Parkinson's AI Monitoring by Elaine Czech

Signed in as:

filler@godaddy.com

Account

The PRoblem

Building an AI Monitoring System to be Used at Home

The programme was building an AI system to monitor Parkinson's movement symptoms at home, combining wearable sensor data with camera (computer vision) data. Capturing data in a way that would help clinicians and pharmaceutical companies capture meaningful movements had not been done. However, engineers, clinicians and people with Parkinson's understand movement capture differently and each development team had its own priorities.

A gap between what a model measures and what matters to patients is a real risk in medical AI. A model can be technically accurate and still produce outputs that patients can't relate to and clinicians can't act on. Fixing that late, under regulatory and clinical trial constraints, is slow and expensive. Furthermore, bringing cameras and sensors into people's homes also raised questions of privacy and trust that a lab setting doesn't.

My Role

User researcher and patient-needs lead

I was responsible for making sure patient needs shaped how the AI was built. In practice, that meant:

co-leading 18 in-depth interviews with people with Parkinson's
leading the co-design workshops: designing and facilitating them with engineers, clinicians and patients
bringing 6 siloed development teams together and managing senior stakeholders' expectations
turning findings into requirements and a working method for the engineering teams

I worked in a team of 6. Three members focused on usability. I worked with my two line managers on co-development, and the whole team ran a joint needfinding study together.

Approach

Listen to patients first

I co-led the team's joint needfinding study: 18 in-depth interviews with people with Parkinson's about how they experience and describe their symptoms day to day.

Bring everyone together

I designed and facilitated participatory workshops with 15 AI engineers, clinicians and patients, so each group heard the others directly.

Align the teams

I brought 6 siloed development teams together around the shared findings and managed senior stakeholders' expectations as priorities shifted.

Change the requirements

The work led to renegotiated system requirements around user needs and privacy, within medical AI regulation and decentralised clinical trial procedures

Make it repeatable

I developed a design method engineers can use to keep model outputs meaningful to patients and clinically relevant, so the thinking outlasts my involvement.

Insights

No two days are the same with Parkinson's

Participants described how their symptoms changed from one day to the next. Current clinical measures only capture a snapshot, which they saw as a major limitation. To be meaningful to patients, monitoring at home has to reflect that day-to-day reality, not just a single moment in clinic.

Each group cared about something different

For people with Parkinson's, how they felt about their movement, such as whether they felt positive about their mobility; Engineers cared about precise capture and ground-truth data; Clinicians wanted to know which signs pointed to progression. A useful system has to connect all three, not optimise for one.

Privacy can be negotiable for many patients

Some participants wanted to share more data and were frustrated by the team's caution. Designing for privacy in health means respecting what each person chooses to share, not just defaulting to the most cautious option.

Impact

Requirements changed

System requirements were renegotiated around user needs and privacy, so patients' priorities are built into the system rather than added at the end.

Six teams, one direction

Teams that had worked in silos now shared an understanding of what patients needed.

Engineering steered towards patients

The workshops I facilitated steered the engineers towards outputs that are meaningful to patients and clinically relevant.

Reflection

The biggest barrier was agency. Most of the team had already made commitments to the project lead and funders, so changing how they worked or what they worked on was too difficult, even when the alternative was more interesting research or better for privacy. If I did this again, I'd bring patients' priorities in before those commitments were made, when there's still room to act on them.

What I'd keep is the involvement of patients in the co-development process. Having them in the room as the team discussed how elements of the system would be brought together helped make privacy and data security concerns more real. It also gave everyone in the room a shared reference point that wasn't any one team's agenda.

Parkinson's AI Monitoring