AI-driven pre-screening for intermediate AMD

A historic pre-screen failure rate of roughly 60% on an intermediate-AMD study — six in ten patients a coordinator reviewed turned out ineligible, discovered only after the manual work of reading their chart was already done. The eligibility signals that would have ruled a patient out earlier sat in unstructured clinic notes and OCT imaging, neither of which a chart reviewer could search at the volume a multi-site trial needs. The criteria themselves were also layered: inclusion and exclusion language in the protocol, then an imaging read on top of that for the patients who passed the first filter — two different skills, applied sequentially, that no single fast pass through a record could replicate.

Deployed against real-world records from a leading European academic eye centre, with source data resident with the site throughout and every record de-identified before the pipeline ever reads it — GDPR-compliant by construction, not by a downstream redaction pass. See data residency for healthcare AI for why "the data never leaves the site" has to be an architectural property of where inference happens, not a clause in a data-processing agreement. For a multi-site academic trial, that also matters operationally: a pattern that only works by exporting patient records to an outside system is a pattern most sites' own ethics and IT review would stop at the door, regardless of what it promises on screening accuracy.

A two-phase pipeline matching the two-layer criteria above. Phase 1 uses generative AI to read de-identified clinical records against the protocol's inclusion and exclusion criteria, producing a shortlist from the full patient population rather than whichever records a coordinator had time to open. Phase 2 takes that shortlist and runs OCT image segmentation against the imaging-specific criteria the records alone can't answer — drusen extent and other intermediate-AMD imaging markers a text-only read would miss entirely. Every patient Phase 1 or Phase 2 flags carries the specific evidence that triggered the flag: the note passage, or the segmented image region, not a bare eligible/ineligible label. The two phases are separate, registered tools rather than one combined model: Phase 1 is a language-model read of structured and unstructured text, Phase 2 is an image-segmentation model tuned specifically for OCT scans and the imaging markers this protocol's criteria actually reference. Keeping them separate means each can be validated, and if necessary replaced, against its own ground truth — a text-eligibility benchmark for Phase 1, a clinician-annotated imaging set for Phase 2 — rather than one opaque end-to-end score covering both.

Because every flag carries its source evidence, a clinician reviewing the shortlist is checking a specific claim against a specific passage or image region, not re-reading the full chart to confirm an opaque AI verdict — the same evidence-attached discipline described in AI hallucination mitigation in regulated settings, applied here to an eligibility judgement rather than a generated dossier claim. The research plan's own targets made the validation bar explicit rather than aspirational: at least double the historical screen rate, and pre-screen failure below 50%, both measured against the same baseline the roughly-60% failure rate came from. See validating AI systems against GxP for the broader standard a system making eligibility-relevant judgements is held to.

Per the research plan: a target of at least twice the historical screen rate, and pre-screen failure brought below 50% from a roughly-60% baseline — GDPR-compliant throughout, with source data never leaving the site. Coordinators review a shortlist with evidence attached instead of opening every chart cold, which is what makes hitting that target possible without adding headcount to the screening effort. The two-phase split does the same work a senior coordinator's instinct does — rule out on the record first, then spend the imaging review only on patients worth that closer look — at a scale no manual process could sustain across a multi-site enrolment.

The site's own research team runs pre-screening against new protocols directly. deidentify · cohort_matching · image_segmentation — a pre-screening pattern extensible to other indications and imaging modalities, not rebuilt per study.