A Personalized Transformer Neural Network for Accurate Recognition of Health-Indicative Complex Activities from Smartphone Sensors

Kavin Chandrasekaran, Luke Buquicchio, Emmanuel Agu, Elke Rundensteiner

Confirmation: I have read and agree with the IEEE BHI 2025 conference submission's policy on behalf of myself and my co-authors.

Keywords: Activities of Daily Living, Personalized Complex Human Activity Recognition, Smartphone sensors, Transformers

TL;DR: This paper proves that personalizing a Transformer model to an individual's unique movement style is critical for accurately recognizing complex daily activities from smartphone sensors, boosting performance from 88% to over 92%.

Abstract: Continuous monitoring of Activities of Daily Living (ADLs) and Instrumental ADLs (IADLs) of the elderly is vital for their safety, independent living, quality of life and overall health. Declines in the ability to perform these tasks, which require the interplay of musculoskeletal, neurological, and cognitive systems, often indicate underlying health problems. Early detection of such declines can prompt timely interventions. Traditional ADL/IADL assessment was manually done infrequently by a skilled expert. Passive monitoring using data from the built-in sensors of ubiquitous devices such as smartphones offers continuous, objective monitoring of an individual's ADLs in their natural environment. However, the recognition of ADLs from sensor data faces challenges including high intra-class variability due to individualized styles of performing ADLs. One-size-fits-all machine learning models often misinterpret individual nuances in ADL performance, resulting in inaccurate health assessments. Personalization of ADL models can make them robust to inter-subject variability. This paper proposes the Personalized Health Activity Recognition Transformer (P-HART), a personalized transformer-based model that captures temporal relationships in sensor data for robust Complex Activity Recognition (CAR) from smartphone sensor data. In rigorous evaluation on a comprehensive complex ADL dataset, P-HART achieved an F1-score of 92.5% with personalization and 87.7% without personalization, outperforming baseline models and demonstrating the substantial benefits of ADL model personalization. P-HART facilitates remote health-indicative activity recognition and monitoring.

Track: 7. General Track

Registration Id: HJN8PCQFLLX

Submission Number: 258