Refusal, not inability
“I know nothing about technology. I refuse to use it, I probably shouldn’t.” The barrier is identity as much as skill. It is a decision already made about what kind of person they are.
Healthcare moved online. It left the people who need it most behind.
Plus is an AI healthcare service for adults over 75, a group facing frequent healthcare needs and lower digital confidence. It uses voice, reminders and plain-language health tracking so someone can manage their own care without asking their children for help every time.
As healthcare services move onto digital platforms, older adults face accessibility barriers, a lack of support and social isolation, at exactly the point in life when they need the service most.
This project asked what AI could do about that, and answered it as a service rather than a feature: understand how people over 75 actually live with their health, then build something that fits into that rather than demanding they adapt to it.
Process
One question, carried through four phases.
The project ran as a double diamond over fourteen weeks, from context analysis through to a tested prototype. The research question was set at the start and everything downstream was checked against it.
How can we use AI to create support systems that promote community integration and meaningful relationships?
01 / Background
The population this fails is about to double.
761 million people were 65 or over in 2021. By 2050 that becomes 1.6 billion, moving from one in ten of us to one in six.
Mapping the factors mattered more than the headline number. Multiple illness, hearing impairment, low self-care ability, widowhood, poor economic conditions, a history of immigration all converge on loneliness, and loneliness is not a soft outcome. It returns as measurable physical decline, mental health harm and reduced cognitive functioning. A healthcare service that only handles appointments is answering a fraction of the problem.
02 / Research
Shadowing, interviews, and what people said about technology.
Shadowing and semi-structured interviews, coded into user understanding and user stories across daily activities, health and well-being, social relationships, and technology.
Help has a cost
“If I have difficulties I’m always asking for help to my children.” Every digital task spends a little of a relationship they would rather keep for something else.
They already search
“I also frequently use Google and Facebook for health-related questions and advises.” They are not disengaged from information, just unsupported in getting good information.
Problem statement
Elderly people (75+) need tailored help to use online healthcare platforms.
Who: people typically aged 75 and above, with low technological proficiency. What: healthcare services are transitioning online, which makes managing their own health independently harder, not easier. Why it matters: they need timely medical care, they want to stay independent, and they should not be marginalised by a digitalisation they did not ask for. How: use AI to give personalised support for reaching medical services, and simplify the process enough that the person stays in charge of it.
03 / The AI system
Four models, one architecture, and an honest question about the reward.
A dialogue manager handles natural language and speech, text detection reads what the camera sees, a decision-support layer learns which suggestions get acted on, and an anomaly layer watches the health data. The architecture had to stay legible to a reader who will never look at the model choice.
Framing both as state–action–reward loops forced an honest question about the reward: the system is not optimising for engagement, it is optimising for a health metric that improved or an appointment that was kept. Input runs through voice and handwriting recognition as well as touch, because the research said reading small print and typing were both real barriers.
04 / The service
One place, and a voice for when the screen is too much.
Voice assistant
Always on, and able to complete the task rather than just answer about it. For someone who has decided they are not a technology person, speaking is not learning an interface.
Camera
Trouble reading the small text on a medicine packet: take a picture and get the information back legibly. This came straight out of Fabio’s pain points.
Anomaly detection
Health tracked in easy-to-read graphs, with the system flagging when something moves in the wrong direction, so noticing does not depend on the person interpreting a chart.
Six features
Each one is the same pipeline with a different model behind it.
Every feature reads as input, analysis and output, so the architecture stays legible without the model choice. What changes between them is the method and what the person actually has to do.
05 / Testing
We took it to the park, because that is where they are.
Rather than recruiting older adults into a lab, testing went to them, with a running ProtoPie prototype on a phone, with a laptop and notes.
Going to the park changed who could take part. Nobody had to travel, register, or be confident enough to walk into a university building, which would have filtered out precisely the participants the project was designed for.
Three things came back consistently, and all three were about the interface rather than the idea: the voice pacing was right for the age group, content was findable without help, and the large-font conversation view was the feature people named unprompted. Nobody asked what the AI was doing.
Takeaways
What the project changed about how I work.
AI as material, not as feature.
The project deepened what I understand about how data is processed, selected and integrated into a design rather than bolted onto one. Working with older users strengthened the technical side and changed the brief at the same time: their needs are what made the case for inclusive healthcare design, not the technology.
For us the algorithms were close to a blank canvas, and making them accessible to people in their eighties was the hard part. We were learning the territory while designing in it, which is uncomfortable and is also where the useful findings came from.
Where I want to take it next.
The direction I want to pursue is a more structured validation process, optimising AI decision-making for accessibility, and a wider exploration of AI-driven healthcare interventions. That is the line running from this project into what I want to research.
The unresolved question is the one the reward function raises. Optimising for a health metric that improved, or a visit that happened, is defensible on paper; establishing that it holds up over months with the people it is built for needs a study this project did not run.
