How to Compare Eligibility Criteria With AI While Preserving the Original Evidence

An eligibility matrix is valuable when it explains why study populations differ. Reducing long criteria to labels such as age, disease and previous treatment can improve readability while losing restrictions that matter to design and feasibility.

Fix the source scope and version first.

Record trial identifiers, document types, retrieval dates and versions. A public registry summary, full protocol and amendment are different source levels. Absence from a summary does not establish absence from the full protocol. Record completeness before comparing.

Use the original clause as the extraction unit.

Each row should include the study, wording, location, criterion type, threshold, timing, exception and uncertainty. AI can propose standardized terminology while retaining a link to the original. Check text extraction from scans and tables, particularly comparison symbols and negation.

Consider a teaching example: “Prior treatment X is permitted if at least four weeks have elapsed since the last dose and toxicity has recovered to the specified level.” Extracting only “X permitted” loses time and safety conditions. Separate the treatment, interval and recovery requirements while preserving their relationship within one clause. This invented clause illustrates a method; it is not an actual protocol requirement.

Keep AND, OR and exceptions distinct.

Requiring A and B can define a different population from requiring A or B. Excluding an untreated condition does not automatically exclude every instance of that condition. Some restrictions apply only to particular study groups. Preserve that logic rather than flattening it into a yes/no column.

In the public record reviewed for CheckMate 9LA, a criterion requires screening PD-L1 results from a central laboratory. That bullet does not specify a positive cutoff. An accurate cell says that results are required and a threshold cannot be inferred from that clause; it should not say “PD-L1-positive patients only.” Standardization needs to preserve semantic scope.

Keep missing information visible.

Useful states include “not specified in public material,” “exception applies” and “full protocol needed.” Unknown is separate from permitted and prohibited. For unclear clauses, identify a precise question, such as the starting date for an interval or whether a threshold applies at screening or before the first dose.

Clinical, statistical and operational reviewers can then consider the matrix’s intended use. Clinical review concerns disease and treatment context, statistical review the population and question, and operational review records and screening delivery. AI may identify differences; the team must judge their rationale and whether they support a design change.

Check errors that could alter the conclusion.

Look for omitted exclusions, lost time windows, reversed symbols, ignored exceptions and missing information treated as permission. Sampling should reflect use and consequences. Accuracy on a few easy clauses does not establish overall quality, and critical content entering a formal protocol requires more scrutiny than a browsing summary.

A useful deliverable has two tables: a discussion matrix and a traceable clause record. Readers can see the conclusion and find the sentence supporting it. Original research on generated citations also shows why the appearance of a reference cannot replace verification.

Use AI to reduce searching and repetitive organization while retaining wording and uncertainty. The resulting matrix can then support design discussions, feasibility work and later updates.