A Transformer-based Model for Older Adult Behavior Change Detection

The advancements in smart home technologies have created new opportunities for precise monitoring of older adults' daily activities to provide timely care, predict health problems, and promote independent living at home. Many recent studies have investigated abnormality detection in Activities of Daily Living (ADL), and some used deep learning methods to handle the problem. In this paper, we leverage Bi-Directional Encoder Representations from Transformers (BERT), as a state-of-the-art method in machine learning, to analyze older adults' ADL sequences. Due to the fine-tuning capability of transformers, they are a good fit for supervised tasks when a large labeled training dataset is not available. Their architecture also allows for Parallel Computation, which is important for (near) real-time abnormality detection. To the best of our knowledge, this is the first effort to represent the older adult's daily behavior as sequences of ADLs with the goal of applying transformers to the behavior change detection problem. We designed four experiments to illustrate the capability of Transformers in detecting individuals' behavior abnormalities. We conducted a case study on a two-resident ADL dataset to evaluate the model in four experiments. Our results show that a BERT-based classifier can effectively detect behavior abnormalities from sequences of ADLs. Also, transfer learning proved to be helpful when it comes to fine-tuning a pre-trained model for a new resident.