Clinical trial recruitment takes 18 months on average. Pretrial.ai compresses feasibility to 48 hours.

Pretrial.ai

Feasibility on statistically validated synthetic patient cohorts.

build a cohortopen dashboard
the problem

A sponsor cannot know whether a protocol recruits until sites have screened real patients against it. That answer arrives months into a study, after the sites are contracted and the budget is committed. When the answer is no, the protocol is amended and the clock restarts.

A synthetic cohort that reproduces the real population's distributions, comorbidity structure, and lab value correlations answers the same question before the first site is opened. The only thing that makes it worth trusting is proof that it matches, which is why validation is a first class service here and not a footnote.

94.7%fidelity
Validated

The single figure a clinical operations director reads to decide whether a cohort is usable. Above 90 it is. Below 70 it is not.

services
servicefunctionmethodport
SyntheCoreGenerates a synthetic cohort from seed data and screens it against the protocol's criteria.CTGAN, conditional sampling8001
FidelityCheckProves the synthetic cohort matches the reference population, or shows exactly where it does not.KS, Jensen-Shannon, correlation delta8002
FHIR-LinkEmits the validated cohort as FHIR R4 bundles a hospital network or CTMS can ingest.R4 Patient and Observation, HTEST tagged8003
what the score is made of
marginal agreement
0.40
Two sample Kolmogorov-Smirnov per continuous feature, total variation distance per categorical feature. This is what inclusion criteria are written against.
shape agreement
0.35
Jensen-Shannon divergence across the full support, base 2. Catches a distribution that occupies the right range with the wrong shape.
dependency structure
0.25
Correlation matrix comparison. Catches a generator that reproduced every variable alone and lost the relationships between them.

The weights are published because the number has to survive scrutiny from a biostatistician who did not build it. The full method is in docs/synthetic-data-methodology.md.