PersonalHealthAIPWA

Epilog

AI caught a drug interaction. The neurologist confirmed it.

A seizure and aura tracker built for a family member. Log an event in seconds, find patterns across months, and surface insights that change a clinical conversation.

Its AI analysis caught a drug interaction their GP had missed: a fiber supplement was interfering with their anti-epileptic medication. They brought the finding to their neurologist, stopped the supplement, and seizure activity went down.

Understanding the constraint

What it took to define the right problem

Someone you love has epilepsy. You watch them try to log an event after a seizure or aura, still foggy, motor control off, cognitive function not fully back online. The event is over, but the aftermath is real. That's not a user story you write on a whiteboard. It's something you understand by being in the room.

Three constraints came out of that proximity:

Log in seconds while still recovering.

Events get logged in the aftermath, when brain function is still impaired. If the logging flow requires concentration, the data doesn't get captured. This wasn't a performance goal. It was a clinical one.

Find correlations without being a data analyst.

The calendar and insights views needed to surface patterns visually, without an interpretation step. A list of events tells you what happened. A calendar tells you what the pattern is.

Serve the caregiver relationship too.

Data collected for personal use is only half the value. The other half is the conversation between a patient and their neurologist. Designing for that conversation was a first-class requirement.

AI can build a health tracker in an afternoon. Knowing which three constraints actually matter requires sitting in the room where the problem lives.

Open the add flow and walk through logging a seizure. Every input is a single gesture: no typing, no scrolling, no decisions that require concentration.

Knowing what to kill

Why the best design decision was deleting a feature

Early on, I built a medication reminder system. Push notifications at dosing times, confirmation flows, the whole pattern you'd expect. It didn't survive first contact with real use.

The problem wasn't the reminders. It was the assumption. Most of the time, medication is taken on schedule. Building a system that demanded confirmation twice a day created friction on the 95% of days when everything was fine. The user stopped engaging with the app entirely.

So I stripped it out and inverted the model: assume adherence, only capture deviations. A "Missed Dose" event type replaced the entire notification system. One tap when something goes wrong, silence when it doesn't.

AI can generate a notification system in minutes. Recognizing that the right move is to delete it requires judgment that only comes from watching someone actually use it.

Tap the + button and look at the event types. "Missed Medication" is a first-class event, not a setting buried in a menu. That's the entire medication tracking system.

The outcome

How the AI caught something a doctor missed

The user's seizure activity had been increasing over several weeks. They'd been logging consistently: seizures, auras, missed doses, sleep data from their wearable. They ran the AI analysis.

The analysis flagged something unexpected: a potential interaction between psyllium husk, a fiber supplement their GP had prescribed for digestive issues, and their anti-epileptic medication. Psyllium husk can interfere with drug absorption when taken at the same time. The GP hadn't considered this. It's not their domain.

The user brought the finding to their neurologist. The neurologist confirmed the concern. They stopped the supplement. Seizure activity decreased.

An AI tool, built by one designer, caught something a doctor missed. Not because the AI was smarter than the doctor. Because it had the right data, in the right context, and surfaced the right question.

The AI didn't replace clinical judgment. The designer's job was knowing what data to collect, how to frame the output, and when to get out of the way. That's the part AI can't do for you.

Open the Insights tab and switch to "AI Analysis." Tap "Analyze my data" to see the kind of output the tool produces. The demo uses a curated dataset, but the structure mirrors real results.

Head to the Export tab and tap "Export PDF." The generated report is structured for a neurologist visit: findings, medication history, and event timeline in a format that respects their time.

Try the live demo ↗