quarters of the cohort used public transport to travel to or from their work or study and 50% reported some walking, cycling or jogging, which can be classified as ‘active travel’. The proportions of ‘active travel’ varied between the housing sectors: the social sector participants were less likely to walk, cycle or jog and more likely to use private transport. Table 15 also shows the time spent in different travel modes at baseline from the GPS monitors. Participants in the intermediate and market-rent sectors were broadly similar in their average daily time spent in each travel mode. In contrast, social sector participants recorded fewer walking, train and underground minutes, and higher vehicle and stationary minutes. Differences between housing sectors for time spent in different travel modes were statistically significant for all travel modes.
| Characteristic | All housing sectors (N = 991) | Housing sector | p-valuea | ||
|---|---|---|---|---|---|
| Social (N = 351; 35%) | Intermediate (N = 446; 45%) | Market-rent (N = 194; 20%) | |||
| Age (years), n (%) | |||||
| 16–24 | 209 (21) | 68 (19) | 82 (18) | 59 (30) | < 0.001 |
| 25–34 | 441 (45) | 91 (26) | 253 (57) | 97 (50) | |
| 35–49 | 275 (28) | 162 (46) | 93 (21) | 20 (10) | |
| ≥ 50 | 66 (7) | 30 (9) | 18 (4) | 18 (9) | |
| Sex, n (%) | |||||
| Female | 559 (56) | 260 (74) | 211 (47) | 88 (45) | < 0.001 |
| Ethnic group, n (%) | |||||
| White | 516 (52) | 70 (20) | 312 (70) | 134 (69) | < 0.001 |
| Black | 232 (23) | 174 (50) | 44 (10) | 14 (7) | |
| Asian | 151 (15) | 69 (20) | 60 (13) | 22 (11) | |
| Other | 92 (9) | 38 (11) | 30 (7) | 24 (12) | |
| NS-SEC, n (%) | |||||
| Higher managerial/professional | 498 (51) | 47 (14) | 318 (72) | 133 (69) | < 0.001 |
| Intermediate occupations | 154 (16) | 49 (14) | 70 (16) | 35 (18) | |
| Routine/manual occupations | 120 (12) | 87 (25) | 25 (6) | 8 (4) | |
| Economically inactive | 211 (21) | 163 (47) | 30 (7) | 18 (9) | |
| Car or van in household | 357 (46) | 132 (50) | 159 (44) | 66 (41) | 0.14 |
| Working or studying | 839 (85) | 232 (66) | 423 (95) | 184 (95) | < 0.001 |
| Work or study away from home | 773 (92) | 213 (92) | 397 (94) | 163 (89) | 0.08 |
| Mode of travel between home and work (modes not mutually exclusive), n (%) | |||||
| Public transport | 590 (73) | 155 (69) | 312 (76) | 123 (69) | 0.09 |
| Private car/motorbike/taxi | 113 (14) | 49 (22) | 48 (12) | 16 (9) | < 0.001 |
| Walk/cycle/jog | 439 (54) | 95 (42) | 246 (60) | 98 (55) | < 0.001 |
| GPS variables (minutes), mean (SD) | p-valueb | ||||
| Walking | 38 (25) | 32 (26) | 41 (22) | 43 (28) | < 0.001 |
| Cycling | 5 (10) | 3 (6) | 6 (11) | 6 (14) | < 0.001 |
| Walking and cycling | 43 (28) | 35 (27) | 47 (25) | 49 (31) | < 0.001 |
| Vehicle | 37 (38) | 46 (47) | 32 (29) | 30 (33) | < 0.001 |
| Train | 14 (20) | 8 (13) | 18 (22) | 15 (24) | < 0.001 |
| Underground | 15 (18) | 8 (15) | 19 (18) | 17 (17) | < 0.001 |
| Stationary | 446 (194) | 500 (179) | 422 (192) | 403 (205) | < 0.001 |
| Total GPS minutes | 554 (206) | 597 (185) | 538 (206) | 515 (228) | < 0.001 |
Changes in mode of travel baseline to follow-up
At follow-up, 714 out of 877 (81%) participants provided GPS data, 681 with ≥ 1 corresponding day with ≥ 540 minutes of accelerometry wear time. Longitudinal analyses were restricted to 578 participants who had valid GPS data at both baseline and follow-up.
Baseline characteristics for the 578 adults with GPS data at baseline and follow-up are shown in Table 16 by East Village/control group and housing sector. Reassuringly, age, sex, ethnicity and NS-SEC patterns were similar to those for the full 877 who were followed up (see Chapter 2). In the social housing sector, the East Village and control group were similar in age, sex and socioeconomic distributions, but the East Village group were more likely to be of black ethnic origin. In the intermediate sector, the East Village group were more likely to be younger, male, of white ethnic origin and economically active. In the market-rent sector, age, sex, ethnic group and socioeconomic status were similar in the East Village and control groups. There were no differences between the East Village group and the control group in the proportions of households who owned a car at baseline. Slightly fewer of the East Village group were working at baseline and 24% were classified as economically inactive compared with 18% of the control group. At baseline, there was greater use of public transport for travel to work or study among those who moved into East Village (p = 0.004 for all housing sectors). Use of private transport and walk/cycle/jog were similar in the East Village and control groups, although social housing sector participants were more likely to use private transport and less likely to walk/cycle/jog than intermediate and market-rent participants.
| Characteristic | All housing sectors (N = 578) | p-value | Housing sector | ||||||
|---|---|---|---|---|---|---|---|---|---|
| Social (N = 201) | Intermediate (N = 283) | Market-rent (N = 94) | |||||||
| Control (n = 285; 49%) | East Village (n = 293; 51%) | Control (n = 74; 37%) | East Village (n = 127; 63%) | Control (n = 141; 50%) | East Village (n = 142; 50%) | Control (n = 70; 74%) | East Village (n = 24; 26%) | ||
| Age (years), n (%) | |||||||||
| 16–24 | 44 (15) | 68 (23) | 0.005 | 7 (9) | 26 (20) | 20 (14) | 29 (20) | 17 (24) | 13 (54) |
| 25–34 | 116 (41) | 133 (45) | 18 (24) | 34 (27) | 65 (46) | 93 (65) | 33 (47) | 6 (25) | |
| 35–49 | 100 (35) | 80 (27) | 44 (59) | 57 (45) | 46 (33) | 19 (13) | 10 (14) | 4 (17) | |
| ≥ 50 | 25 (9) | 12 (4) | 5 (7) | 10 (8) | 10 (7) | 1 (1) | 10 (14) | 1 (4) | |
| Sex, n (%) | |||||||||
| Female | 173 (61) | 159 (54) | 0.12 | 58 (78) | 91 (72) | 82 (58) | 57 (40) | 33 (47) | 11 (46) |
| Ethnic group, n (%) | |||||||||
| White | 152 (53) | 161 (55) | 0.001 | 17 (23) | 24 (19) | 87 (62) | 117 (82) | 48 (69) | 20 (83) |
| Black | 48 (17) | 78 (27) | 22 (30) | 75 (59) | 21 (15) | 3 (2) | 5 (7) | 0 (0) | |
| Asian | 57 (20) | 30 (10) | 29 (39) | 13 (10) | 23 (16) | 15 (11) | 5 (7) | 2 (8) | |
| Other | 28 (10) | 24 (8) | 6 (8) | 15 (12) | 10 (7) | 7 (5) | 12 (17) | 2 (8) | |
| NS-SEC, n (%) | |||||||||
| Higher managerial/professional | 163 (57) | 130 (45) | 0.03 | 17 (23) | 16 (13) | 97 (69) | 99 (70) | 163 (70) | 15 (63) |
| Intermediate occupations | 43 (15) | 56 (19) | 12 (16) | 21 (17) | 18 (13) | 29 (21) | 13 (19) | 6 (25) | |
| Routine/manual occupations | 29 (10) | 33 (11) | 16 (22) | 24 (19) | 11 (8) | 9 (6) | 2 (3) | 0 (0) | |
| Economically inactive | 50 (18) | 71 (24) | 29 (39) | 64 (51) | 15 (11) | 4 (3) | 6 (9) | 3 (13) | |
| Car or van in household | 111 (46) | 104 (46) | 0.96 | 31 (53) | 47 (51) | 55 (46) | 49 (43) | 25 (41) | 8 (40) |
| Working or studying | 245 (86) | 240 (82) | 51 (69) | 80 (63) | 128 (91) | 138 (97) | 66 (94) | 22 (92) | |
| Work or study away from home | 217 (89) | 226 (94) | 0.04 | 44 (86) | 74 (92) | 118 (92) | 132 (96) | 55 (83) | 20 (89) |
| Mode of travel to or from work or study (not mutually exclusive), n (%) | |||||||||
| Public transport | 156 (67) | 182 (79) | 0.004 | 31 (66) | 56 (74) | 83 (67) | 110 (82) | 42 (67) | 16 (76) |
| Private car/motorbike/taxi | 35 (15) | 28 (12) | 0.36 | 10 (21) | 16 (21) | 19 (15) | 10 (7) | 6 (10) | 2 (10) |
| Walk/cycle/jog | 141 (61) | 128 (55) | 0.27 | 19 (40) | 40 (53) | 82 (67) | 77 (57) | 40 (63) | 11 (52) |
| GPS variables (minutes), mean (SD) | |||||||||
| Walking | 40 (26) | 38 (24) | 0.003 | 31 (21) | 31 (25) | 41 (24) | 43 (22) | 47 (31) | 38 (28) |
| Cycle | 6 (13) | 4 (7) | 0.06 | 2 (3) | 2 (4) | 7 (14) | 5 (9) | 8 (18) | 2 (4) |
| Walking and cycling | 46 (30) | 41 (26) | 0.84 | 33 (22) | 34 (26) | 48 (28) | 48 (25) | 55 (36) | 40 (28) |
| Vehicle | 37 (38) | 38 (39) | 0.98 | 43 (47) | 49 (48) | 36 (32) | 29 (28) | 34 (38) | 32 (30) |
| Train | 15 (21) | 15 (19) | 0.55 | 8 (12) | 9 (14) | 16 (23) | 20 (22) | 17 (25) | 12 (12) |
| Underground | 14 (17) | 15 (18) | 0.58 | 12 (20) | 8 (15) | 15 (17) | 21 (18) | 16 (16) | 19 (15) |
| Stationary | 440 (193) | 449 (188) | 0.73 | 507 (192) | 496 (184) | 421 (190) | 418 (179) | 407 (184) | 380 (207) |
| Total GPS | 552 (207) | 558 (195) | 0.02 | 602 (201) | 596 (190) | 537 (206) | 537 (188) | 529 (210) | 482 (225) |
Table 17 shows the change in the neighbourhood perceptions and built environment variables for the East Village and control groups. Compared with the baseline data, those participants who had moved to East Village lived closer to their nearest park, had better access to public transport and lived in a more walkable area. They also reported more positive perceptions of their local area. These differences were most marked for social housing sector participants.
| Measure | All housing sectors, mean (95% CI); p-value | Housing sector, mean (95% CI); p-value | ||
|---|---|---|---|---|
| Social | Intermediate | Market-rent | ||
| Built environment characteristics | ||||
| Control group | ||||
| Total participants (n) | 256 | 73 | 127 | 56 |
| Distance to closest park (m)a | 28 (–23 to 78);0.28 | 27 (–44 to 97);0.45 | 31 (–39 to 101);0.39 | 21 (–123 to 166);0.77 |
| Access to public transport (PTAL)b | –0.1 (–0.3 to 0.1);0.35 | –0.2 (–0.5 to 0.1);0.20 | –0.1 (–0.4 to 0.3);0.63 | 0.0 (–0.5 to 0.4);0.94 |
| Walkabilityc | 0.3 (0.0 to 0.6);0.09 | –0.2 (–0.7 to 0.3);0.38 | 0.5 (0.0 to 1.0);0.04 | 0.3 (–0.2 to 0.9);0.25 |
| Land use mixd | 0.02 (0.00 to 0.04);0.08 | –0.02 (–0.05 to 0.01);0.22 | 0.04 (0.01 to 0.08);0.03 | 0.02 (–0.02 to 0.06);0.25 |
| Residential densitye | 1.8 (1.0 to 2.7);< 0.001 | 1.2 (0.2 to 2.3);0.02 | 2.1 (0.7 to 3.5);0.003 | 2.0 (0.2 to 3.8);0.03 |
| Street connectivityf | –0.1 (–0.2 to 0.1);0.38 | –0.2 (–0.4 to 0.0);0.13 | 0.0 (–0.2 to 0.2);0.93 | –0.1 (–0.4 to 0.2);0.63 |
| East Village group | ||||
| Total participants (n) | 277 | 124 | 130 | 23 |
| Distance to closest park (m)a | –547 (–597 to –498);< 0.001 | –463 (–529 to –397);< 0.001 | –602 (–673 to –530);< 0.001 | –693 (–952 to –435);< 0.001 |
| Access to public transport (PTAL)b | 1.2 (0.9 to 1.5);< 0.001 | 2.0 (1.5 to 2.5);< 0.001 | 0.7 (0.2 to 1.1);0.007 | 0.3 (–0.9 to 1.5);0.61 |
| Walkabilityc | 2.4 (2.1 to 2.7);< 0.001 | 2.7 (2.3 to 3.1);< 0.001 | 2.0 (1.5 to 2.5);< 0.001 | 2.4 (0.4 to 4.3);0.02 |
| Land use mixd | 0.38 (0.36 to 0.41);< 0.001 | 0.40 (0.37 to 0.43);< 0.001 | 0.37 (0.34 to 0.41);< 0.001 | 0.35 (0.23 to 0.47);< 0.001 |
| Residential densitye | 14.4 (12.8 to 16.0);< 0.001 | 14.7 (12.3 to 17.0);< 0.001 | 13.3 (11.0 to 15.6);< 0.001 | 19.1 (12.4 to 25.9);< 0.001 |
| Street connectivityf | –1.1 (–1.2 to –0.9);< 0.001 | –0.9 (–1.1 to –0.7);< 0.001 | –1.2 (–1.4 to –1.0);< 0.001 | –1.2 (–2.0 to –0.4);0.006 |
| Neighbourhood perceptionsg | ||||
| Control group | ||||
| Total participants (n) | 285 | 74 | 141 | 70 |
| Crime score | 0.6 (0.1 to 1.0);0.02 | 1.3 (0.4 to 2.2);0.005 | 0.0 (–0.7 to 0.7);1.00 | 1.0 (0.0 to 1.9);0.04 |
| Quality score | 0.7 (0.2 to 1.2);0.01 | 1.5 (0.6 to 2.4);0.001 | 0.1 (–0.6 to 0.9);0.73 | 0.9 (0.0 to 1.9);0.06 |
| East Village group | ||||
| Total participants (n) | 293 | 127 | 142 | 24 |
| Crime score | 4.7 (4.1 to 5.2);< 0.001 | 6.1 (5.2 to 7.1);< 0.001 | 3.5 (2.8 to 4.2);< 0.001 | 3.8 (1.7 to 5.9);0.001 |
| Quality score | 7.0 (6.4 to 7.5);< 0.001 | 7.9 (7.0 to 8.7);< 0.001 | 6.3 (5.5 to 7.1);< 0.001 | 6.5 (5.1 to 7.9);< 0.001 |
The effect of moving to East Village on the time spent in different travel modes is shown in Table 18. Overall, there was little change in participants’ walking or cycling minutes. However, vehicle minutes decreased overall by 7–8 minutes per day, with greater effects in the intermediate housing sector (10 minutes’ decrease, 95% CI –17 to –2 minutes’ decrease; p = 0.01), and time spent travelling by underground increased by 3–4 minutes, particularly in the market-rent sector (12 minutes, 95% CI 4 to 19 minutes; p = 0.001). There were large decreases in both stationary and total minutes; however, this was as a result of issues with the GPS signal being blocked by East Village housing, discovered after follow-up had been completed.
| Change | All housing sectors (N = 578), effect (95% CI); p-value | Housing sector, effect (95% CI); p-value | ||
|---|---|---|---|---|
| Social (N = 201) | Intermediate (N = 283) | Market-rent (N = 94) | ||
| Daily walking minutes | ||||
| Model 1a | –1.3 (–4.9 to 2.3); 0.49 | –0.3 (–6.4 to 5.8); 0.92 | 0.1 (–5.1 to 5.3); 0.97 | –4.9 (–15.5 to 5.7); 0.36 |
| Model 2b | –1.7 (–5.4 to 2.0); 0.36 | –1.8 (–8.4 to 4.8); 0.59 | –1.9 (–7.5 to 3.8); 0.52 | –4.2 (–15.4 to 7.0); 0.46 |
| Model 3c | –1.4 (–5.3 to 2.5); 0.48 | |||
| Daily cycling minutes | ||||
| Model 1a | 0.7 (–0.8 to 2.2); 0.34 | 0.5 (–0.5 to 1.5); 0.37 | 1.6 (–0.9 to 4.1); 0.21 | 1.9 (–3.3 to 7.0); 0.48 |
| Model 2b | 0.7 (–0.8 to 2.2); 0.36 | 0.3 (–0.8 to 1.4); 0.58 | 1.0 (–1.6 to 3.7); 0.44 | 1.9 (–3.6 to 7.4); 0.50 |
| Model 3c | 1.1 (–0.5 to 2.7); 0.17 | |||
| Daily walk plus cycle minutes | ||||
| Model 1a | –0.7 (–4.7 to 3.3); 0.75 | 0.1 (–6.2 to 6.4); 0.98 | 1.6 (–4.2 to 7.3); 0.59 | –2.2 (–15.4 to 11.0); 0.74 |
| Model 2b | –1.1 (–5.2 to 2.9); 0.59 | –1.5 (–8.4 to 5.4); 0.68 | –0.9 (–7.1 to 5.4); 0.78 | –2.4 (–16.0 to 11.3); 0.74 |
| Model 3c | –0.4 (–4.7 to 3.8); 0.85 | |||
| Vehicle minutes | ||||
| Model 1a | –7.6 (–13.0 to –2.2); 0.01 | –7.2 (–18.9 to 4.5); 0.23 | –11.4 (–18.2 to –4.7); < 0.001 | –5.9 (–18.2 to 6.4); 0.35 |
| Model 2b | –7.0 (–12.5 to –1.5); 0.01 | –6.2 (–18.9 to 6.6); 0.34 | –9.6 (–16.9 to –2.2); 0.01 | –6.9 (–19.9 to 6.1); 0.30 |
| Model 3c | –8.3 (–14.0 to –2.5); 0.01 | |||
| Train minutes | ||||
| Model 1a | –1.4 (–4.2 to 1.4); 0.32 | –0.1 (–3.9 to 3.7); 0.95 | –1.3 (–5.4 to 2.8); 0.53 | 2.1 (–7.5 to 11.7); 0.67 |
| Model 2b | –1.7 (–4.6 to 1.1); 0.23 | –1.7 (–5.8 to 2.4); 0.41 | –1.5 (–6.0 to 3.0); 0.51 | 0.2 (–10.1 to 10.6); 0.97 |
| Model 3c | –0.8 (–3.7 to 2.2); 0.62 | |||
| Underground minutes | ||||
| Model 1a | 2.6 (0.0 to 5.1); 0.05 | 2.9 (–0.9 to 6.6); 0.14 | 2.2 (–1.8 to 6.2); 0.27 | 10.9 (4.1 to 17.6); 0.002 |
| Model 2b | 2.6 (0.1 to 5.2); 0.04 | 4.0 (–0.1 to 8.1); 0.06 | 1.6 (–2.7 to 6.0); 0.46 | 11.5 (4.4 to 18.6); 0.001 |
| Model 3c | 3.9 (1.2 to 6.5); 0.005 | |||
| Stationary minutes | ||||
| Model 1a | –248 (–278 to –218); < 0.001 | –346 (–393 to –299); < 0.001 | –199 (–243 to –155); < 0.001 | –143 (–229 to –56); 0.001 |
| Model 2b | –248 (–279 to –217); < 0.001 | –354 (–405 to –304); < 0.001 | –205 (–252 to –157); < 0.001 | –127 (–217 to –38); 0.01 |
| Model 3c | –247 (–279 to –214); < 0.001 | |||
| Total minutes | ||||
| Model 1a | –256 (–289 to –223); < 0.001 | –347 (–400 to –295); < 0.001 | –211 (–259 to –162); < 0.001 | –139 (–234 to –44); 0.004 |
| Model 2b | –255 (–289 to –222); < 0.001 | –357 (–412 to –302); < 0.001 | –217 (–269 to –164); < 0.001 | –126 (–223 to –29); 0.01 |
| Model 3c | –253 (–288 to –217); < 0.001 | |||
Discussion
Using the ENABLE London study accelerometry and GPS data, we have developed an accurate predictive algorithm that identifies five travel modes, including the active travel modes walking and cycling, and identifies each travel mode correctly > 90% of the time. Our levels of accuracy in cross-validation outperform recent similar studies.160,162 The models developed here are freely available to apply to similar data in the statistical software environment R.165 We have applied this algorithm to the complete baseline and 2-year follow-up ENABLE London data sets to characterise the travel patterns of a large group of adults in London, identifying time spent in different modes of travel. Furthermore, we have examined changes in time spent in different travel modes following a move to East Village. Despite the built environment variables showing that East Village is a more walkable area, we found no change in the time spent walking or cycling compared with those living elsewhere. However, there was some suggestion that vehicle travel had decreased, particularly in the intermediate housing sector, and that underground travel had increased, more so in the market-rent sector.
Strengths of the algorithm developed
Our study has a number of important strengths. Applying the algorithm to the STAMP-2 data set resulted in similar levels of accuracy of travel mode prediction within a very different group of participants living in a different environmental setting. This finding suggests that our method may be generalisable to other data sets and could be used by other researchers without the time-consuming steps of creating new training data. However, until more robust tests have been completed, we would recommend a manual identification checking stage similar to our methods to verify the generalisability of the method.
We exhibit similar levels of accuracy to the personal activity location measurement system (PALMS),173 which is a freely available method to process physical activity data. The purpose of the output is somewhat different though, with PALMS identifying journeys and our method identifying each data point. The preferred method will depend on the research question. One advantage of our R package that the server-based system in PALMS does not have is that it can be run on a researcher’s own machine. Data used for the present (and other similar) analyses are subject to strict data privacy and ethical conditions. Running the analyses on a researcher’s own machine, rather than uploading to a server for remote processing, can help avoid problems related to data privacy. Furthermore, our method is open source, which means that all code is freely available online.165 As a result, other researchers can suggest edits and improvements and contribute to the development of our method, and its utility to the research community.
The prediction of active travel modes should complement existing analyses of physical activity using accelerometers. Accelerometers have been used with great effect to objectively quantify activity, but they are not without limitations. A well-known problem of traditional physical activity analyses using accelerometers is that cycling is not recognised as a form of MVPA because cycling generates relatively low readings on a waist-worn accelerometer compared with other active modes. Identification of cycling from combined accelerometer and GPS data will allow better quantification of cycling as a form of physical activity. Traditional physical activity analyses will still miss cycling as a form of activity whereas our method will quantify it, albeit with moderate precision. Furthermore, assessment of active travel using our method will help to understand how people are active. For example, activity at a single location, such as at home or at the gym, will not be classed as active travel using our method but as a stationary travel mode; therefore, a participant may show high levels of overall physical activity but low levels of active travel. Incorporation of this extra information will help to understand participants’ overall physical activity patterns. For example, if a participant shows increased physical activity but not active travel, it is likely that they have increased their activity at locations such as at home or at the gym. If we see no change in physical activity but an increase in active travel, then participants have replaced some of their physical activity at a location with active travel.
Limitations of the algorithm developed
It is unsurprising that our accuracy scores are lower when compared with manual identification of both participants within this study and from elsewhere than in cross-validation of the training data set. Identification is most likely to be accurate during journeys of a consistent mode, which is how the training data were defined. In full days of data there are likely to be other more ambiguous forms of movement that may be short in duration and, therefore, difficult to identify, or which do not fully represent one of the forms of travel we have included. Our model performs least well for the detection of cycling and, therefore, is probably overfitted to our training data. As a result, we would recommend manual checking of cycling data to improve accuracy. In this study, cycling represents a relatively small amount of total time; therefore, manual checking of this subset represents a much smaller time investment than in the full analysis.
Visual inspection of the data in the GIS revealed that much of the disagreement between prediction and manual classification was found at the start and the end of journeys. This highlights the challenge of identifying modal shift, that is when to switch from one mode to another. This is a limitation of the current method and all other indirect methods. However, the strength of identifying individual data points rather than full journeys is that the imprecision in identifying modal shift leads to small numbers of misclassified points, rather than entire misclassified journeys. Small numbers of misclassified points will only have a small effect on total time in each travel mode. Our inaccuracy at the start and end of trips means that any prediction to a data set where many short-duration trips are expected would be predicted to yield lower accuracy. Conversely, within a data set containing longer journeys, travel mode should be able to be identified with greater precision.
It is also worth highlighting that some disagreement within our test data sets may not be true errors. For example, if a participant is stationary but on a train it is questionable whether they should be classified as using a train or stationary; this may not matter as long as the rule is applied consistently. However, when comparing the predicted and manually derived methods this causes disagreement because during manual classification stationary points on train lines were termed stationary (e.g. standing at a train station), yet the algorithm identifies them as ‘train’. A number of the misclassifications between our manual identification and predicted data sets may, therefore, be open to interpretation and may not be true misclassifications.
Using the model
First, we were only able to identify the active modes of walking and cycling: there is no consideration of running or any other activity. This limitation is based on our study sample, where no commuters consistently used these modes so we could not include them in our training data. However, walking and cycling represent the most frequently used active transport modes and other modes were rarely reported by our participants. Any form of running will most likely be identified as walking using our method because of the high acceleration that running causes on an accelerometer, and so will still be identified as active travel. We therefore feel that this limitation will have little impact on our study, although we would recommend caution in applying our method to a data set where a large quantity of running is expected. Second, we have demonstrated our capacity to identify major travel modes but were not able to discriminate car travel from public transport. Consequently, if a study is attempting to quantify the use of public transport then the current method is inadequate. To address this, we have developed a second model that discriminates bus travel. However, this leads to reduced overall accuracy, because the pattern of speed/activity can be confused between buses and other vehicles. Further work in this area would need to assess the generalisability of our methods. Third, although we assess different geographical contexts, both are in the UK. We do not know how well the method would perform elsewhere in the world, where other transport options may be available. Furthermore, we have only tested the current model on adults, and further research could assess how well our model performs on older adults or children, to potentially be of use in a wider group of studies.
Applying the model to the ENABLE London baseline and follow-up data
Using the algorithms developed on a subset of the baseline ENABLE London GPS and accelerometry data sets, we have been able to characterise the travel use of a large group of adults in London and identify the time spent in different modes of travel. Furthermore, we have examined changes in the time spent in different travel modes following a move to East Village. However, despite the built environment variables showing that East Village is a more walkable area, we found no change in the time spent walking or cycling compared with those living elsewhere. However, there was some suggestion that vehicle travel had decreased, particularly in the intermediate housing sector, and that underground travel had increased, more so in the market-rent sector.
Strengths of applying the model to the ENABLE London data
The East Village neighbourhood was designed to encourage active living. Restrictions are placed on parking facilities in East Village but it has good public transport links, being close to Stratford station with both overground links and underground links. Half of the cohort were living in East Village at 2-year follow-up and half had either stayed at their baseline address or moved elsewhere. Following up the same individuals 2 years later enabled us to examine change in behaviour at an individual level with participants acting as their own controls. The GPS monitors were worn at the same time as the accelerometers, which provided direct and time-stamped comparisons of physical activity and location. GPS monitors rely on having good signal links with satellite receivers and the signals can be blocked by some buildings and when travelling underground, but the concurrent use of accelerometry data helped to identify different modes of travel and daily wear time. The eligibility criteria for moving into East Village, set externally by the housing association and companies, meant that we could examine differences between three housing sectors: social, intermediate and market-rent. Eligibility for these different housing sectors was based on income, which provided a measure of socioeconomic status. Finally, our method of analysis, regressing follow-up-values on baseline values, allows direct comparison of change in outcomes between the East Village and the control groups and minimised the issues of regression to the mean.
Limitations
Lack of adequate wear time led to a reduced sample of 548 participants with GPS data at both baseline and follow-up. However, baseline demographic characteristics were similar to the full cohort. After data collection had been completed and the GPS data were being processed, it was discovered that the East Village housing blocked the transmission of the GPS signal leading to low GPS-recorded time in the East Village group. However, as the accelerometers were worn at the same time as the GPS monitors and provided an estimate of wear time, we were able to use this as a proxy for wear time for the GPS monitor. Additionally, the decrease in total GPS minutes was almost identical to the decrease in stationary time in the East Village group, confirming that the GPS time lost was most likely to be spent stationary by participants in their East Village homes.
Conclusions
In summary, we have developed a method to identify travel modes from accelerometer, GPS and GIS data for the ENABLE London study, which successfully predicts > 90% of points tested in a range of contexts. This method can be of use to complement existing analyses of physical activity data and assess active travel alongside physical activity. All code necessary to replicate the analysis, apply the method to other data sets or predict from our models to other data sets is provided to facilitate usage by other researchers.