Real-Time Stream Learning System for Monitoring Activities of Daily Living in Older Adults

Paula Sofía Muñoz Ordoñez,1 Ana Sofía Orozco Orozco,1 Ricardo Salazar-Cabrera,1 Diego M López1,2 1Department of Telematics, Universidad Del Cauca, Cauca, Colombia; 2Faculty of Health Sciences, Universidad Del Cauca, Cauca, ColombiaCorrespondence: Ricardo Salazar-Cabrera, Email [email protected]: Monitoring activities of daily living (ADLs), such as walking, sitting, and stair climbing, is an important indicator of functional status, autonomy, and overall well-being in older adults. Traditional assessment approaches, such as questionnaires or offline machine learning models, struggle to adapt to dynamic environments and to the natural variability of human movement. This study aims to design, implement, and evaluate a real-time data collection and processing system that supports the training and assessment of Stream Learning (SL) models for ADL classification, and to compare their performance with conventional offline models. This study aims to design, implement, and evaluate a real-time data collection and processing system for ADL classification using Stream Learning (SL) models. Additionally, the study analyzes whether SL-based monitoring can provide more reliable and responsive metrics than traditional offline approaches under real-world conditions.Materials and Methods: The study involved nine adult participants during offline and online experimental phases aimed at evaluating real-time ADL monitoring under controlled and uncontrolled conditions. A mobile application captured accelerometer and gyroscope signals from smartphones and transmitted the data streams to a cloud server for window segmentation, label alignment, storage, and incremental model updates. Both offline models and SL algorithms were evaluated using precision, adaptability, and prediction stability metrics. The data captured were obtained through experiments approved by an Ethical Committee, and those involved signed an informed consent.Results: The system successfully enabled continuous real-time data acquisition and incremental model training. Offline models achieved competitive accuracy but limited adaptability to variations in movement patterns or acquisition conditions. SL models demonstrated more robust and stable predictions, adapting better to the natural variability of daily activities in real-world scenarios.Conclusion: The proposed system demonstrated the feasibility of integrating mobile sensing and SL for continuous ADL monitoring under real-world conditions. Compared with traditional offline models, SL approaches achieved superior adaptability, stability, and responsiveness during continuous operation, with Hoeffding Tree showing the best overall balance between accuracy and latency. These findings support the potential of SL-based systems for scalable and low-cost functional monitoring applications. Further validation using larger and clinically characterized populations is still required.Keywords: activities of daily living, stream learning, inertial sensing, incremental machine learning, online classification, offline classification, data stream processing, mobile health monitoring