Real-life assessment of physical activity – Insights from the WEALTH European project

Overall abstract for the symposium

Purpose: The primary purpose of this symposium is to discuss innovative aspects of physical activity measurement provided by work performed in a currently ongoing collaborative European research project (WEALTH).

Description: The symposium will start with an introduction by the Chair (Jean-Michel Oppert, France) on the topic of physical activity monitoring in real-life settings and the needs for improved objective assessment in health surveillance surveys at national and international level. The first presentation (Alan Donnelly, Ireland) will provide an overview of the WEALTH project which main objectives are to develop standardized machine learning (ML) data processing techniques for accelerometer data and to test the feasibility of employing Ecological Momentary Assessment (EMA) to link physical activity and dietary data. The recruitment of participants across four European centers, and the data collection protocol using four accelerometer devices and an Ecological Momentary Assessment (EMA) app during a 9-day monitoring phase will be detailed. The second presentation (Tomas Vetrovsky, Czech Republic) will explain the EMA protocol combining time-based and event-based surveys to label free-living data measurement and assess environmental, social, and psychological contexts of physical and eating behaviors. The third presentation (Christoph Buck, Germany) will detail how ML models are trained and externally validated for prediction of physical and eating behaviors, based on wrist-worn accelerometry. EMA information is then used to refine classifiers of physical behaviors. The fourth presentation (Greet Cardon, Belgium) will focus on the feasibility and acceptability of methods under study and will use an online voting tool to interact with the audience and get the perspective of the researcher/policy maker. The final part of the symposium led by the Discussant (Sebastien Chastin, UK) will strongly involve the audience to get feedback on methods and results from the project and why, when and how they could be used to improve health surveillance surveys.

Chair: Jean-Michel Oppert, Sorbonne Paris Nord University, France
Presenter 1: Alan Donnelly, University of Limerick, Ireland
Presenter 2: Tomas Vetrovsky, University of Hradec Kralove and Charles University, Czech Republic
Presenter 3: Christoph Buck, Leibniz Institute for Prevention Research and Epidemiology – BIPS, Germany
Presenter 4: Greet Cardon, Ghent university, Belgium
Discussant: Sebastien Chastin, Glasgow Caledonian University, UK

Individual abstracts for each presenter

Presenter 1: ALAN DONNELLY, Limerick University, Ireland

Title: The WEALTH Project

Background: The WEALTH project (wearable sensor assessment of physical and eating behaviors) is a five-country European partnership focused on the development of improved measurement methods for physical behaviors and their impact on eating behaviors.

Purpose: To develop standardized machine learning data processing techniques for research grade and commercially accessible accelerometer device data, and to test the feasibility of employing Ecological Momentary Assessment (EMA) to link physical activity behaviors and dietary behaviors.

Methods: 600 adult participants were recruited (150 from four European centers) to simultaneously wear four accelerometer devices including two wearables (ActiGraph GTX3, ActivPAL3M, FitBit Charge 5 and Skagen Falster Gen 6) for nine days. Physical and eating behaviors were labelled using feedback from event-triggered and random EMA questions. Anthropometric measures including height, weight and grip-strength were recorded, and participants completed three web-based SACANA 24-hour dietary recall questionnaires. Accelerometer and EMA data were employed to: 1) develop machine learning data-processing techniques to identify activity behaviors, and 2) test the feasibility of using triggered EMA to assess the link between physical and eating behaviors. Participants completed study feasibility questionnaires at the end of the measurement period.

Results: Data collection will be completed by March 2024, and data processing and analysis is currently underway. The large-scale data collection will generate a database which can be interrogated to develop novel methods to explore the relationship between physical and eating behaviors.

Conclusions: The WEALTH project has illustrated the feasibility of employing combined accelerometry and event triggered and random EMA questions to study physical and eating behaviors in large populations.

Practical implications: The WEALTH project methods will be shared at the end of the project and will demonstrate that such methods can be used in national and international surveillance studies.

Funding: ERA-Net HDHL-INTIMIC (grant agreement No 727565).

Presenter 2: TOMAS VETROVSKY, University of Hradec Kralove and Charles University, Czech Republic

Title: Ecological Momentary Assessment (EMA): Real-Life Instantaneous Behavior Assessment

Background: Accurate measurements of physical behaviors (PB) and eating behaviors (EB), along with their context, as determined using Ecological Momentary Assessment (EMA), are crucial for understanding the determinants of healthy lifestyles. The WEALTH project aims to develop data processing methods for accelerometer data through a combination of machine learning and EMA.

Purpose: To report on the EMA data collection process undertaken within the WEALTH project.

Methods: The study included a 9-day free-living EMA data collection using the HealthReact system. Participants aged 18-64 years were recruited from four centers (Ireland, Germany, France, Czechia). The EMA protocol combined self-initiated, time-based, and event-based surveys to assess the environmental, social, and psychological contexts of PB and EB. Participants were instructed to self-report EB after each meal, snack, or drink (excluding water). Additionally, they received six time-based surveys daily. Finally, event-based surveys were triggered by near-real-time data from the Fitbit activity tracker following episodes of walking (5min, ≥60 steps/min, max. 4/day), running (5min, ≥140 steps/min, max. 4/day), or prolonged sitting (20min, 0 steps, max. 3/day).

Results: Among the enrolled participants (n=455), an average of 4.4 self-initiated reports of EB per day were recorded. Beyond the six daily time-based surveys, participants were prompted to complete, on average, 4.6 event-based surveys daily, of those x triggered by walking, y by running, and z by prolonged sitting. The response rates for the time-based surveys and those triggered by walking, running, and prolonged sitting were 58%, 65%, 59%, and 66%, respectively.

Conclusions: The collected EMA data will be instrumental in developing and validating machine learning algorithms for detecting PB and EB and their contexts.

Practical implications: The results will be applicable for surveillance and monitoring at the population level, as well as in public health interventions promoting healthy PB and EB.

Funding: ERA-Net HDHL-INTIMIC (grant agreement No 727565).

Presenter 3: CHRISTOPH BUCK, BIPS, Germany

Background: The development of new machine learning (ML) algorithms for behavior classification creates the need for further collection of high-quality data. Based on scripted studies for labeled data assessment, recent ML behavior classifiers show high accuracy mostly in a similar study design, but perform lower based on free-living data.

Purpose: To develop and improve ML models for physical behavior classification considering scripted and free-living labeled data.

Methods: In the WEALTH study, we collected data from 600 participants in four European countries conducting a two-hour scripted study were participants followed a behavior protocol, followed by a nine-day free-living data assessment, where data labelling was supported by ecological momentary assessment (EMA) to identify time stamps of similar behaviors. Classifiers of physical behaviors were modeled based on raw tri-axial accelerometer data from three different sensors, i.e. Actigraph GT3x, ActivPal and 25hz smartwatch data. We used multiple ML methods ranging from random forests as a benchmark model to convolutional neural networks (CNN) as deep learning applications. Based on the scripted study, holdout validation was conducted for model development, and models were further calibrated using free-living sensor data identified by EMA labels.

Results: Our models will be able to classify physical behaviors such as sitting walking, cycling and running, with high accuracy, whereas first results indicate lower specificity for high intensity behaviors due to extensive variation in raw accelerometry data.

Conclusions: Based on EMA labeled free-living sensor data, development of ML models for physical behaviors will show improved
performance for classification from tri-axial sensor data based on free-living study designs.

Practical implications: The developed ML models will allow behavior classification particularly for easy to apply wrist worn wearables, particularly smartwatches, that can be implemented in large surveillance studies.

Funding: ERA-Net HDHL-INTIMIC (grant agreement No 727565).

Presenter 4: GREET CARDON, Ghent University, Belgium

Title: Feasibility and acceptability of technology to provide physical activity and dietary behavior.

Background: Technology to provide real-time, valid physical activity and dietary behavior data holds potential for advanced monitoring and surveillance in adults. The WEALTH project advances the assessment and monitoring of physical and dietary behaviors by maximizing the use of data collected from multiple commonly available measurement devices by standardized data analysis and using a Machine Learning data processing platform and by linking advanced processing of activity data with triggered Ecological Momentary Assessments questioning.

Purpose: We aimed to assess the feasibility and acceptability in study participants and in researchers / policy makers when applying these methods for health behavior assessment.

Methods: As part of the WEALTH project a convenience sample of about 150 participants from 4 countries (Ireland, Germany, France, Czech Republic) filled out a questionnaire (29 questions, 5-point likert scale) after wearing 4 devices (Fitbit, ActivPAL, Skagen, Actigraph) and receiving smart phone based surveys (self-initiated and time based) for 9 days.

Results: Data collection in participants is currently ongoing and will be finalized in February 2024. Results will be presented for e.g.: comfort to wear the devices, ease of handling, comfort of data sharing, ease of responding to smart phone surveys, reactivity on smart phone surveys, technological ease, ease of compliance, privacy issues, general acceptance of the study. Making use of an online voting tool (Mentimeter) in the current presentation the audience will furthermore be asked to rate the feasibility and acceptance of the presented WEALTH methodology to get the perspective of the researcher/policy maker.

Conclusions: Conclusions will be drawn on the feasibility and acceptability of the methodology both from the perspective of study participants as from the researcher/policy maker perspective.

Practical implications: Findings will inform improvements needed in order to harness the potential of technology to provide real-time, valid data for monitoring and surveillance.

Funding: ERA-Net HDHL-INTIMIC (grant agreement No 727565).