Identify: LLM-based EHR screening with the evidence attached

What does Bond read in the EHR?

Bond connects to all the major EHRs, including Epic, Oracle Health (Cerner), MEDITECH, athenahealth, eClinicalWorks, NextGen, Veradigm and OncoEMR,^⁠[2] through FHIR R4 APIs, HL7 v2 feeds where applicable, or an integration partner.^⁠[1] The integrations page lists what each EHR needs from site IT. What Bond reads depends on what the site's connection exposes:

• Structured data: diagnoses and problem lists, medications and prescriptions, lab results, vitals, procedures and demographics. These settle criteria like age, a lab threshold or a current prescription.
• Clinical notes and reports: progress notes, history and physicals, discharge summaries, and pathology and imaging narratives. All five are among the eight clinical note types in version 1 of the federal USCDI data standard.^⁠[11]
• Imaging data and other documents: Identify also uses imaging data and other unstructured documents, including pathology, radiology and molecular reports.^⁠[2]

Notes matter because many criteria are never coded; the answer sits in unstructured clinical data. For a heart failure trial at Mass General Brigham, the team behind the RECTIFIER screening tool found that structured EHR data could determine 5 of 6 inclusion criteria but only 5 of 17 exclusion criteria. The rest required manual chart review.^⁠[4]

How does Bond decide whether a patient meets a criterion?

Bond's matching pipeline, described in its August 2026 technical report (an internal preprint), runs in three stages. Each language-model stage receives an evidence packet: the clinical concepts involved, their synonyms, their mappings across coding systems and their hierarchy relations, each tagged with where it came from.^⁠[3]

1. 1Relevance. The first pass asks whether the study concerns this patient's condition at all.
2. 2Criterion-level evaluation. Each inclusion and exclusion criterion is checked against the patient's record and the graph-grounded evidence, and the evidence used is kept with the decision.
3. 3Ranking. Criterion results are combined into graded ranking signals. Scoring and gating follow deterministic rules rather than free-form model output.^⁠[3]

The terminology layer lets one check catch the many ways a chart says the same thing. Bond's ClinText Graph holds 3,270,078 nodes drawn from 18 biomedical terminologies, including SNOMED CT, RxNorm, LOINC, ICD-10-CM and MeSH. Tested on 51,055 eligibility criteria from ClinicalTrials.gov, it resolved at least one exact multi-token concept in 82.0 percent of criteria, against 5.7 percent for ICD-10-CM alone.^⁠[3]

How are the criteria configured and validated?

Screening starts from the study's own inclusion and exclusion criteria. Setting them up is part of the workflow configuration covered by Bond's platform fee, and it happens inside the implementation plan, which takes 48 hours for full EHR integration depending on the EHR, IT review and interface method.^⁠[1]

1. 1

   Settle ambiguous criteria

   Phrases like "adequate organ function" need the thresholds the sponsor actually uses. The PI and study team decide those interpretations, not the software, and write them down.

2. 2

   Resolve criteria to concepts

   Bond's terminology graph maps each criterion to concepts, synonyms and codes across vocabularies.^⁠[3] State time windows explicitly so that an old lab or a resolved diagnosis does not count.

3. 3

   Check a sample against the chart

   Before relying on the ranked list, coordinators check a sample of criterion decisions against the evidence each one cites. Spend the most time on criteria that depend on dates or on combining several findings.

4. 4

   Keep measuring after launch

   Matching accuracy, screen-failure signals and coordinator hours saved are among the outcomes Bond reports, and decisions stay in the audit trail.^⁠[1] When an amendment changes eligibility, repeat the sample check on the changed criteria.

What does the coordinator see when reviewing matches?

Coordinators work from a ranked list in Bond's real-time dashboard. Each candidate comes with the criteria, the decision on each, and the note, lab or medication behind it, so the coordinator checks that evidence rather than re-reading the whole record.^⁠[1]

• Gaps the chart cannot fill. Some criteria, such as willingness to use contraception, are rarely documented. Those still have to be asked in pre-screening, by the voice and SMS/text agents in Engage or by staff.
• Audit trail. Decisions and their rationale are recorded, which helps when a monitor asks why a patient was or was not approached.

How the output fits the coordinator's workflow matters as much as the model. In a 2026 randomized evaluation of 355 oncology charts from a community practice, AI assistance raised coordinators' chart-level accuracy from 71.1 to 76.5 percent but did not shorten full chart abstraction: 37.4 versus 37.8 minutes per chart.^⁠[9] A 2014 systematic review of 79 recruitment support systems concluded that success depends more on workflow integration than on sophisticated algorithms.^⁠[10]

Reduced our chart review time significantly while improving the quality of patients we bring in for screening.

How accurate is LLM-based eligibility screening?

Bond publishes two kinds of numbers: a matching accuracy figure on its website, and benchmark results from a technical report that is an internal preprint, not a peer-reviewed paper.^⁠[1],[3] Independent results are listed alongside.

How does manual chart review compare with Bond screening?

In a 2012 prospective study at Virginia Commonwealth University's cancer center, the largest share of eligibility evaluations (35.8 percent) took 10 to 30 minutes, and more than 10 percent took 2 to 4 hours. Finding, screening and enrolling one patient took an average of 3.4 to 8.8 staff hours, depending on study phase.^⁠[8]

How is PHI handled during screening?

Screening reads protected health information. Bond signs a business associate agreement with the site.^⁠[1] Bond is HIPAA compliant and SOC 2 Type I compliant, and its SOC 2 Type II and ISO 27001 audits are underway.^⁠[2]

• Encryption at rest and in transit, AES-256 where applicable^⁠[1]
• Role-based access control with SSO support
• Audit logging
• Penetration testing and employee security training
• A Trust Center hosted on Vanta

As of September 2026, the HIPAA Privacy Rule's research provisions include two routes for reviewing records before patient contact: review preparatory to research, where the researcher represents that no PHI will leave the covered entity during the review, and a waiver of authorization approved by an IRB or privacy board.^⁠[13] Which route applies, and how it covers an outside vendor's processing, is for the site, its privacy office and its IRB to decide, not Bond. See security and IRB and HIPAA rules for patient outreach.

What does Bond not do?

Identify narrows the search. It does not replace the people who run the study.

• It does not decide eligibility. Bond ranks and explains. The coordinator and investigator make the call, and the screening visit confirms it.
• It does not see what is not in the chart. Care received elsewhere and anything never documented still need a conversation.
• It does not contact patients. Outreach is a separate step in Engage, with scripts the site configures.
• It does not obtain consent. Consent supports the process; the site and PI still obtain consent.
• It does not screen charts without data access. Sites that want to start before the EHR connection is live can begin with list-based outreach.^⁠[1]
• It does not remove the need to watch for bias. A 2026 JAMIA study of nine LLMs, using physician-validated patient vignettes, found eligibility judgments largely stable across patient identities. Homelessness produced the largest negative shift, and disparities appeared where a model had to infer behavior or resources.^⁠[12] Criteria that turn on adherence or resources are good candidates for human review.