A clinical trial can follow a patient for ten years. The record of that patient's experience, though, is usually built from a series of isolated moments.

At enrollment, the study knows a lot. Baseline measurements, medical history, a careful conversation with a coordinator who has time set aside for exactly this. At the first scheduled follow-up, it knows a little more. Then six months pass. Then another visit window opens, and the study asks the patient to fill in everything that happened in between.

The interesting story is what happened inside those gaps. And that's the part we're worst at capturing.

A decade, one participant

Picture a single hypothetical participant moving through a ten-year follow-up study.

Year 1. They're engaged. The study is new, the procedure is recent, and appointments feel important. They answer the phone. They complete the questionnaires. When a coordinator asks how they've been feeling, the answer is right there, because it's what they've been thinking about.

Year 3. Life has moved on. The procedure isn't something they consider daily anymore. A coordinator calls periodically and asks them to remember how the past several months have gone. They give an honest answer, shaped by whatever the last few weeks happened to look like.

Year 6. Maybe they've moved. Maybe they've changed primary care doctors, developed an unrelated health condition, switched phone carriers, or simply become harder to reach. Their study record still exists. But it's increasingly a stack of check-ins with wide, quiet spaces between them.

Year 9. The study that was once a significant event in their life is now something they did almost a decade ago. The trial still needs to know how they're doing. The participant still matters enormously to the data. But the relationship, in any practical sense, has thinned to a name in a database and a phone number that may or may not still work.

Nothing went wrong here. No one made a mistake. This is what a well-run long-term follow-up study looks like.

The problem isn't volume

It's tempting to frame this as a data collection problem, as though the fix is more instruments, more timepoints, more questions per timepoint. But most long-term studies collect plenty. The problem is structural: the patient's story has to be reconstructed from snapshots.

A questionnaire at Year 5 tells you how someone answered at Year 5. It doesn't tell you what happened between Year 4 and Year 5. It doesn't tell you whether a score that dropped four points did so gradually over eleven months or fell off a cliff in March when something specific happened. It doesn't tell you that the participant had a rough six weeks, recovered, and answered on a good day.

Two participants with identical Year 5 scores can have had completely different years. The instrument can't see the difference. And we compensate by asking patients to be the memory system; to recall, summarize, and self-report across intervals no one recalls accurately.

This also compounds. Attrition and missing data are usually treated as statistical problems to be handled downstream, but a lot of them start as relationship problems upstream. A participant who hasn't heard anything meaningful from a study in eight months is easier to lose than one who has an ongoing, low-friction connection to it. By the time someone is formally "lost to follow-up," the loss usually happened earlier and quietly.

Reframing 83 calls

Here's the number that reframed this for me: over a ten-year follow-up protocol, a single participant might be contacted 83 times.

The default way to read that number is as workload. Eighty-three tasks. Eighty-three dials, eighty-three voicemails, eighty-three attempts to reach someone at a number that hasn't been updated since 2019, spread across a coordinator's caseload of hundreds of patients. Under that framing, the goal is efficiency: complete the 83 tasks faster and cheaper.

But there's a second way to read it. Eighty-three calls is also eighty-three opportunities to maintain one continuous relationship and one continuous record. Same number, completely different design problem. The first framing optimizes for closing out touchpoints. The second optimizes for the thread that runs between them.

Most follow-up infrastructure is built for the first reading. Each contact is a discrete event that produces a discrete record, and whatever connective tissue exists lives in a coordinator's head or in a free-text note that nobody queries.

What continuity would actually mean

This is where conversational AI stops being a staffing story and starts being a continuity story.

The interesting question isn't "can a system place calls instead of a coordinator." It's: what if the trial didn't have to remember the patient from one follow-up to the next, because the conversation never fully disappeared?

A participant who mentions in Year 3 that they've started walking again isn't starting from zero in Year 4. A participant whose answers have been drifting in one direction across four consecutive contacts isn't just a Year 7 data point. A participant who used to answer on the first try and now takes five attempts is telling you something before any instrument records it.

Each interaction contributes to a developing picture instead of a freestanding entry: how they're feeling, what's changed since last time, patient-reported outcomes, engagement patterns, and sometimes the context surrounding a change in responses, the part that usually evaporates between calls.

None of this replaces human judgment. It creates room for it. When a response pattern warrants real attention, that's exactly when a coordinator should be on the phone, and exactly when they should have more context than a name and a due date. The routine touchpoints that consume most of the capacity are the ones that don't need a person to hold them together.

The decade perspective

Ten years is a long time to ask someone to remember a study.

It's an even longer time to ask a study to remember a patient. Across a decade, coordinators turn over, sites change hands, systems get migrated, and institutional memory quietly leaks out of the protocol. The participant's continuity is expected to survive all of it, held together by whoever happens to be making the call this quarter.

We've spent a lot of effort trying to improve recall on the patient side: better instruments, shorter intervals, clearer prompts, diaries, reminders. That work matters. But it's addressing the wrong end of the problem.

The future of long-term follow-up may not be about asking patients to remember more. It may be about building systems that forget less.