Room-Aware Transformer for Activities of Daily Living Recognition in Adults with Mild Neurocognitive Disorder Using Ambient Sensors: Development and Evaluation Study
This study demonstrates that a room-aware Transformer architecture, which preserves location-specific sensor structures, offers context-dependent improvements in recognizing activities of daily living for adults with mild neurocognitive disorder using ambient sensors, though performance varies significantly across homes and activity types, highlighting the need for personalized model development before clinical application.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer
In the quiet rhythm of daily life, the way a person moves through their home tells a story. For adults with mild neurocognitive disorder, a condition where thinking skills decline slightly more than normal aging but daily independence remains mostly intact, these patterns can shift in subtle ways long before a doctor notices. Traditional methods of tracking these changes rely on infrequent clinic visits or the memories of family members, which can miss the slow, steady drift of a routine. To catch these changes early, researchers have turned to unobtrusive sensors—small, unobtrusive devices placed in homes that track motion, temperature, light, and door openings without cameras or microphones. These sensors create a continuous, minute-by-minute record of a household's activity, offering a window into how a person functions in their own space. The challenge, however, is that every home is different; layouts vary, sensor placements differ, and people move in unique ways, making it difficult to build a single computer program that works everywhere.
A team of researchers set out to solve this problem by teaching a computer to recognize daily activities, such as eating, sleeping, or moving from room to room, using data from five real homes of adults with mild neurocognitive disorder. They tested whether a new type of computer model, which pays specific attention to the location of each sensor, could understand these routines better than older, more general models. The researchers collected nearly 93,000 minutes of data from motion detectors, temperature gauges, light sensors, and door contacts across five different houses. They paired this sensor data with detailed logs written by the participants and their caregivers, which recorded exactly what activity was happening and where. To make the data manageable, they grouped the original thirteen types of activities into eight broader categories, such as "Sleep/Rest," "Eating," and "Leisure/Social."
The team trained three different types of computer models to predict these activities. One model was a standard approach that treated all sensor data as a single, mixed stream. Another was a more advanced model that looked at the entire sequence of events at once. The third, and most novel, was a "room-aware" model. This model was designed to understand that a motion sensor in a kitchen tells a different story than one in a bedroom. Instead of mixing all the data together immediately, this model first processed the signals from each room separately, learning the specific patterns of movement and environmental changes for that space, before combining them to understand the full day's story. They tested these models using time windows of 10, 30, and 60 minutes to see how much past context the computer needed to make an accurate guess about what was happening right now.
The results showed that no single model was perfect for every situation, but the room-aware approach offered distinct advantages in specific contexts. In nine out of fifteen comparisons between different homes and time windows, the room-aware model achieved the highest accuracy, correctly identifying the activity more often than the other models. For instance, in one home, the room-aware model correctly identified activities 37% of the time over a 60-minute window, while the other models managed only 33% and 23%. However, the study also found that the room-aware model did not consistently win in every category. When measuring the model's ability to distinguish between different activities using a different statistical score, the standard model that ignored room locations sometimes performed slightly better. This suggests that while knowing where a sensor is helps, it does not automatically solve every problem, and the best approach depends heavily on the specific layout of the house and the habits of the person living there.
The researchers also discovered that the length of the time window mattered, but not in a simple way. Sometimes, looking at just the last 10 minutes of data was enough to get the best result, while in other homes, looking back 60 minutes helped. There was no universal rule that longer observation periods always led to better predictions. The models were also much better at recognizing common activities like sleeping or resting, which made up a large portion of the data, than they were at spotting rare or brief activities like dressing or using the toilet. In fact, for some of these less frequent tasks, the models struggled to make reliable predictions, often scoring very low. This highlights a significant hurdle: the computer needs a lot of examples to learn how to recognize rare events, and in real-world homes, some activities simply do not happen often enough to teach the system well.
Ultimately, the study suggests that building a system to monitor daily life for people with mild cognitive decline requires a personalized approach. A model that works well in one house may not work in another because of differences in how the home is arranged or how the person moves. The room-aware model showed promise by respecting the unique structure of each home, but it was not a magic solution that worked everywhere. The researchers concluded that before such technology can be used to guide medical care or support for families, it must be tested in much larger and more diverse groups of people. The path forward involves refining these models to handle the messy, varied reality of real homes, ensuring that the technology can reliably spot the subtle signs of change that matter most to the people living there.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.