1D-ResNeXt Deep Learning Architecture for Recognizing Activities of Daily Living in Elderly Populations Using Wearable Sensors
Sakorn Mekruksavanich, Datchakorn Tancharoen, Anuchit Jitpattanakul · 2025
This study proposes an unexplored deep learning framework, known as 1D-ResNeXt, tailored explicitly for detecting daily activities among elderly individuals using data collected from wearable sensors. The architecture leverages the powerful pattern recognition capabilities of the ResNeXt model, adapted for sequential one-dimensional signals through convolutional blocks structured around cardinality—a mechanism that improves the model’s capacity to extract complex temporal features. To evaluate the proposed model’s effectiveness, we conducted experiments on a comprehensive dataset consisting of sensor streams from accelerometers, gyroscopes, and physiological devices gathered from 18 older adults performing a variety of routine tasks. The 1D-ResNeXt model achieved a classification accuracy of 97.46%, outperforming conventional deep-learning approaches and integrated hybrid models by margins of 5.17% and 6.41%, respectively. Moreover, the framework demonstrated superior performance in distinguishing between transitional movements and activities with overlapping motion signatures, often posing significant classification challenges in geriatric populations. Its capability to learn layered feature representations while maintaining efficient parameter usage makes the architecture particularly advantageous for deployment in wearable platforms with constrained hardware resources. The high prediction accuracy and low computational overhead support for real-time data processing position this system as a viable solution for incorporation into intelligent home ecosystems and continuous health surveillance infrastructures. Ultimately, it holds promise for enabling early detection of physical deterioration and supporting independent living strategies for older people.