Can AI Detect Early Signs of Illness in People With Dementia?

AI can potentially detect early signs of illness in people living with dementia by analysing changes in sleep, movement, heart rate, breathing and everyday behaviour. Smart-home sensors and wearables can collect these signals continuously, while machine-learning systems can compare them with a person’s normal pattern and flag unusual changes. Research and early clinical programmes suggest this approach may help identify deterioration—including some infections—before it becomes obvious to caregivers. But the technology is not yet a general-purpose early-warning diagnostic system, and an AI alert should be treated as a reason for human assessment rather than a diagnosis.

The distinction matters because many early signs of illness are nonspecific. Someone who sleeps differently, moves less, gets up repeatedly at night or shows a change in routine may be developing an infection, but the same pattern could result from medication, pain, poor sleep, emotional distress or an ordinary change in circumstances. Current reviews describe promising results from passive monitoring and AI, but also report small studies, inconsistent accuracy and limited external validation in real-world homes.

How AI can spot a change that humans might miss

Traditional healthcare often relies on snapshots: a clinic appointment, a caregiver’s observation or a person’s description of how they have been feeling.

Sensors can provide something different—a continuous record of what happens at home.

A monitoring system might learn that a particular person normally:

• Sleeps for a fairly consistent period
• Gets out of bed once or twice during the night
• Moves around the house regularly in the morning
• Prepares breakfast at approximately the same time
• Has a characteristic level of daytime activity

If several of those patterns change simultaneously, AI can potentially identify the deviation.

The important word is potentially. The algorithm does not necessarily know why the pattern changed. It identifies something unusual that may justify a closer look.

Infections are one particularly interesting application

Infections can be difficult to recognise in people with dementia because symptoms may present differently or be communicated less clearly. A change in behaviour or function may sometimes be one of the first indications that something is wrong.

Researchers working with the UK’s Minder programme have investigated whether continuous home monitoring can identify such changes.

The Alzheimer’s Society reports that researchers developed a system using an under-mattress sensor to measure nighttime heart rate, breathing and behaviour. According to the researcher quoted by the organisation, an AI model learned to identify signs associated with infection, including COVID-19 and other respiratory infections. The system was subsequently being offered to people with dementia in west London for home monitoring.

This is significant because the sensor does not have to identify a conventional symptom such as coughing. Instead, the model can look for a combination of physiological and behavioural changes.

That approach is still being evaluated, however. Evidence from one research programme should not be interpreted as proof that commercially available smart-home systems can reliably diagnose infections.

What kinds of changes can AI monitor?

Several types of information can contribute to an early-warning system.

The advantage of combining signals is that a single measurement rarely tells the whole story.

A slightly elevated heart rate may be insignificant. A change in heart rate accompanied by disrupted sleep, altered breathing and reduced daytime activity could be more worthy of investigation.

Why a personal baseline is important

AI monitoring works differently from a conventional medical test.

A conventional test may compare a person’s measurement with a clinically established threshold. A smart-home AI system may instead compare today’s behaviour with that person’s own historical behaviour.

This can make the system more sensitive to individual differences.

One person may naturally sleep for eight hours. Another may normally sleep six. One may walk around the house frequently; another may spend much of the day sitting.

An algorithm trained on an individual’s baseline can potentially detect a meaningful deviation without assuming that everyone behaves identically.

Research into sensor-based dementia monitoring increasingly focuses on this type of personalised anomaly detection. Reviews have identified applications involving sleep, agitation, activities of daily living and other behavioural changes, although predictive performance remains insufficiently established for widespread clinical use.

AI may detect deterioration before a caregiver notices

This is one of the most important potential benefits.

A family member might visit on Saturday and notice that an older relative seems slightly different. But if sensors have been collecting information throughout the week, a monitoring system may already show that activity declined progressively over several days.

That could give a clinician or caregiver a longer timeline to consider.

Passive monitoring has attracted attention precisely because it can collect information without requiring the person to remember to operate a device or answer questions repeatedly. A 2025 systematic review found evidence that passive sensing can capture behavioural, physiological and environmental measurements continuously in real-world settings and may have applications in detecting acute deterioration.

But earlier detection only becomes clinically valuable if someone can respond appropriately.

Detecting a change is not the same as diagnosing illness

This is the most important limitation to understand.

Suppose an AI system detects:

“Nighttime heart rate and breathing have changed significantly from baseline.”

That does not mean:

“The person has pneumonia.”

The change could have many explanations.

Similarly, reduced movement could indicate illness, but it could also reflect a sore joint, fatigue, a change in weather or a new daily routine.

For this reason, the appropriate role for AI is generally triage and monitoring rather than autonomous diagnosis.

The technology can potentially move an important observation earlier in the care process:

Sensor data → unusual pattern → alert → human assessment → appropriate clinical action

It should not be:

Sensor data → AI diagnosis → treatment without assessment.

Dementia makes early detection especially complicated

Behavioural changes can be particularly difficult to interpret in dementia because the condition itself can affect sleep, activity, mood and daily routines.

A change therefore needs context.

Research into sensor technologies for dementia has found early promise in monitoring behavioural symptoms, sleep and activities of daily living, particularly when multiple data sources are combined with machine learning. However, a 2026 systematic review found only 17 eligible studies in this area and noted that most were small and observational, with inconsistent accuracy and limited evidence that results generalise between settings.

This means researchers still need to establish important questions such as:

• How large must a change be before an alert is meaningful?
• How long should the change persist?
• How frequently do false alarms occur?
• Does performance remain reliable in different homes?
• Does an alert actually lead to earlier treatment?
• Does earlier intervention improve outcomes?

Those are much harder questions than simply demonstrating that an algorithm can identify a pattern in research data.

Smart sensors can monitor more than health measurements

Early illness detection does not necessarily require sophisticated medical hardware.

A smart home can potentially observe changes in ordinary behaviour.

For example, a person who is becoming unwell might:

• Spend longer in bed
• Visit the kitchen less often
• Stop preparing meals
• Move more slowly
• Use the bathroom at unusual times
• Become unusually active during the night
• Spend less time outside their bedroom
• Show a sustained reduction in overall activity

Researchers have studied motion sensors, pressure mats, cameras, inertial sensors and other technologies for detecting changes in movement and activities of daily living. A 2026 systematic review found potential for motion-based indicators of dementia-related cognitive decline, while emphasising the need for larger longitudinal datasets and greater standardisation.

In practice, a combination of small changes may be more informative than one dramatic measurement.

AI can also monitor changes associated with dementia itself

There is another, related application: detecting cognitive or functional decline rather than an acute illness.

Researchers are investigating whether AI can analyse gait, speech, eye movements, sleep, activity patterns and other digital biomarkers to identify subtle cognitive changes.

A 2026 systematic review of wearable technology included 49 studies and found associations between disrupted sleep, fragmented circadian rhythms, irregular activity and poorer cognitive outcomes. Machine-learning and deep-learning models showed promising classification performance in some studies, but the authors highlighted short monitoring periods and limited external validation as important limitations.

This is a different question from detecting an infection.

Early illness detection asks: “Has something changed that may indicate an acute health problem?”

Dementia detection or progression monitoring asks: “Are there persistent patterns that may indicate cognitive or functional change?”

The technologies can overlap, but the clinical questions are different.

False alarms are a serious problem

An early-warning system has two ways to fail.

It can miss a genuine deterioration, or it can raise an alert when nothing is seriously wrong.

The second problem may sound less serious, but repeated false alarms can place additional demands on caregivers and potentially cause alarm fatigue.

Recent reviews of anomaly-detection systems in dementia care specifically identify false positives, privacy and user compliance as continuing challenges.

An effective system therefore needs more than high laboratory accuracy. It needs to work reliably in ordinary homes, with different layouts, routines, internet connections, devices and users.

It also needs an alert system that tells the right person the right information at the right time.

Privacy becomes more important as monitoring becomes more powerful

The same sensors that can help detect deterioration can also reveal intimate details about someone’s life.

A system monitoring sleep, movement, room occupancy and physiological measurements may create a detailed behavioural record.

That makes data governance a central part of the technology rather than an afterthought. A systematic review of remote monitoring for Alzheimer’s disease and related dementias identified privacy, security, usability and user-centred design among the major issues that need stronger evaluation.

Before adopting such a system, families and care providers should establish:

• What information is collected
• Whether processing happens locally or remotely
• Who can access the data
• How long it is retained
• Who receives health alerts
• What happens when the system fails
• Whether the person being monitored understands and accepts the arrangement

More data is not automatically better care.

Where the technology stands now

The evidence supports a promising but still developing picture.

Smart-home sensors and wearables can continuously capture information that conventional appointments may miss. AI can analyse that information for deviations from established patterns. Research has demonstrated potential for identifying behavioural changes and, in some programmes, early signs associated with infection.

At the same time, systematic reviews continue to identify small study populations, inconsistent accuracy, limited external validation and gaps between experimental systems and routine clinical care.

That makes it too early to regard AI monitoring as a replacement for conventional clinical assessment.

The real opportunity is earlier human attention

The most useful question may not be whether AI can “diagnose illness” at home.

It is whether AI can notice a meaningful change early enough for a person to investigate it.

That is a more realistic and potentially valuable role. A sensor can notice that sleep has changed. An algorithm can recognise that the change is unusual for that individual. A monitoring service can alert a caregiver or healthcare professional. The human can then consider the wider context and decide whether assessment is necessary.

If research can demonstrate that this chain consistently leads to earlier recognition and better outcomes—while controlling false alarms, cost and privacy risks—AI-assisted monitoring could become an important additional layer of dementia care at home. For now, its strongest role is as an early-warning aid, not an independent medical diagnosis.