Anticipating Protocol Amendments

Aika's first draft already anticipates half of what amendments add later

Building on protocol costs, we have seen that protocol amendments are one of the most expensive problems in clinical trial development. Many of them simply add a new eligibility rule, a requirement the trial team did not think of when they first wrote the protocol. We asked a simple question: could our clinical AI, Aika, have written those rules from the start? Can amendments be anticipated and cut easily?

Why amendments cost so much

Eligibility criteria decide who can join a trial. When investigators realize a rule is missing, they file an amendment, triggering a slow, costly chain of internal review and regulatory approval.

The key insight: many added criterias might be predictable. They reflect needs that could have been anticipated but were overlooked in the initial version of the protocol.

What our in-house study on amendments found

We compared Aika's first-draft eligibility criteria against the criteria that real amendments later added.

We looked at 47 real clinical trials.
Those amendments added 126 new eligibility criteria in total.

Using semantic similarity, Aika's first draft already covered a substantial share of what the amendments would later add. To check that this reflected genuine anticipation rather than generic trial language, we ran a negative control, we repeated the test using Aika's drafts from unrelated trials:

At a similarity threshold (ADD@85), Aika matched 84% of added criteria versus 75% for the negative control (p=0.001).
The advantage over control remain significant (p=0.001) across stricter (ADD@90 and ADD@95) thresholds too.

That gap matters. It shows Aika was anticipating what a specific protocol would need, not restating boilerplate.

Changed criteria: an early signal

Amendments don't only add rules; they also reword existing ones. Here the picture is softer but pointing the same way:

Aika matched about 55% of modified criteria, versus roughly 50% for the unrelated-trial control.
This difference was not statistically significant (p = 0.238).

So we treat modifications as a tendency, not a proven result, a hint that Aika may anticipate reworded rules too, worth testing in larger samples.

What this means, and its limits

If half of amendment-added rules can be caught at the drafting stage, a meaningful share of protocol amendment cost and delay may be avoidable, before a single patient is ever screened.

A few honest caveats:

Aika drafts; a human always reviews. The process can be paused or stopped at any point.
Results depend on the similarity threshold we chose.
Findings need confirmation in larger, independent samples the model never trained on.

Even with those caveats, the direction is clear: protocol amendments are not entirely inevitable. However, there are tools and ways to anticipate and cut protocol amendments. Better first drafts, supported by clinical AI, could prevent many amendments from ever being needed.