Data Labelling for Free-Living Physical Activity Recognition using Thigh-Worn Wearables and Event-based Ecological Momentary Assessment

Accurate assessment of physical activity (PA) is critical for developing effective public health guidelines and interventions. Although wearable devices and machine learning approaches have advanced PA recognition, models trained on laboratory data often fail to generalize to free-living conditions due to variability in activity patterns and data acquisition. Overcoming this limitation requires the development of models based on real-world datasets, despite the challenges inherent in obtaining accurate labelled data. Here, we present a framework that integrates thigh-worn accelerometer data with event-based Ecological Momentary Assessment (EMA) surveys to label different types of PA in free-living settings. Data were collected from 589 participants over a seven-day monitoring period as part of the WEALTH study. EMA responses were synchronized with accelerometer signals to create a sparse labelled dataset, which was compared against ground-truth labels generated by the proprietary ActivPAL CREA algorithm. The framework successfully labelled six PA categories with up to 97\% agreement with ground-truth data. A machine learning algorithm trained on the resulting dataset achieved classification accuracies of up to 73.6%. Overall, the framework demonstrates high labelling accuracy for activities such as sitting and running, and offers a promising approach for generating reliable, real-world datasets to advance sensor-based PA monitoring.