Initiated in 2022, the WEALTH project was funded by the EU Joint Programming Initiative Healthy Diet Healthy Life [20] under the call on STAMIFY (standardised measurement, monitoring and/or biomarkers to study food intake, physical activity and health).

In the framework of the WEALTH project, a multicountry mixed methods cross-sectional survey was conducted, which focused on enrolling and monitoring PBs and EBs of adult participants from 5 research centers across Europe, in the Czech Republic, France, Germany, and Ireland. Participants were invited to participate between March 2023 and March 2024 via stratified and convenience sampling methods through letters, emails, telephone calls, posters, flyers, and social media. France was the exception, recruiting their participants from the NutriNet-Santé study. With a target sample size of 600 participants (150 participants: 50% (n=75) female, aged 18‐64 years per country), enrollment was distributed across the centers meeting the project’s eligibility criteria (see below). A small cash reward incentive of €30-€40 (US $35-$47) per participant was given in the Czech Republic, Germany, and Ireland to increase response rates, participant compliance, and ensure the return of all devices. Participants who dropped out during the study or provided insufficient data were eliminated. Participants who dropped out before the center total was reached were replaced. A total of 627 adults completed the study (Table 1).

To be deemed eligible for inclusion, participants were required to be aged between 18 and 64 years. To maintain protocol homogeneity and ensure the relevance of the data processing methods applied, individuals with overt chronic diseases, physical impairments, and those engaged in shift work were excluded from the study. These exclusion criteria were implemented to uphold the integrity of the prescribed semistructured activity protocol and to prevent potential disturbances in the free-living period arising from participants’ unique circumstances and, consequently, enhance the fidelity of the ensuing analytical outcomes.

The WEALTH data collection encompassed 2 integral components (Figure 1). The first component involved a standardized semistructured lab–based activity protocol, specifically designed to collect labeled data that simulated common PBs and EBs. The second component extended the data collection beyond the laboratory setting. It incorporated a 9-day EMA data collection combined with worn accelerometer and wearable devices, thereby enabling the continuous data collection of individuals’ behaviors during free-living. To record the participants’ experience of the study, an end-of-study feasibility instrument was completed when the participants returned their devices. To uphold methodological rigor and ensure conformity across all stages of the study, a WEALTH standard operating procedures manual was developed to define the WEALTH research group’s study protocol consisting of different modules (wearables, anthropometric measurements, dietary assessment, and general questionnaire), field access with interview guidelines, order of measurements, measurement procedures, equipment requirements, prerequisites, and data entry. The manual served as a comprehensive reference tool, offering researchers clear guidelines to enhance the precision and effectiveness of the data collection processes across the various study locations.

All data comprising questionnaires and device data were compiled, processed, and stored in a quality-controlled common database in accordance with European Union data protection regulations. The accuracy of the EMA reports was evaluated using the activPAL’s CREA classification, which served as a gold standard for scarcely labeled free-living data [43]. The activPAL device was selected due to its unobtrusive and strategic placement on the body, enabling reliable continuous monitoring over 24-hour periods [44]. Data completeness was verified on a daily basis for each sensor to ensure integrity, and potential nonwear time was identified through 2 complementary approaches: participant self-reports collected via questionnaires and automated predictions derived from either the activPAL’s CREA algorithm or the GGIR package. This approach provided a systematic framework to assess data loss, ensure data quality, and mitigate limitations associated with the accuracy of self-reported labels. Fitbit device data and EMA prompted via the HealthReact software are stored on a European Union server with project-specific access. Fitbit data were initially transmitted and stored on a US server before retrieval by the HealthReact app. In accordance with the WEALTH standard operating procedures specifications and to ensure data quality management measures were upheld, daily monitoring processes were implemented to evaluate participant adherence or nonadherence across all centers. All study sites received email notifications regarding data inconsistencies, and the subsequent resolution of queries was conducted to increase data completeness and data quality. Identified discrepancies included inconsistent data, absent data (relevant EMAs, context-based questionnaire), range validations, and deviations from the protocol. In addition, weekly rollouts for monitoring the compliance with 24HDR entries and the general questionnaire were provided to all centers.

The data collected in the WEALTH project will be used initially for 2 purposes. For hip- and wrist-worn devices, ML models will be developed to classify PB and EB from raw tri-axial accelerometer data. Model development will be based on strict labeled data from the semistructured protocol as well as labeled data from EMA time stamps with valid behavior reports of participants during the free-living survey. Classifiers for PBs and EB will be modeled using multiple ML methods ranging from random forests to convolutional neural networks or more complex deep learning applications. In the German subsample, the data on energy expenditure from the semistructured protocol will be used to develop ML models estimating metabolic equivalent of task units to each behavior.

Predictive models will be developed using the lab-based protocol and subsequently externally validated and calibrated using free-living sensor data labeled via EMA responses. To ensure robust evaluation, a unified validation strategy will be applied, combining subject-independent holdout validation, k-fold cross-validation, and performance testing on EMA-labeled free-living data. Model performance will be assessed using established metrics, including balanced accuracy, precision, recall, and F-score, computed only on data not used for training.

To ensure comparability across devices, accelerometer data will be harmonized by resampling to 20 Hz, converting to gravitational units (g), and synchronizing all sensors using the start time recorded during the laboratory session. EMA, accelerometer, and dietary recall data will be aligned using time stamps, enabling consistent temporal integration of self-reported and sensor-derived behaviors.

EMA-labeled data used to train ML models will be further validated by comparing them with external references, including the activPAL CREA algorithm and established ActiGraph classification methods. Low-agreement segments will be excluded using threshold-based criteria, ensuring that only high-confidence labels contribute to model development as done in Sigcha et al [43]. Model success will also be assessed through consistency across devices and quantitative agreement between EMA reports, accelerometer-derived predictions, and dietary-recall events. Agreement will be evaluated using temporal overlap metrics, Bland-Altman comparisons, intraclass correlation coefficients, and concordance indices, such as Cohen’s or weighted kappa for categorical outcomes.

The second purpose of the WEALTH study is to evaluate the potential of the use of EMA methods to study the interrelation between PB and EB. Interrelationships will be examined between PB and EB as predicted from ML models and data on context as assessed by EMA time- and context-sensitive results on incident behaviors. Multibehavioral patterns on the data from these different sources will be identified to investigate the association on body composition, well-being, and health status (self-rated health and health-related quality of life). Based on the feedback questionnaires, feasibility will be evaluated for the included measurement methods, and factors (eg, country, age, gender, activity level) associated with the degree of acceptance will be defined.

Sample size in each test center was guided by an estimate for a sample needed to develop the ML models. The sample size required to train a ML model successfully depends on the complexity of the model, the use of prior knowledge within the model, and the acceptable difference between performance observed during training and performance that can be expected on yet unseen data. It was estimated that a sample of at least 100 participants would allow for a substantial range of ML models to be trained successfully. This estimate is further supported by evidence from large-scale free-living accelerometer datasets, where empirical learning-curve analyses demonstrate that model performance improves markedly with increasing sample size up to approximately 60‐100 participants, after which performance gains plateau, with only marginal improvements (<1%‐2% in macro-F1) for both classical ML algorithms (eg, random forest, gradient boost), and deep learning algorithms (eg, convolutional or recurrent neural networks) [45]. These findings suggest that a cohort of 100 participants provides sufficient data to train robust, population-level models while maintaining generalizability across diverse activity contexts. To ensure that the amount of available data would not be a limiting factor in the choice of ML models, and to safeguard against data loss or participant attrition, 150 participants were recruited in each center.

This multicenter study is being conducted in accordance with globally accepted standards of good research practice in agreement with the Declaration of Helsinki, and with local institutional research policies, procedures, and regulations. The study design was developed by academic and clinical expertise underpinned by scientific evidence that emphasizes the requirement for standardized measurement of PBs and EBs. Patients or the public were not involved in the development of the study protocol.

Ethics committee approval was granted in each of the 4 study centers prior to study commencement. In the University of Limerick, ethics committee approval was granted by the Education and Health Sciences Faculty Research Ethics Committee (approval no 22_09_10_EHS_); in Bremen, approval was granted by the Ethics Committee of the University of Bremen (approval no 2022‐25); in the Czech Republic, approval was granted by the Committee for Research Ethics at the University of Hradec Kralove (approval no 11/2022); and in France, approval was by Comité de Protection des Personnes CPP Ile-de-France VI (approval no 2022-A02208-35).

All participants received detailed information about the purpose, procedures, risks, and benefits of the study prior to participation. Written informed consent was obtained from all participants before data collection commenced. Participants were informed of their right to withdraw at any time without penalty. Participants’ privacy and confidentiality were strictly protected. Identifying details (including names, initials, , or other personal identifiers) have been omitted.

JPI-WEALTH partners abided by the guidelines on consent [46] and transparency [47] as adopted by the European Data Protection Board and will remain in line with any new guidance that may be issued by the European Data Protection Board. The JPI-WEALTH project partners conform to the relevant convention and protocols of the Council of Europe, including the Convention on Human Rights and Biomedicine [48], and protocols on Human Rights and Biomedicine [49]. In addition, data collection was in accordance with the International Epidemiological Association’s Guidelines for Proper Conduct of Epidemiological Research [50] and the Council for International Organizations of Medical Science’s International Ethical Guidelines for Health-related Research Involving Humans [51].

The WEALTH project is expected to generate one of the most comprehensive multicountry datasets integrating accelerometer-derived PBs, wearable-derived EBs, and EMA data collected in both controlled and free-living settings. Based on the study design, we anticipate producing (1) harmonized, labeled datasets for PB and EB classification, (2) ML models capable of classifying behavioral states from raw accelerometry, and (3) an evidence-based methodology to support the simultaneous measurement of PBs and EBs in real-world contexts. These data will support the development of harmonized, labeled datasets and ML models capable of classifying PBs and EBs from raw accelerometry. This is essential given longstanding concerns about the lack of standardized device protocols, heterogeneity in cut points, and poor comparability across studies and countries [5,6,45,46]. The outputs from WEALTH will be accessible through a dedicated website outlining the project’s aims, methods, and products [52]. Critically, the ML models developed will be publicly available, enabling researchers to apply them to similar data. Providing openly accessible models and processing frameworks will facilitate detailed monitoring and assessment of PBs, EBs, and their determinants, thereby supporting future research beyond the project’s duration and contributing to data sharing and protocol harmonization.

The wearable dataset collected in this study will be made publicly available following the completion of the study and the end of an embargo period, in line with data protection and ethical approvals. The data will be shared following the Findable, Accessible, Interoperable, and Reusable (FAIR)principles, starting with a comprehensive data catalog and detailed data description, to facilitate transparency, reproducibility, and further research in this domain. During the embargo period, these data are available upon reasonable request in agreement with the research team.

Prior research has demonstrated the potential of ML to improve the classification of PB beyond traditional cut-point methods [11,12,45]. However, most existing work has been constrained to single-device datasets, laboratory environments, or narrow demographic groups. Similarly, while emerging studies have shown that wrist-worn sensors can help detect eating episodes and components of meal behaviors [15-18], free-living detection remains challenging due to variability in movement patterns and context.

Existing EMA studies have demonstrated strong value in capturing contextual, emotional, and environmental correlates of behavior [4,19] but have rarely integrated automated event-triggered surveys based on wearable sensor data. Previous work has also been limited by participants needing to self-initiate logs, contributing to recall bias and incomplete data [4]. The WEALTH design advances this field by combining 4 wearable devices, automated EMA triggers, multicountry standardization, 24HDR data, and cross-device timestamp harmonization. To our knowledge, this integrated approach has not previously been attempted at this scale in free-living behavioral research. This level of methodological integration directly addresses persistent methodological challenges documented in the literature, including cross-device comparability and harmonization of accelerometry outputs [5,6], the limited contextual fidelity of self-report and EMA-based behavioral monitoring [4,19], and the substantial complexities associated with accurately detecting and classifying EBs using wearable sensing technologies in free-living settings [15-18].

The aspiration is for WEALTH methodologies to be implemented in national and international nutritional and PA surveillance efforts. The ultimate goal is to establish a more robust system for assessing real-world data on PBs, EBs, and their determinants. This enhanced assessment capability will aid in designing public health interventions that promote health-enhancing PA and healthy diets, particularly in at-risk populations. These methodological advancements are expected to build a stronger evidence base on the health impacts of PBs and EBs, informing the development of improved national and European guidelines and recommendations aimed at enhancing citizen health. Furthermore, this accessibility is intended to support national and international efforts to achieve the United Nations Sustainable Development Goals by informing public health interventions and policies targeting PA and healthy diets in at-risk populations.

The project will deliver several key products and methodologies:

These outputs are designed to maximize the use of new knowledge and methods developed within WEALTH by project stakeholders and European Union citizens, ultimately contributing to public health improvement and policymaking. By ensuring both ML models and datasets are publicly available, WEALTH actively promotes reproducibility, transparency, and wider scientific collaboration.

We used a standardized protocol across multiple countries to capture both lab-based and free-living PBs and EBs. The combination of research-grade and consumer-grade wearable devices enabled the comprehensive monitoring of participant activities and physiological responses. The 9-day EMA and 24HDR provided rich, real-time data for examining behavioral patterns in real-life contexts. The backend engineering protocols of HealthReact are proprietary due to commercial licensing, which may limit direct replicability. The intensive data collection and device reliance may have introduced participant burden and technological challenges, affecting data quality. Self-reported EMA and dietary data could be subject to recall and social desirability bias, despite correction efforts.

This protocol outlines the design and methodology of the WEALTH project, a multicountry initiative to develop standardized, integrated approaches for assessing PBs and EBs using wearable sensors, ecological momentary assessment, and ML techniques. Although the findings are forthcoming, the anticipated outputs include harmonized datasets, openly accessible ML models, and an integrated data processing framework. These outputs have the potential to strengthen behavioral surveillance, support epidemiological and interventional research, and inform public health guidance. By combining rigorous standardization with advanced sensing and analytical methods, WEALTH is positioned to make a substantive contribution to the harmonization of PB and EB measurement across Europe and to future digital health innovation.