Dementia care is a matching problem

1We are thinking about people with dementia

Our team at the Tokyo Forum Youth Session is studying how AI affects people with disabilities, and we chose dementia as our focus. Japan is the clearest place to see why: the numbers are large, the treatment window is narrow, and much of the risk is modifiable.

The scale in Japan

A Ministry of Health, Labour and Welfare (MHLW) study led by Professor Toshiharu Ninomiya (Kyushu University) estimated 4.432 million older adults with dementia in 2022 (12.3% of people aged 65+) and 5.585 million with mild cognitive impairment, MCI (15.5%). Together that is about 28% of older adults. The estimate projects 5.842 million with dementia and 6.128 million with MCI by 2040. Professor Ninomiya notes that MCI caught early can partly return to normal through lifestyle change, which is why prevention matters.

The treatment window

The two antibody drugs approved in Japan, lecanemab (Leqembi, on sale from December 2023) and donanemab (Kisunla, from November 2024), are indicated only for slowing progression in MCI due to Alzheimer's disease and mild dementia. In July 2026 APOE genetic testing was added to insurance coverage, and in August 2026 PMDA revised both drugs' warnings to recommend considering the test before treatment. The stage that families tend to miss is exactly the stage where treatment is possible.

Much of the risk is modifiable

The 2024 Lancet Commission estimates that about 45% of dementia cases could be prevented or delayed by addressing 14 risk factors across life: less education, hearing loss, high LDL cholesterol, depression, traumatic brain injury, physical inactivity, diabetes, smoking, hypertension, obesity, excessive alcohol, social isolation, air pollution and untreated vision loss. Hearing loss accounts for about 7%, LDL cholesterol about 7%, and vision loss about 2%.

Safety: people who go missing

Japanese police received missing-person reports for 17,345 people with dementia in 2025, about one in five of all 80,486 missing-person reports. The record was 19,039 in 2023. Of the 16,729 cases resolved in 2025, GPS devices were involved in only 139, and 118 of those people (84.9%) were found on the day the report was filed. Once 15 or more days pass, confirmed deaths outnumber survivors. In 2024, 491 people died, and 382 of them (77.8%) were found within 5 km of where they went missing.

2Four stakeholders, four agendas

When I think about dementia, the most important thing is the match and mismatch between what each group wants.

A product can be clever and still fail if it serves one group at the expense of another. These are the four groups I keep in view.

So I researched whether their needs actually match. Sections 3 to 6 are the evidence; section 7 sums it up.

3What people with dementia say they need

The most consistent finding is a perception gap. People with dementia report fewer needs than their family carers report for them, and the needs they do name are about company and meaning, not safety.

They report about half the needs their carers report

In the European Actifcare study of people with middle-stage dementia living at home, the person rated an average of 0.95 unmet needs, the family carer rated 1.66 for them, and a trained researcher rated 1.85. For met needs, the person rated 5.5 and the carer 8. Carers reported almost twice as many needs, in line with earlier research.

A study by Miranda-Castillo and colleagues (2013) found the same pattern: people with dementia reported fewer needs than caregivers and professionals, and the person most often named psychological distress as their top unmet need, while caregivers most often named memory (32.5%), daytime activities (16.0%) and company (13.1%). An earlier study cited in the same paper found people with dementia naming information (34.5%), company (28%) and daytime activities (25.2%), while their caregivers named daytime activities (40%), company (36.8%) and memory (36.6%).

Unmet needs matter for outcomes. In a study of 152 people with dementia living at home, the most frequent unmet needs were daytime activities (50.7%), company (39.5%) and help with psychological distress (30.9%). More behavioural symptoms, weak community networks, a younger carer and higher carer anxiety predicted more unmet needs, and people whose carers report more unmet needs are more likely to move into a care home or die.

Living alone raises every need

Part of the gap is lost awareness

Anosognosia, being unaware of one's own deficits, is estimated to affect 20% to 80% of people with Alzheimer's dementia depending on how it is measured, and up to 80% at the dementia stage in an early study. One study found it in 13.6% to 17.3% of mild cases but 44.2% to 55.8% of moderate cases; another found 39.5% of 177 people with mild or moderate Alzheimer's at baseline. A meta-analysis suggests awareness is mildly reduced at the MCI stage and severely reduced in dementia.

Design implication: at the MCI or early stage most people can still judge their own needs. That is the window to ask for consent and record what they want, before awareness fades.

Going out alone: autonomy with support

In a UK study of 12 people with dementia, 3 caregivers and 5 older people, participants said a GPS device would help them go out independently with more confidence. They were less worried about the ethics of "tagging" than academics and the media, but wanted the device to be discreet so it would not add stigma. In a Swedish study, everyone with dementia did daily outdoor activities, mostly alone. UK users also rejected devices that beep or flash, imagining an alarm going off in a supermarket.

4How much autonomy, and where people live

Japan measures autonomy directly. The "independence in daily living of older adults with dementia" scale (認知症高齢者の日常生活自立度) is assigned during long-term care certification.

National picture

In 2010, 2.80 million older adults were at rank II or above (9.5% of people 65+), about 60% of everyone with a care certification. Including rank I and people never certified, the estimated dementia total was about 4.39 million, plus about 3.8 million with MCI. The count excludes people who never applied for certification, who are often early-stage: exactly the people a wearable for early detection would serve.

Tokyo: the share living at home falls as autonomy falls

Tokyo's FY2013 survey found 66.2% of people at rank I or above and 59.8% at rank II or above living at home, with rank II or above projected to grow from about 270,000 to about 440,000 by 2025. About half of people at home with suspected dementia live alone or as a couple only. Tokyo repeats the survey every three years with municipal breakdowns; the latest uses a 1 November 2025 reference date.

Osaka City, April 2026

Korea, and gaps in the other regions

Korea estimated 954,789 people aged 60+ with dementia in 2024 (6.76%), publishes estimates down to district level on data.go.kr, and opened 256 dementia relief centers (치매안심센터) in every district by December 2019. Hong Kong has a high institutionalization rate compared with other developed economies, partly because community care capacity is limited. I did not find comparable public breakdowns for mainland China or the Philippines; teammates from those regions are better placed to find local sources.

5What families and the public fear

In the Cabinet Office survey published in 2025 (1,551 responses from 3,000 people aged 18+), the top worry about getting dementia oneself was burdening one's family physically or mentally (74.9%). When a family member gets dementia, the top worry was that they would cause trouble for people around them (46.5%). In 2025, 60.5% of respondents had met someone with dementia.

Shimane Prefecture's 2022 survey with Meiji Yasuda Life found the same: 75.4% feared burdening their family, and 61.5% wanted local government to focus on reducing family burden. In the 2023 survey, 82% of teenage respondents thought living in the community would be difficult with dementia. Japan's Dementia Basic Act took effect in January 2024 and promotes a "new view of dementia": that people can keep doing what they want and live with hope in their community after diagnosis. In a 2025 Nagasaki survey, 74.1% had never heard of that idea.

Fear of burden can work as common ground. A person who fears burdening their family may accept support that is framed as giving the family peace of mind, rather than as watching them.

6What companies and researchers are building

Most products target detection, monitoring and safety. The science behind passive sensing is getting strong.

What wearable signals can predict

Gait speed and stride regularity measured this way also track brain structure: in 6,412 UK Biobank participants they were associated with white matter lesions and the volume of frontal, hippocampal, insular and amygdala grey matter.

Phone and watch at scale

The Biogen and Apple "Intuition" study turned 126,640 app downloads into 23,004 enrolled US adults and collected up to 24 months of iPhone and Apple Watch data, including passive device use, surveys and repeated cognitive tests. It showed remote tests correlate with clinic tests, but it was stopped early when Biogen's priorities changed.

What the reviews warn

• A 2026 JMIR systematic review of 49 studies since 2020 concludes that wearable behavioural markers show promise for early risk stratification. Some headline accuracies are from tiny samples: 93.5% with 222 people, and about 94–95% with only 33 people (5 with dementia).
• A 2026 npj Digital Medicine review found 21 studies of 16 wearable EEG devices for MCI. MUSE 2 was the most used (7 of 21), followed by EPOC X (3 of 21). Support vector machines were used in 13 of 21 studies (61.9%). Accuracy was reported in 76.2% of studies and AUC in 38.1%, with wide variation in protocols.
• A 2026 JMIR Aging review screened 2,824 records and kept 17 studies on sedentary behaviour in MCI and dementia; evidence on differences in sedentary time was inconsistent.
• An earlier review of digital biomarkers from computerized memory tests found, for detecting MCI, sensitivity of 42.0–85.8%, specificity of 66.0–93.3% and AUC of 0.53–0.93.

Products and programs in 2025–2026

Acceptance is the bottleneck

• In a 2025 survey, 74.4% of older adults were interested in wearable health monitoring, but 54.9% reported anxiety, along with worries about over-surveillance.
• In interviews, older people linked smartwatches to surveillance and loss of control in their own space, and said feedback and alerts could make them doubt their own decisions.
• A 2025 meta-synthesis of 58 studies on wearables in dementia research (only 11, or 19%, were trials) found privacy concerns in 15 studies (26%). Adherence depended on whether the person could keep remembering what the device was for, and on keeping devices simple.
• A 2025 focus-group study of 31 adults aged 65+ found reluctance to adopt fall-detection devices often came from stigma, because the devices were seen as symbols of frailty.

7Matches and mismatches

Putting sections 3 to 6 together, this is what I found.

8What I realized: everyone needs a service that fits

We all need personalized services for each person with dementia. AI should not only make services better; it should make them the most suitable for each person.

The evidence shows that needs differ by person, by stage and by place: living alone versus with others, rank I versus rank III, Tokyo versus Osaka versus Seoul. A one-size product gets low adoption, the way GPS does. A good match also has to answer to all four stakeholders at once.

Examples of the autonomy tension

Fit changes with stage

9So first, maybe we need a marketplace

If the goal is the best fit for each person, the first step is a place where the person, their family and their professionals can see and choose among services: companionship, activities, information, going-out support, prevention, diagnosis, home care and respite. The difference from a directory is that the ranking is by fit to this person, not by who pays most for placement.

10Ways to build it, and who is already trying

I found several ways to create this. Each example solves one part of the problem.

None of these yet ranks services by fit to one person with dementia across all four stakeholders. That is the gap.

11How AI already combines medical data: MIT's HAIM

To match well, the marketplace needs data about the person. Hospitals already have a pattern for combining data: the Holistic AI in Medicine (HAIM) framework from MIT, published in npj Digital Medicine in 2022.

The numbers

The authors combined MIMIC-IV (de-identified records of 383,220 patients from the ICU and emergency department of Beth Israel Deaconess Medical Center in Boston, 2008–2019) with MIMIC-CXR (377,110 chest X-rays from 227,835 imaging studies). The resulting dataset had 34,537 samples from 7,279 hospital stays and 6,485 patients. They trained 14,324 models across 4 data types, 11 data sources and 12 tasks. Combining data improved results by 6–33% over single-source models. Images contributed most to chest diagnoses; time series contributed most to length-of-stay and 48-hour mortality. Each added data type helped, with diminishing returns as sources overlapped.

What HAIM collects in each data type

Each patient file also contained admissions, transfers, provider orders, microbiology cultures, medication administrations, prescriptions, intravenous and fluid inputs, sensor outputs and billing codes. The final model was XGBoost, chosen after comparing logistic regression, decision trees, random forest, multi-layer perceptron, LightGBM and TabNet.

The box HAIM left empty, "other sources" such as wearables and behaviour, is where dementia happens: at home, between hospital visits. That is the box our project wants to fill.

12What I have to think next

How do we integrate all the medical data a marketplace needs to give each person with dementia the best service?

Below is every type of data that came up in this research, grouped the way HAIM groups data, plus the two groups a marketplace adds: care context and the person's own wishes. Integrating them raises the questions I will work on next: which data each matching decision really needs, who consents to sharing what, and how to keep the person's wishes at the start of the chain.