Investigating heart rate responses of children to active travel: a mixed effects modeling analysis – BMC Public Health

In this study, we utilized a multiday mixed survey with physiological measurements following the approach proposed by Stark et al. [2] to gather data on the physiological responses of children in Austria in the context of their everyday mobility. Subsequently, we employed mixed effects models for exploratory analysis to investigate the impacts of various factors on these physiological responses.

Survey process and participants

Our sample includes 73 children in the 6th and 7th grades (average age of 13) from three secondary schools of a comparable type, but differently located (A, B, and C). Schools A and B are located in the densely built 19th district of Vienna and they are very well accessible with metro, tram and bus services. School C is in Korneuburg, a small town north of Vienna in Lower Austria with 14,000 inhabitants, but it has a larger catchment area. The schools were contacted during the application phase of the project in which this study is embedded. Together with the heads of the schools, cooperation classes were defined in the desired age groups. It was not possible to extend the sample to include additional school classes; therefore, there is no claim to representativeness.

The data collection took place between April and May 2023. The weather during this period was typically spring-like for this time of year in Austria. There were no anomalies in terms of precipitation or temperature. The school children participated in an online-survey for seven consecutive days. The children were enrolled in a workshop by the researchers for the online questionnaire, which was developed in cooperation with the children to ensure comprehensibility and practicability for this target group. The children were supported in completing the questionnaire throughout the survey period, e.g., in finding the right addresses of trip origins and destinations.

During the week of the survey, the children wore a PA tracker. In a workshop, the activity tracker was explained and the children were instructed to wear the tracker as continuously as possible (including during possible training sessions and bedtime). They were asked not to change any functional settings of the tracker. To avoid burdening the children with charging the device, the researchers took care of the charging process. For this purpose, the tracker was charged twice during the survey week at school. A data backup was also made during this process. Each tracker was given a unique ID and assigned to a child for the study period. At the end of the study period, anonymization was performed for data protection reasons by deleting the link between the activity tracker ID and the child according to EU General Data Protection Regulation [34]. Comprehensive permits had to be obtained for all activities within the project in which this study was embedded. This refers to authorizations from the directorate of education and from the parents. In addition, all survey processes in the study have been approved by the ethics committee.

Measurement design

Survey questionnaire

The basic structure of the questionnaire was composed of two parts: (1) socio-demographic and household information; (2) travel diary of seven reporting days. The reporting days information included trips, trip stages (travel diary), and PA at the destination as well as well-being on that day. Only the parts of the questionnaire that are relevant to the present study are explained below.

On social demographic level, the questionnaire was divided into three sections: (i) me and my home, (ii) my travel modes, and (iii) my health. In the first section, variables were collected such as age, gender, number of persons in household (adults/children), and living situation [apartment, single-family/row house, other]. In the next section, we collected information on the number of cars in the household, the availability of a bicycle, scooter, e-scooter/e-bicycle, discount ticket for public transit, walking minutes to the next public transit stop from home. In the third section (my health), we asked questions on the general health status, bodily fitness and collected self-assessments on the children’s exercise behavior.

The travel diary was structured like this: Per trip, we collected start and end time of local UTC-timestamp (including year, date, hour and minutes) as well as origin/destination addresses. Further, the mode choice [walking, bicycle, scooter, e-scooter/e-bike, public transport, in a car, other] and the duration of each trip stage were queried, as well as the accompaniment [I was alone, friends/siblings, parents/adults, I accompanied someone (e.g. younger siblings)]. For each trip, respondents were asked who made the mode choice [me alone, parents, adults, joint decision] and what the weather was like [rain/snow, windy, sunny, warm, cold, multiple answers possible]. We also asked how the children felt along the way of their subjective perception and how physically demanding the trip was: In doing so, multi-part scales were used, which were recoded to binary scales during model development [stressed/annoyed, unhappy, anxious/uncomfortable, exhausted]. We collected information on the trip purpose [was at home, school, sports (e.g. soccer, dancing, swimming), shopping/café/cinema, was out: chilling, walking around, was e.g. with friends/relatives, errands (e.g. doctor, shopping), other]. During data input, all trip origins and destinations were encoded with GPS coordinates to allow geo-referenced analysis and getting the information of land use (density). Undirected trips, such as loop walks and leisurely strolls, were also included in the analysis. However, location-based activities like horse riding lessons, soccer training, and similar sports were classified as sports activities rather than trips.

Activity tracker

In selecting activity tracker watches for the study, we prioritized models with strong battery life, simple functionality, and lightweight design to ensure continuous wear and minimize participant interference during the extended survey period. This approach aimed to reduce the need for frequent charging and mitigate potential influences on data collection caused by the allure of advanced features, especially among children. Based on previous studies [35, 36], that demonstrated the reliability of collecting heart rate and activity data, wrist-mounted Vivosmart 3 and 4 trackers were selected for this study. Both models offer same functionalities for accurate data collection.

Before the beginning of survey, the models were tested to check its suitability for everyday use and for children. Each tracker consists of a small measuring instrument integrated into a robust rubber wristband. The battery life of trackers are up to five days, and data can be stored for up to three weeks in embedded memory card. The device is certified waterproof, which records the average heart rate and walk steps each 1 min.

Dataset and variables

Dataset of repeated-measures establishment

The survey and the objective sensor data of physiological measurement in this study were based on repeated observations of each participant over time, making it suitable for the repeated-measures data pattern [37]. Given the goals of this research, it was crucial to establish a dataset with a repeated-measures structure. To accomplish this, we followed a sequential data merging process. Initially, a new data frame containing timestamps with a 1-minute interval was generated for each survey day, covering a 24-hour period. Subsequently, this data frame was merged with the raw tracker data, utilizing local UTC-Timestamp as the common identifier. In a similar manner, we incorporated raw data from online questionnaires by matching participant IDs and local UTC-Timestamps.

Dataset preprocessing

During the course of the survey, questionnaire data of each participant was checked and error and missing values were directly marked through online survey system every day by the research team to ensure the data accuracy and validity of the study’s outcomes. Unfortunately, due to technical constraints, the review of the tracker data could only be conducted at the end of the survey. To avoid disturbing continuous measurements of participants, access to the data during the survey period was not feasible. Consequently, main efforts were dedicated to the preprocessing of the raw heart rate data, where two primary issues were identified. Firstly, errors in the tracker system resulted in the presence of outliers. Secondly, some participants failed to appropriately wear the trackers or forgot to wear them during the survey period, leading to some missing data. To ensure data accuracy and reliability, we initially addressed the outlier problem, mainly focusing on the heart rate data. As specific heart rate ranges for children during activities were unavailable in existing literature, we filtered out obvious erroneous data points, such as zero values, and temporarily labeled them as missing values while retaining other data points for further investigation. Subsequently, we excluded trips with a data missing rate exceeding 80% from the analysis. Finally, for the remaining missing values, we followed a principle of using the nearest neighbor estimation approach to filling missing values for trips with one or two consecutive missing data points. However, if a trip had two or more consecutive missing data points, we chose to delete the entire trip as it was not possible to infer the pattern of heart rate data changes. Following this statistical process, approximately 2.5% of trips were removed from the dataset.

It is also worth noting that two participants were not motivated to continue their participation and data collection prematurely. Therefore, all the collected data from them were not considered in the further analysis. Additionally, we also found another participant wore the watch for very little time in all survey days; thus, this person was deleted from the sample. Finally, this resulted in a total of 1,146 trips with 28,206 records data by 70 participants, with each record representing one-minute data.

Variables selection

In this study, heart rate serves as an objective indicator to represent both PA intensity and emotional responses from psychological perspective, which is treated as the outcome variable. Our main objective was to systematically analyze the relationship between active travel and children’s physiological responses at trip level.

• First, we processed the mode choice variable in each trip stage into the categories “active” and “non-active” travel modes, according to the taxonomy concept by Cook et al. [38]: We classified walking, bicycle, kick-scooter, and power-assisted e-bike as “active” (without a specific intensity), and car as well as public transport (e.g. bus, tram) as “non-active”. Mode choice in the category “other” has been classified as “non-active” according to the existing trip characteristics (trip purpose, origin, destination, and duration) (this refers to nine cases).

• Second, we calculated the duration of “active” and “non-active” travel modes per trip.

• Third, based on the duration, we introduced our primary exposure variable “Trip Active Level”; it consists of four levels: (i) Level 1 represents a trip with the lowest duration of “active” as no active travel mode was used; (ii) Level 2 represents a trip with a low duration of “active”, in which the “dominant” travel mode(s) were “non-active”, but some “active” was still included. Here, “dominant” means the maximum share of total trip duration; (iii) Level 3 corresponds medium, where “active” was “dominant” (in terms of duration) in the trip, but still “non-active” was used; and (iv) Level 4 indicates high active. This means all travel modes used in the trip were “active”.

• Fourth, to enable linear regression modeling, the “Trip Active Level” was further transformed into four dummy variables as “Trip active level 1 (None)”, “Trip active level 2 (Low)”, “Trip active level 3 (Medium)”, and “Trip active level 4 (High)”.

Moreover, we incorporated a variety of variables from the children’s travel diary data into the analysis. These variables referred to specific trip characteristics like the trip purpose, the land use of destination, accompany status, and the day of the week, etc. The subjective perceptions reported by children during each trip, including feelings such as anxiety, unhappiness, stress, and exhausted, were transformed into dummy variables. Socio-demographic factors like gender and age, were selected. Furthermore, weight and height of each child were processed to Body Mass Index (BMI) which serves as a general indicator of health [39].

Model development

To estimate the relationship between the Trip Active Level and physiological responses of children, specifically heart rate values, we conducted repeated measurements. The repeated nature of these measurements introduced dependencies between observations that make it necessary to consider regression assumptions to accurately model the data. Specifically, we needed to account for the random heteroscestadicity resulting from differences between 70 individuals (Fig. 1), ensuring that the effects identified in the model represent the true impacts of the independent variables on the dependent variable and not unobserved variation between children.

To address this issue, we employed mixed effects models incorporating both fixed and random effects, which allowed us to model variation between individuals and within individuals across the repeated measurements. In this case, a random intercept term was introduced to capture the variation between children in their multiple physiological measures of heart rate, where such individual variation can be removed from the fixed effects, making the effects more realistic. The linear mixed regression model is represented in Eqs. 1 and 2:

Where:

Y_ij = physiological measurement (heart rate value) i of individual j

β_0j = intercept of individual j

x_mj = exposure variable m of individual j

β_m = model coefficent for variable m

β₀ = a fixed component

u_0j = specific component of individual j

e_0ij = error

From the Eqs. 1 and 2, we assume that physiological measurement (heart rate value) ‘\(\:i\)’ of individual ‘\(\:j\)’, is influenced by its exposure to various factors denoted as\(\:{\:x}_{mj}\), and these exposures have specific coefficients (\(\:{\beta}_{m}\)) that determine the strength and direction of their impact on heart rate. The equation incorporates both fixed components \(\:{\beta}_{0}\) that are constant for all individuals and random components \(\:{u}_{0j}\) that are specific to each individual, accounting for its variations. Random errors (\(\:{e}_{0ij}\)) are also added.

By implementing this approach, we developed four mixed linear models (Model 1 to 4) using a stepwise method, incorporating different types of independent variables to predict physiological responses. In Model 2, in addition to the Trip Active Level variables, we included variables of trip characteristics as fixed effects. Model 3 extended this by including socio-demographic variables, which were expected to have a systematic and predictable influence on the data. Finally, in Model 4, subjective perception variables were introduced into the model. The independent variables, along with their definitions and descriptive statistics, are presented in Table 1.

The aim of this stepwise approach was to gain comprehensive insights into children’s physiological responses during travel transitions from various perspectives. By taking into account trip characteristics, socio-demographic factors, and subjective perceptions, we aimed to draw reliable and meaningful conclusions about the factors influencing children’s physiological responses and also to compare the explanatory power of each model.

Multicollinearity was assessed in the regression analysis and the variance inflation factor (VIF) for all the variables was found to be within an acceptable range. Fitting and comparison of these four models was carried out using Akaike’s Information Criterion (AIC) and the R-squared coefficient of determination (R²). The latter was preferred as the primary measure of model fit, as it serves as an indicator of both model goodness-of-fit and complexity. All modeling was carried out in R version 4.2.2 using the lme4 package.