AI can learn a person’s normal daily routine by analysing patterns collected from sensors over time rather than judging individual events in isolation. In dementia care, motion detectors, door sensors, appliance sensors, bed sensors and other connected devices can record when someone moves between rooms, sleeps, gets out of bed or carries out familiar household activities. Machine-learning systems can then identify recurring patterns and establish an individual baseline against which later behaviour can be compared.
The important point is that AI does not need to decide what a “normal” routine should look like. It can learn what is normal for that particular person. Someone may wake at 6 a.m., make tea shortly afterwards, spend most of the morning in the kitchen and go to bed at 9 p.m.; another person’s routine may look completely different. In dementia care, the potential value lies in detecting meaningful departures from an individual’s established pattern and bringing those changes to the attention of caregivers or clinicians.
Why learning the individual baseline matters
Daily life is naturally variable.
A person may sleep later on a Sunday, skip breakfast one morning or spend an afternoon with family. A system that treats every difference as a warning would generate too many alerts to be useful.
AI can instead examine behaviour over a longer period and distinguish recurring routines from occasional variation.
Research using smart-home sensors has demonstrated that longitudinal sensor data can be used to identify repeating patterns in daily routines. One study analysed 50–100 days of data from older adults and used clustering techniques to capture changes in how consistent their routines were over time.
This creates a personalised baseline:
What does this person’s ordinary day usually look like?
Only after establishing that baseline does the more difficult question become possible:
What has changed?
What information can AI use?
A smart-home system can combine multiple streams of information.
| Data source | What it can reveal about routine | Example of a meaningful change |
|---|---|---|
| Motion sensors | Movement between rooms and activity levels | Less daytime movement than usual |
| Door sensors | Entries and exits | Repeated nighttime door opening |
| Appliance sensors | Interaction with household equipment | A familiar kitchen activity occurs less often |
| Bed sensors | Bedtime, waking and nighttime movement | Increasingly fragmented sleep |
| Physiological sensors | Measures such as heart rate or breathing | A sustained change alongside behavioural changes |
| Wearables | Steps, movement and sometimes location | Reduced activity compared with the person’s baseline |
| Combined sensor data | Relationships between activities | Several parts of the person’s routine changing together |
The advantage of combining these measurements is context.
A single door opening tells an AI system very little.
A pattern showing repeated nighttime exits, disrupted sleep and increased movement over several nights tells it considerably more.
How AI builds a baseline
The process can be thought of as several stages.
1. The system observes ordinary life
Sensors collect data without requiring the person to record everything manually.
For example, a motion sensor might register movement in the hallway, while an appliance sensor records activity in the kitchen.
Over time, the system begins to see repeated sequences.
2. It identifies recurring patterns
Machine-learning algorithms can group similar periods together.
Perhaps the person usually has a relatively consistent morning pattern:
Bedroom → bathroom → kitchen → living room
The exact timing may vary, but the sequence occurs frequently.
AI can learn these recurring relationships without requiring somebody to manually program every activity.
Research into smart-home monitoring has shown that machine-learning approaches can analyse sensor observations to learn individual activity patterns and identify unusual deviations.
3. It allows for normal variation
A useful system cannot expect every day to be identical.
The person’s routine may shift depending on the day, visitors, appointments, weather or other circumstances.
This is why longer-term data matters.
A change that occurs once may be ordinary variation. A change that persists for several days or weeks may be more significant.
4. It detects deviations
Once a sufficiently detailed baseline exists, the system can look for behaviour that differs substantially from the person’s established pattern.
This might include:
- Sleeping at unusual times
- Spending much less time in the kitchen
- Moving around the home less frequently
- Increasing nighttime activity
- Leaving the home at unusual times
- Taking longer to complete familiar activities
- Changing the sequence of everyday activities
The AI does not necessarily know why the change occurred.
It identifies the change.
AI can learn routines without cameras
One of the important developments in this area is the use of passive environmental sensors.
These can monitor aspects of activity without recording video or audio.
MinderCare, the NHS-linked dementia monitoring service developed through Imperial’s UK Dementia Research Institute programme, uses discreet sensors around the home to monitor daily routines, movement and other signals. Imperial says the system’s devices do not record sound or video.
That means the system can potentially recognise that someone has moved through a hallway or opened a door without creating a video record of what they were doing.
This can provide useful information while reducing some of the intrusiveness associated with cameras.
For a broader explanation of the sensors involved, see From Smart Beds to Door Sensors: The Technology Behind AI Dementia Care at Home.
The difference between an event and a pattern
This is one of the most important concepts in AI-assisted dementia monitoring.
Suppose a door sensor records that someone left the house at 2 a.m.
That is an event.
It might be completely harmless. Perhaps the person was visiting the bathroom, responding to a noise or briefly stepping outside.
But suppose the system has learned that the person normally sleeps continuously between 10 p.m. and 6 a.m., and it now observes repeated nighttime exits over several days.
That is a pattern.
The second situation is potentially more useful because the system can place the event in the context of the person’s established behaviour.
This is why AI is particularly suited to large amounts of longitudinal sensor data: the important information may lie in relationships between hundreds or thousands of individual observations.
AI can recognise changes in everyday activities
Daily activities can also provide clues about functional changes.
Researchers have explored smart-home systems capable of recognising activities such as preparing meals, beverages and snacks using fixed sensors in the home. A 2024 pilot study found that activity duration could differentiate healthy controls from people with subjective cognitive decline or mild cognitive impairment in one of the studied meal-preparation tasks, although the study was small and conducted in a simulated smart-home environment.
This illustrates an important possibility.
Instead of asking someone during an appointment:
“Are you still managing to prepare meals normally?”
a future system might provide objective information about how frequently and how consistently particular activities are being performed.
That does not mean the AI can determine whether someone has dementia from kitchen activity.
It means technology may provide another way of measuring functional behaviour.
AI can learn the relationship between activities
Daily routines are not just collections of isolated actions.
They contain sequences.
Someone might typically:
- Get out of bed.
- Go to the bathroom.
- Prepare breakfast.
- Take medication.
- Sit in the living room.
A disruption in one part of that chain may affect everything that follows.
AI can analyse these relationships.
For example, if breakfast preparation normally occurs shortly after waking but suddenly stops occurring, the system may identify the change even if the individual still moves around the house normally.
Research into smart-home monitoring has explored both large-scale daily patterns and smaller activity sequences, allowing algorithms to examine changes at different time scales.
The baseline can become increasingly personalised
The longer a system observes someone’s routine, the more information it potentially has about individual variation.
This matters because dementia care is highly personal.
A generic rule such as:
“Getting out of bed twice at night is abnormal.”
would be unsuitable for many people.
A personalised system could instead recognise:
“This person normally gets out of bed once between midnight and 7 a.m., but has been getting up repeatedly over the past four nights.”
The second approach is more informative because it is based on the person’s own history.
That is one reason researchers describe these measurements as potential digital biomarkers: continuously collected data may provide objective indicators of changes in behaviour or function that are difficult to capture through occasional assessments.
What happens when the routine changes?
The next stage is interpretation.
MinderCare’s current model provides an example. Sensors transmit information to a secure platform, where data is analysed for trends and patterns. A monitoring team then reviews the information and can escalate concerns for further clinical review or intervention.
This creates an important division of responsibility:
Sensors collect information.
AI identifies patterns.
Clinicians interpret what those patterns might mean.
Care teams decide what action, if any, is appropriate.
That distinction prevents AI monitoring from being mistaken for autonomous medical diagnosis.
A change in routine can have many explanations
An unusual pattern does not necessarily indicate dementia progression or illness.
A person may become less active because:
- They have a temporary illness.
- They are experiencing pain.
- Their medication has changed.
- A family member has altered their routine.
- They have had visitors.
- They are sleeping poorly.
- They are experiencing anxiety or another emotional difficulty.
- Their environment has changed.
The AI may be able to detect the behavioural change, but it generally cannot determine the cause from sensor data alone.
This is why human context remains essential.
AI can combine routine data with physiological information
Routine monitoring becomes potentially more informative when behavioural data is combined with physiological measurements.
For example, a system might detect:
Reduced daytime activity + disrupted sleep + altered heart rate + changed breathing pattern.
Each signal has limitations.
Together, they may give a clinical team a stronger reason to investigate whether something has changed.
MinderCare’s home-monitoring system combines environmental sensors with an under-mattress sensor that can measure sleep disturbances, movement, breathing and heart rate.
The objective is not simply to collect more measurements. It is to build a richer picture of how the person is functioning.
AI does not need to understand the whole person
There is sometimes a misconception that an AI system must understand a person’s life in the same way a caregiver does.
It does not.
Its role can be much narrower.
It may simply need to recognise that:
“The current pattern is substantially different from this person’s usual pattern.”
That can still be valuable.
A nurse or caregiver can then ask the questions the sensor cannot answer.
This division can make the technology more practical: machines handle continuous observation and pattern detection, while people provide context, empathy and judgement.
How much data does AI need?
There is no universal number of days that establishes a person’s routine.
The amount of information required depends on the complexity of the behaviour being studied, the variability of the person’s routine, the sensors available and the algorithm being used.
Research has demonstrated routine detection using data collected over weeks and months. One study examined 50–100 days of motion-sensor data to identify recurring patterns, while other dementia-monitoring research has followed participants for substantially longer periods.
The key principle is that more data is useful only if it represents meaningful everyday behaviour.
A sophisticated algorithm cannot compensate for poor sensor placement, missing information or an environment that does not reflect the person’s normal life.
Privacy remains part of the equation
Learning someone’s routine requires collecting information about their life.
Even without cameras, a system may reveal when someone sleeps, cooks, leaves home or moves around the house.
That information can be sensitive.
Research into smart-home systems involving people with dementia has therefore emphasised user-centred design, privacy, trust and usability alongside technical performance.
A good monitoring system should make clear:
- What is being collected
- Why it is being collected
- Who can see it
- How it is protected
- How long it is retained
- What happens when the system generates an alert
The objective should be useful monitoring, not surveillance for its own sake.
The biggest challenge is deciding what counts as meaningful
Learning a routine is technically different from understanding its significance.
An AI system may become very good at predicting what someone normally does tomorrow morning.
That does not automatically mean it can determine whether a deviation is clinically important.
This is one of the central challenges in dementia monitoring.
Research has shown that smart-home sensor data can capture everyday activities and routine patterns, but studies continue to examine how reliably those measurements translate into clinically meaningful information.
The most useful systems therefore need to minimise two opposing problems:
Too many alerts: caregivers become overwhelmed by changes that are harmless.
Too few alerts: meaningful changes are missed.
Getting that balance right may be more important than simply increasing the number of sensors.
Where this technology is heading
Current research is moving toward richer, more continuous forms of monitoring.
Imperial’s 2026–2027 MinderCare Radar project, for example, is testing radar sensing in 30 homes to add physiological and movement information to the existing monitoring system. The research aims to evaluate whether contactless sensing can operate reliably in real homes and contribute to earlier identification of health changes.
At the same time, researchers are exploring how AI can combine environmental, physiological and wearable data into increasingly detailed profiles of everyday functioning.
This development is closely connected to the broader question explored in AI and Smart Sensors Could Help Detect Illness Earlier in People Living With Dementia.
It also connects with Dementia Care Moves Into the Home as AI Monitors Sleep, Movement and Daily Routines, which looks at the wider shift toward continuous home-based monitoring.
The goal is not to predict every action
The most useful AI dementia-care systems may not be those that attempt to understand every detail of a person’s behaviour.
Their value may come from something simpler:
knowing what is usual, recognising when it has changed, and helping the right person notice that change.
That approach turns months of otherwise disconnected sensor readings into a personalised picture of everyday life. Research already shows that machine-learning techniques can identify recurring activity patterns and deviations, while current NHS-linked programmes are testing how these capabilities can be integrated into clinical care.
The unresolved question is what happens after the AI detects the change. If caregivers and clinicians can reliably distinguish meaningful deterioration from ordinary variation—and respond appropriately—the technology could become a useful layer of dementia care at home. If the systems generate unreliable alerts or collect more information than care teams can meaningfully use, additional data will not necessarily improve care. The long-term value of AI routine learning will therefore depend less on whether machines can recognise a person’s habits than on whether those insights lead to earlier, more personalised and genuinely useful support while protecting the person’s privacy and independence.