SmartADAPT-Net: Decision-Level Adaptation for Free-Living Smartwatch-Based Activities of Daily Living Recognition

Everyday life is a sequence of small, often repetitive movements: brushing teeth, washing dishes, walking to the kitchen, or mopping a floor. For doctors and caregivers, tracking these "activities of daily living" is vital. A person's ability to perform these tasks independently is a primary indicator of their health and safety. Traditionally, measuring this relies on asking people to remember what they did or having a nurse observe them at specific times. These methods are imperfect; human memory is fallible, and a snapshot observation misses the rhythm of a whole day. To get a true picture of how someone is functioning, we need a way to watch them continuously without being intrusive. This is where the smartwatch comes in. Worn on the wrist, it carries sensors that can feel the subtle shifts of movement, offering a potential window into a person's daily routine. But there is a catch: a computer program trained to recognize these movements in a quiet, controlled lab often fails when the same person wears the watch in the messy, unpredictable reality of their own home.

Researchers at Queen Margaret University and Edinburgh Napier University set out to solve this specific problem. They asked whether a smartwatch system could learn to recognize a person's daily activities as it was being used, without needing to be completely retrained from scratch. They developed a system called SmartADAPT-Net, designed to work in two distinct phases. First, it uses a general "population" model—a set of rules learned from many different people in a lab—to make an initial guess about what activity is happening. Second, it slowly learns from the specific person wearing it, but only when that person confirms or corrects the watch's guess. The goal was to see if this two-step approach could bridge the gap between a lab-trained robot and a helpful, personalized assistant in the real world.

The researchers began by building the foundation of their system in a simulated home environment. Thirty volunteers performed six common tasks: brushing teeth, mopping or vacuuming, shelving items, walking, washing dishes, and washing their face. They wore Apple Watches that recorded the motion of their wrists. The team used this data to teach a computer model to recognize the patterns of these movements. This initial model was "frozen," meaning its core brain was locked in place and would not change later. However, the researchers added a clever layer of adjustment called "complexity-aware modulation." This feature looks at how chaotic or irregular a movement is. For instance, walking is usually rhythmic, while washing dishes might be more erratic. By measuring this complexity, the system could make a quick, smart adjustment to its initial guess, improving its accuracy before it even knew who the user was.

The true test came when they deployed this system to a new group of seventeen people in their actual homes for twenty days. These participants wore the watch and started a recording whenever they performed one of the six target activities. The system did not know these people beforehand. When the watch made a prediction, the user was asked to confirm if it was correct or to select the right activity from a list. This feedback was the key. The system did not retrain its main brain, which would have been too heavy for the watch's battery and processor. Instead, it stored simple "prototypes"—compact summaries of what that specific person's movements looked like for each activity. As the user confirmed more actions, the system slowly adjusted its final decision, blending the general population knowledge with the specific user's habits.

The results showed a clear and powerful progression. When the system relied only on its initial, frozen brain, it got the activity right about 62 percent of the time. This drop from the lab performance was expected, as real life is messier than a controlled experiment. However, the moment the system applied its complexity-aware adjustment, accuracy jumped to 78 percent. This improvement happened immediately, without any personal data from the user. As the users continued to confirm their activities over the twenty days, the system's personalization kicked in, and the final accuracy rose to 89 percent. This means that by combining a smart, immediate correction with a slow, steady learning process, the system recovered from its initial mistakes and became highly reliable.

Not every activity was equally easy to master. Walking and mopping were recognized with high accuracy right from the start. Washing face, however, remained the most difficult. Even after twenty days of personalization, the system still struggled with this task, getting it right only about 79 percent of the time. The researchers found that while the system could correct many of its early errors, some confusion between washing a face and other movements persisted. This suggests that the wrist sensor, while excellent at tracking arm motion, sometimes lacks the fine detail needed to distinguish between very similar hand actions. Despite this, the system proved that it could learn from a user without needing a massive amount of data or a complex retraining process.

The study also confirmed that this approach is practical for everyday use. The entire system, including the memory needed to remember a user's specific habits, took up very little space on the watch—roughly the size of a small photo file. It ran quickly, making decisions in less than a tenth of a second, and the extra battery drain was minimal, consuming about 3.5 percent of the battery per hour of active use. This efficiency is crucial; a system that drains a watch in a few hours is useless for long-term health monitoring.

The researchers were careful to note the limits of their findings. The "correct" answers used to test the system came from the users themselves, who saw the watch's guess before confirming it. This means the system might have been influenced by the user's own confidence or bias, rather than an independent observer. Additionally, the study focused on six specific activities that the user started manually, rather than a system that watches a person all day and automatically figures out what they are doing. The system also did not try to learn new types of activities beyond the six it was taught.

Nevertheless, the core discovery is significant. It demonstrates that a wearable device can start with a general understanding of human movement and then adapt to an individual's unique style through simple, human-in-the-loop feedback. It does not need to be a blank slate waiting for weeks of data, nor does it need to be a rigid, unchangeable program. By separating the immediate, smart adjustments from the slow, personal learning, the system offers a path toward smartwatches that are not just sensors, but truly adaptive companions for monitoring health in the real world. The study suggests that the future of wearable health monitoring lies not in making the computer smarter in isolation, but in creating a partnership where the device learns from the person wearing it, one confirmed activity at a time.