This study used data from the WEALTH project, a large, multicenter cohort conducted in Czechia, France, Germany, and Ireland. The study supported the development of standardized methods for assessing daily PA, sedentary behavior, and eating behaviors by integrating structured laboratory assessments with free-living monitoring to ensure ecological validity.

Participants first attended a laboratory session involving a 75-minute semistructured activity protocol, during which well-defined movement patterns were recorded under controlled conditions using 4 different sensors including activPAL 3 micro (PAL Technologies), ActiGraph GT3X (ActiGraph LLC), Skagen Falster Gen 5 smartwatch (Skagen, Fossil Group Inc), and a Fitbit Charge 5 (Fitbit Inc., Google LLC). This was followed by a 9-day free-living phase, during which the 4 different sensor devices collected accelerometer data in combination with event-based EMA to capture contextual information about participants’ behaviors in real-world settings [29].

A cohort of 627 adults was recruited across 4 European study sites using a convenience sampling strategy through university, community, and institutional advertisement channels, resulting in a selective but demographically diverse sample. Sample size was determined according to the WEALTH study protocol [29]. The target sample was selected to support machine-learning model development and to accommodate expected data loss arising from device noncompliance and data quality exclusions. Participants were required to have no mobility impairments and to be able to complete both the laboratory and free-living protocols. From the initial cohort, 32 participants did not provide concurrent activPAL and ActiGraph recordings, and an additional 5 participants were excluded due to significant temporal drift that precluded reliable sensor synchronization. The final sample consisted of 590 participants with valid and concurrent activPAL and ActiGraph data. Participant characteristics are summarized in Table 1 and reflect variability in age, sex, country, and BMI, including representation of individuals classified as obese (BMI ≥30 kg/m²).

A combination of research-grade and consumer-grade wearable devices was used to capture movement patterns, posture, and activity-related events throughout both the laboratory session and the 9-day free-living period. As part of this multimodal setup, an activPAL 3 micro was placed on the midline of the anterior aspect of the right thigh using a hypoallergenic adhesive tegaderm dressing. The device recorded continuous triaxial acceleration at 20 Hz, provided estimates of activity intensity expressed as METs, and behavioral outputs through its proprietary CREA algorithm, classifying sedentary, standing, stepping, cycling, and nonwear, which offered a reference for refining EMA-derived annotations.

To complement the thigh-worn sensor, participants also wore an ActiGraph affixed to the right waist. This device provided higher-frequency (100 Hz) movement data, which contributed to the development of dual-sensor models by enabling an additional perspective on upper-limb and whole-body movement patterns.

Finally, participants wore a Fitbit Charge 5 on the nondominant wrist to trigger event-based EMA surveys. Although Fitbit data were not used as model input, step count and heart rate data were used to detect predefined behavioral conditions and automatically initiate EMA prompts. An overview of the body-worn wearables, data sources, and processing workflow used in this study is provided in Multimedia Appendix 1.

EMA assessments were implemented using the HealthReact mobile app [28], which delivered three types of surveys throughout the 9-day free-living period: (1) event-based (sensor-triggered) surveys, (2) time-based surveys, and (3) participant-initiated surveys. Only data from event-based surveys were used, as they were automatically triggered in response to specific sensor-detected activities relevant to this study.

Event-based survey prompts were generated automatically based on Fitbit-recorded data, triggering an EMA survey after ≥20 minutes of zero steps for sedentary behavior, ≥5 minutes of walking at 60‐139 steps/minute, or ≥5 minutes of running at ≥140 steps/minute. Participants reported the activity and posture being performed, while timestamps were automatically logged by the HealthReact app. Prompt allocation rules were pilot-tested following Janek et al [23], with up to 4 daily prompts for sedentary behavior and up to 3 each for walking and running.

The raw accelerometer data were processed using ActiLife (v6.8) for the ActiGraph devices and PAL Analysis (v9.1) for the activPAL devices. For activPAL, the activity annotations and MET-based activity intensity estimates (expressed in MET-seconds) were extracted using PAL Analysis software and the proprietary CREA algorithm. An epoch length of 60 seconds was applied to ensure consistency with prior studies and established protocols for MET-based activity intensity assessment [30,31].

To harmonize the recordings across devices, all acceleration signals were expressed in gravitational units (g). ActiGraph data (100 Hz) were resampled to 20 Hz to match activPAL’s native frequency. Initial synchronization was performed using device timestamps and the recorded start of the laboratory session. A fine-grained alignment was subsequently performed and verified by visual inspection using MATLAB (R2020a). Temporal offsets attributable to internal clock drifts were compensated by inserting a small, fixed number of samples (80 samples per day) into the ActiGraph data at a daily interval. This correction accounted for systematic clock drift between activPAL and ActiGraph devices during prolonged free-living recordings, which can produce cumulative timing misalignment between posture and activity signals despite simultaneous initialization [32].

Following device-level synchronization, the CREA-derived MET estimates were temporally aligned with accelerometer time-series. Activity intensity estimates were converted from MET-seconds to MET values and subsequently mapped to categorical intensity. In this study, AIC were operationalized using MET-derived categories rather than direct kcal-based energy expenditure estimates to enable standardized intensity classification for population-level free-living analyses. This enabled classification into sedentary behavior (1.0‐1.5 MET), LPA (1.6‐2.9 MET), and MVPA (≥ 3 MET) categories using standard MET-based cut-points [7,8].

Sparse free-living labeled data were generated following the EMA-based labeling framework proposed in Sigcha et al [24]. During the 9-day data collection period, 26,166 EMA activity surveys were recorded. After removing duplicates and surveys lacking valid posture or activity information, 16,907 surveys were retained for wearable data annotation. EMA triggers were used to segment the continuous accelerometer data streams, and each resulting data segment was labeled using the participant’s response to the associated EMA prompt.

The resulting segments were annotated according to predefined PB categories established within the WEALTH study: sitting, standing, walking, running, sports, cycling, and lying, resulting in temporally sparse but highly specific annotations. This EMA-based approach prioritizes ecological validity and label specificity over label density, thereby providing high-quality annotations under free-living conditions.

To strengthen label reliability, only segments showing at least 75% sample-wise agreement with CREA were retained for model training and evaluation. In line with the findings reported in [24], this threshold was selected as an appropriate trade-off between data quality and quantity, reducing less confident segments while preserving sufficient free-living coverage across activity types, ensuring that the final labeled dataset was both ecologically valid and of high quality.

Following EMA-based annotation and quality control using a ≥75% agreement threshold with the CREA algorithm, 2108 hours of labeled free-living data from 583 participants were retained for model development and validation. The reduction in the number of participants from 590 to 583 was primarily due to EMA surveys that could not be synchronized with the corresponding accelerometer recordings because of time-stamp mismatches, which placed the EMA entries outside the valid recording range and prevented reliable segment construction. Detailed distributions of the EMA-labeled data before and after quality control are provided in Multimedia Appendix 2.

In addition to EMA-labeled data, the 9-day free-living (unlabeled) recordings from activPAL and ActiGraph devices were synchronized to enable the implementation and evaluation of a dual-sensor approach for generating model-derived daily PB and AIC estimates, as well as their corresponding 9-day averages. To ensure strict temporal correspondence between the 2 devices, each pair of recordings was then trimmed to match the sensor file with the lowest sample count and visually inspected using MATLAB.

Following this synchronization procedure, complete unlabeled free-living data were obtained from 590 participants (Table 1), representing the full 9-day monitoring period and providing the basis for subsequent analyses of daily PB and AIC profiles.

To prepare the data for model development, raw time-series signals were segmented using nonoverlapping windows. Multiple window lengths (5, 10, and 15 s) were evaluated to determine their influence on classification performance, reflecting common practices in human activity recognition [24,25]. For each window, a PB label (task 1) and AIC label (task 2) were assigned using 2 different approaches. The PB label was directly taken from the EMA-reported activity for the corresponding prompt, reflecting the participant’s self-reported behavior within that data segment. In contrast, because MET-based estimates were available at a 60-second epoch resolution, a majority voting technique [33] was applied to assign a single AIC label to each window. Specifically, each window was labeled as Sedentary, LPA, or MVPA based on the AIC most frequently observed in the specific window.

Three feature-extraction strategies were implemented to support classical ML pipelines. First, a set of 40 features was computed following the CAPTURE-24 specification [25], capturing statistical, spectral, and temporal characteristics. Second, an adapted version of the HARTH feature set [34] was generated for single-sensor use, extending its applicability beyond the original dual-sensor design. Third, the original HARTH feature set [34] was extracted for dual-sensor configurations to leverage complementary information from thigh- and wrist-based measurements. A detailed description of the set of features extracted in this study is provided in the Multimedia Appendix 3.

In parallel, raw accelerometer data were used to train a multihead convolutional neural network (MH-CNN), implemented in both single- and dual-sensor variants. This DL-based approach enabled the extraction of hierarchical temporal patterns directly from the raw signals, providing an alternative to feature-engineered models.

To ensure a robust assessment of model performance, a subject-independent hold-out strategy was implemented using the EMA-labeled dataset. Of the original cohort (N=590) participants, data from 583 individuals were available for model training after excluding cases where EMA responses could not be synchronized with accelerometer data. The resulting participants were allocated to training (407/583, 70%), validation (87/583, 15%), and test (87/583, 15%) subsets, with each individual assigned to only 1 data subset to prevent information leakage across subsets. Stratification by age, gender, and country preserved demographic balance across subsets to support a reliable model evaluation. The number of participants and the characteristics of the data subsets used to train and evaluate the ML models are detailed in the Multimedia Appendix 6.

Model performance for both PB and AIC classification was quantified using classification metrics, including precision, recall, and F₁-score [36]. These complementary indicators allowed for a comprehensive evaluation across the imbalanced activity classes. Figure 2 presents the overall methodology for ML model development and agreement analysis.

ML and DL models were trained on a workstation equipped with an Intel Xeon 2.30 GHz processor, 25 GB RAM, and a 12 GB NVIDIA Tesla K80 GPU. ML models were implemented in Python (v3.12) using Pandas (v2.2), NumPy (2.0.2), and Scikit-learn (v1.6). To address class imbalance, the ML training subset was undersampled so that all activity classes matched the size of the smallest class. Undersampling was chosen because the dataset is highly skewed toward a single category; RF models can be biased by class imbalance; it also reduces training time, mitigates the risk of overfitting, and avoids potential artifacts introduced by synthetic data generation techniques such as the synthetic minority oversampling technique (SMOTE).

DL models were developed using Keras (3.10) and TensorFlow (2.19). Unlike the ML pipeline, undersampling for the DL model was applied selectively by reducing only the most frequent classes (sitting to 30,000 samples; lying to 25,000; walking to 20,000) to preserve the natural distribution of daily activities while controlling class dominance. As with ML, undersampling was performed exclusively on the training data subset.

The MH-CNN architecture was trained using a multitask learning framework with task-specific weighting prioritizing PB over AIC classification (ratio 10:1), together with class-weighted losses and early stopping to reduce overfitting. Optimization was performed using the adaptive moment estimation (ADAM) optimizer [37]. Additional details regarding training configuration and procedures are provided in Multimedia Appendix 5.

Different experiments were conducted to evaluate the effects of window size, sensor type, and sensor configuration on PB and AIC classification.

The effect of window size was evaluated by segmenting synchronized accelerometer data into nonoverlapping windows of 5, 10, and 15 seconds. These analyses were performed using the dual-sensor configuration to identify an appropriate temporal resolution for subsequent experiments.

Single-sensor experiments were then carried out separately for activPAL and ActiGraph. For each device, RF models trained with hand-crafted features (CAPTURE-24 and an adapted single-sensor HARTH feature set) were compared with MH-CNN trained on raw accelerometer signals. This design allowed for a direct comparison between feature-engineered ML approaches and end-to-end DL models.

Finally, dual-sensor experiments were performed by combining activPAL and ActiGraph data. In this setting, RF models were trained using the original (dual-sensor) HARTH feature set and compared with the MH-CNN trained on synchronized raw data from both sensors.

In all experiments, PB and AIC classification were treated as separate outcomes but evaluated within the same experimental framework. Results of these experiments are reported as mean and SD across 5 repeated runs to identify the model stability during training. In addition, bootstrap resampling (n=100) was used to estimate 95% CIs for the F₁-scores across the evaluated window lengths following the methodology proposed by Efron [38].

Prior to the study’s commencement, ethics committee approval was obtained from all participating institutions: The Education and Health Sciences Faculty Research Ethics Committee, University of Limerick (22_09_10_EHS); the Ethics Committee of the University of Bremen (2022‐25); the Comité de Protection des Personnes CPP Île-de-France VI, Paris (2022-A02208-35); and the Committee for Research Ethics at the University of Hradec Králové, Czech Republic (11/2022). All participants provided written informed consent before data collection and were informed of their right to withdraw from the study at any time without providing justification. To acknowledge participants’ time and contribution, monetary compensation was provided: €40 in Ireland and €20 in the other countries. All collected data were deidentified for analysis and stored securely to maintain participant confidentiality.

This study provides a comprehensive development and evaluation of multitask ML models for simultaneous PB and AIC classification using multisensor accelerometry under free-living conditions. Leveraging data from 590 adults across 4 European countries monitored over a 9-day period, the findings demonstrate that integrating thigh- and waist-worn sensors within a multitask learning framework improves model performance and the stability of daily behavioral profiles compared with single-sensor approaches. Collectively, these results support the feasibility of scalable long-term wearable monitoring under real-world conditions, thereby providing ecologically valid evidence at a scale rarely reported in prior free-living wearable studies.

When applied to free-living data, the dual-sensor configuration produced PB distributions characterized by high sedentary time and comparatively limited engagement in ambulatory and vigorous behaviors. These estimates fall within ranges commonly reported in adult populations, where sedentary behaviors account for approximately 60%‐70% of daily time [31,43] and walking typically averages 1‐2 hours per day [44]. Similarly, low-prevalence behaviors such as running, sports, and cycling contributed minimally to daily activity, consistent with epidemiological observations that vigorous or sport-related activities occupy only a small fraction of daily time in adults [43,45].

For AIC, the dominance of sedentary time and the relative contributions of LPA and MVPA were consistent with population-based accelerometer studies reporting 8‐10 hours/day of sedentary behavior [46,47], 10%‐20% of time in LPA [48], and 20‐40 minutes/day of MVPA [39]. The dual-sensor configuration consistently produced intermediate estimates between single-sensor approaches, suggesting reduced sensor-specific bias and improved behavioral characterization across heterogeneous activity patterns. Similarly, the inclusion of participants with a broad range of BMI values, including individuals classified as obese, supported the representation of diverse PB and AIC patterns under free-living conditions.

Agreement analyses further showed that the dual-sensor approach aligned most closely with CREA for ambulatory behaviors (standing, walking, and cycling), while greater discrepancies were observed for prolonged sedentary behaviors, particularly lying. This pattern suggests that multimodal sensing enhances recognition of dynamic behaviors under free-living conditions but remains challenged by extended low-movement postures, a limitation also reported in prior work [49].

From a monitoring perspective, these findings have important implications for real-world implementation of wearable-based monitoring systems. First, they demonstrated that dual-sensor wearable monitoring could reduce modality-specific biases inherent to single-sensor approaches, particularly for posture-related behaviors. The dual-sensor configuration mitigated systematic over- and underestimation observed with activPAL-only and ActiGraph-only models, producing more balanced and interpretable daily PB profiles that are better suited for longitudinal monitoring and population-level analyses. The scalability of the dual-sensor approach warrants consideration for large-scale deployment. Although the use of 2 devices may increase participant burden, it proved feasible and was well accepted in our sample [50]. Furthermore, jointly labeled thigh- and hip-worn data enable the development of more accurate and generalizable models that can subsequently be applied in single-sensor configurations, supporting scalable mHealth apps [27].

Second, the integration of EMA-informed labeling provided a scalable and pragmatic solution to overcome the scarcity of reliable free-living training data. Event-based EMA enabled the collection of relevant behavior labels at the moment of occurrence, thereby improving ecological validity, supporting the development of models that generalize beyond controlled laboratory environments [23,24,26]. This approach is particularly relevant for mHealth studies requiring extended monitoring periods, where intensive manual annotation is not feasible.

Third, the relative stability of AIC estimates across sensor configurations suggests that multitask learning may be especially valuable for digital health apps focused on energy balance, PA guidelines, and surveillance of activity intensity distributions. Even when discrepancies were observed in PB classification, aggregated AIC estimates remained consistent with epidemiological norms reported in large cohort studies [39,46-48], supporting their use in public health monitoring and evaluation of intervention outcomes.

Previous studies on wearable-based activity recognition have frequently relied on laboratory-based protocols, short-term recordings, or single-sensor configurations, which may limit generalizability to free-living settings [22,25,26]. Although datasets such as CAPTURE-24 have advanced free-living activity recognition using video-based behavioral labels [16], reliance on wearable cameras and manual annotation may reduce scalability and raise privacy concerns for longer-term monitoring. In contrast, this study uses 9-day free-living data and EMA-informed labeling from a multicenter cohort, providing a rigorous assessment of model performance under realistic conditions.

Previous research has demonstrated the feasibility of EMA for behavioral annotation and the potential of multitask learning for activity recognition [17,18]. Extending this work, this study combined multitask learning with multisensor accelerometry in a large multicenter dataset, offering empirical evidence of their combined value for mHealth monitoring in real-world settings.

The findings are consistent with previous studies reporting that feature-engineered ML models can remain competitive with DL approaches for free-living activity recognition [21]. In this study, the MH-CNN achieved the highest PB classification performance, whereas the RF model achieved the highest AIC classification performance, suggesting that engineered temporal and statistical features remain informative for AIC classification under free-living conditions [25]. Although more complex sequential architectures may further improve temporal modeling, the selected MH-CNN offered a balanced compromise between predictive performance, complexity, and scalability for long-term monitoring.

Several limitations should be acknowledged. AIC estimates were derived from the activPAL CREA algorithm and operationalized using MET-derived categories rather than direct kcal-based energy expenditure, thereby limiting the assessment of continuous energy expenditure. In addition, the CREA algorithm contributed to the label-refinement procedure and was subsequently used as a comparator for daily estimates; therefore, comparisons with CREA-derived estimates represent assessments of agreement rather than independent validation. Although gold-standard methods such as doubly labeled water or other forms of indirect calorimetry would provide stronger absolute validation of the AIC estimates [3], these approaches were beyond the scope of this study and warrant further investigation within multitask wearable monitoring frameworks.

The behavioral taxonomy was limited to 7 PB categories and 3 activity intensity levels, which, although appropriate for large-scale monitoring, do not capture the full diversity of real-world activities. In addition, the EMA-labeled dataset was imbalanced, with low-frequency behaviors such as running and sports remaining underrepresented despite mitigation strategies including selective undersampling and class-weighted losses. Future research may benefit from data augmentation, self-supervised learning, or transfer learning approaches trained on large unlabeled datasets to improve recognition of infrequent behaviors [15].

Furthermore, although the MH-CNN improved PB classification performance, the comparatively smaller gains observed for AIC classification may reflect class imbalance and limitations in CNN-based temporal modeling under free-living conditions. Future work should evaluate sequential architectures such as LSTM- or Transformer-based models using larger (real-world) datasets. Finally, although the dual-sensor configuration improved performance, practical deployment in some mHealth contexts may require adaptation to single-sensor or consumer-grade devices, which warrants further investigation.

This study demonstrates that multitask ML models combining thigh- and waist-worn accelerometry can provide consistent estimates of PB and AIC under free-living conditions. The dual-sensor approach produced stable, epidemiologically coherent daily profiles and reduced modality-specific biases observed in single-sensor methods. By integrating multimodal wearable sensing with EMA-informed labeling at scale, this work advances methodological foundations for real-world mHealth monitoring and supports more consistent and comparable assessment of PB in population and public health research.