One path from cognitive screening to daily training
HAII's AI reads brain scans at 88% accuracy, but its three screening apps ran as three separate products. Users signed up three times, and every app ended at its own score with nothing after it. I tested all three with 30 adults aged 50 to 75, then connected them under one account so each app hands people to the next.
Lab Intern, Yonsei HCI Lab
HAII ($8M raised, K-FDA cleared)
July to September 2025
Research, UT, IA, UX/UI direction
01 | overview
Older adults want early screening. They abandon tools that feel clinical and unclear. Brain Check ran 5 minutes, Saemirang 20 and DACI 35, and none of them handed the user to the next one.
Alzguard keeps the three services and connects them, so one account carries a person from screening to training.
3 apps, 0 links
1 account, 3 hand-offs
Finishing one service now opens the next.
Eye-tracking and voice, taken at home.
Sessions picked from the last result.
Nearby centers and verified information.
02 | research
Over 3 million Koreans are projected to live with dementia by 2050, and the only trusted way to catch it early is a PET scan. That scan costs more than $3,000 and takes six months to schedule, which is why digital screening exists, but the digital option added up to an hour spread across three disconnected apps.
So the problem was never test accuracy. It was the missing path between the tests.
The reliable test is out of reach for most people.
So the question was never whether digital screening should exist. It was what happens once someone opens it.
An hour of testing, split across three apps that never hand the user on.
Every app ends the same way. A score, and no reason to open anything else.
Reflecting upon this core issue, I started to question:
💭 How might one workflow carry a user across three tools without losing a single sense of purpose?
Accuracy was already proven in the clinic, so the open question was whether the same tests earn trust at home. Five phases traced where that trust broke, moving from usability testing to segmentation, personas, context and market. Each phase had to end in an IA or UI decision instead of a report, so the next phase always started from a design problem.
usability testing
Instead of counting who finished, I watched for the exact moment each person stopped feeling sure.
Moderated in person, one participant at a time, on the live products.
Three apps, three breaking points
Across all three apps people were not avoiding the test. They were avoiding the verdict. Clinical language and ranked scores turned a preventive check into a judgment they could not interpret. So trust needed three things at the same time: clear guidance, emotional reassurance and a human voice on results.
Users were eager for early dementia screening but lost confidence when tests felt clinical or confusing. Building trust required clear guidance, emotional reassurance, and a human tone that balanced accuracy with empathy.
Identifying key user groups
To act on that, I segmented twice, first by clinical stage and then by emotional response to being tested. KLoSA data on 6,057 older adults puts 40% in the high-worry group while only 20% are proactive accepters, so half of the people who worry most never test. Those Anxious Avoiders became the primary target, so the product now reduces fear before it measures cognition.
Medical Condition
Preclinical/Early AD
Beta-amyloid accumulation begins,
but cognitive function is still preserved
Psychological Response
Anxious Avoider
High levels of worry
but avoidance of testing
🎯 Core target users
Individuals for whom brain changes have started but are still within the golden window, who delay testing due to anxiety.
Persona development
Screening is never one person's act. Someone is tested, someone urges them and someone follows up. Three personas covered those roles, a 55-year-old at risk, a 20-year-old family member and a 55-year-old caregiver, and each one fails at a different point. So the product later split into a user view and a caregiver view.
“Mom, you need to start taking care now.”
Found the app in an ad, and never uses it herself. Needs something she can hand over, not just recommend.
“I tried it too. You should try it.”
Tests herself, then passes it on. A peer trying it first is what makes it safe to try.
“My mother had dementia. How do I prevent it?”
Avoids checkups, and does not trust her own hands on a phone. Will start, but only once someone else has.
Nobody in this chain starts alone, so every ending has to hand the person to someone.
Segmentation explained who avoids testing. Context explained why the results could not be trusted. A cognitive score is only as stable as the room it was taken in, so I mapped the hospital, the home and the social attitudes around both. Each one produced a single design rule that went straight into onboarding.
Secondary Research highlights
Korean Dementia Society data shows 81.2% agree that early dementia treatment should not be delayed, and 78.3% want the government to expand early diagnosis support. But only 27.7% know the term MCI and 42.91% do not know it is the golden window for prevention, so agreement runs far ahead of recognition. So Alzguard explains MCI at the moment of the result, where the question actually occurs, instead of in a brochure.
Screening for MCI is not one service, it is a chain of separate ones. Every break in that chain costs a trip, a guess, or a decision the user will not make alone.
Competitive Analysis Highlights
I benchmarked six competitors on early screening, AI personalization and family integration to see whether anyone had closed that gap. Each covered detection or training and rarely both, and almost none were free to start. So Alzguard took the open position, covering the full journey with a free entry point that removes the first excuse to postpone.
03 | Defining the Service
Users arrive already believing something different. So the opening changes with the path, and everything after it stays identical.
Mapped every entry point, ran a best and worst case through each, then wrote the counterplan the opening had to carry.
Four openings, one product. Which opening a person gets is settled before the app is ever installed, by where they met Alzguard.
Service Delivery Channels
As part of defining the service ecosystem, we identified all possible access routes for Alzguard and categorized them into four channel types: daily-life touchpoints, corporate/community networks, welfare-facility channels, and online/offline healthcare institutions.
Developed Business map canvas
Getting someone to try it once is only half the job. After the first test, the app keeps teaching them about brain health, and a family member or caregiver can stay connected. Results come back right away, and from there a person can share them or book a hospital visit.
Value Proposition
Easy access to health insights
Trust-driven experience
Instant results, sharing, booking
👥
Customer
50s–80s health-conscious adults
Avoiders who prefer non-hospital options
Channel
🤝
Relationship
QR-based onboarding
Ongoing education
and caregiver connection
Information architecture
Brain Check, Saemirang and DACI are still three separate apps. I didn't merge them. What changed is how each one ends. It used to stop at a score, and now it sends the person on to the next step. I laid out all three flows with ChatGPT and Claude to test different structures, and HAII adopted the final one into its product direction.
04 | solution
Each step answers something the research found. The test runs at home, so people who would never book a hospital visit for it can still take it. The result is written in plain words, because a score like “Top 50%” left people asking whether they were sick.
05 | Validation
01
"Learning Healthcare Design Beyond the Interface"
Understanding clinical rigor and real-world validation in medical services
Working inside a clinical product taught me that validation and reliability are judged before usability. Because regulation and risk decide what can ship, I learned to design inside those limits instead of around them. And treating flows as research artifacts rather than deliverables is what let me test structure instead of screens.
