Triomics

Reading whole oncology charts is the product. OncoLLM, a family of oncology-tuned models, reads every note, report, scan and fax in a patient's record and produces structured, cited answers that power three products: Prism screens every scheduled patient against every open trial, Symphony prepares visit summaries and draft pre-charting notes, and Harmony abstracts registry and research data.

The company's founding point is that the structured part of the record was already tractable and the value sits in free text. With the models removed there is no product left to sell, only an integration into the record system.

The oversight structure is described and one part is missing. Every output is a proposal a person checks: Prism delivers a prioritized, cited worklist of trial candidates that a coordinator confirms; Symphony's pre-chart notes are 'a starting point, not a finished product' for the physician; Harmony auto-populates registry cases so 'the work that remains is verification' by registrars, with 94 percent of fields accepted as-is in the company's published work.

Each value carries a citation that opens the source line in the chart, and audit trails are native. The company describes this as 'cited, auditable, never a black box', a claim about the basis of each output that the citations back up. What is missing is the rule for what happens when evidence is thin or absent, and any threshold at which a patient is withheld from or added to a worklist. The failure a coordinator cannot see is the eligible patient the model never surfaces.

The approach is named and published; the models, their bases and versions are not. OncoLLM is described as a family of oncology-tuned models, a constellation of specialized models from about 3 billion to 72 billion parameters working agentically at the patient level, some trained from scratch and some fine-tuned from open-source models, then tuned further on each institution's own records with human feedback and deployed as a customized model per partner.

The PRISM trial-matching method is peer reviewed (npj Digital Medicine, 2024), and a preprint describes the approach. What is not published is which base models are used, the version running at a given customer, or how and when institution-tuned models are updated.

The model layer is the company's own, which keeps that part of the chain short. OncoLLM is described as a family of models the company trained from scratch or fine-tuned from open-source models, tuned further on each institution's records and deployed per institution, rather than calls to an outside model service; the company reported that its extraction matched leading commercial models at a fraction of the cost. No hosting provider, subprocessor list or support-access arrangement is published, and the open-source base models are not named.

Peer-reviewed and independently evaluated performance, with the methods available to read. The PRISM trial-matching system was published in npj Digital Medicine (October 2024), reporting 95 percent accuracy in matching patients to trials against expert review. An ASCO 2025 abstract co-authored with a leading academic medical center reports enterprise screening with OncoLLM producing 40 percent more trial matches and a 30 percent lift in trial accruals.

The Regenstrief Institute has published an evaluation of AI-enabled trial screening in real-world oncology workflows, and a preprint on structuring unstructured records reports 94 percent accuracy across more than 100 data elements against expert curation and 95 percent acceptance by expert abstractors. Named deployments include Memorial Sloan Kettering, Mount Sinai Tisch Cancer Center, a Connecticut deployment and more than ten NCI-designated centers. Each study comes from a single setting or a small number, and much of the work is co-authored by the company.

Categorical positions are published on how patient data enters the product, and the training position is stated in the direction of use. Evaluations run on a synthetic-data sandbox or a de-identified record by default, and real data requires a signed business associate agreement first.

The company says its models are tuned on each institution's own records with human feedback and deployed as a customized model per partner, and Symphony learns each physician's note style; the privacy page points to a General Privacy Policy on HIPAA. Nothing published says whether tuning on one institution's data improves the shared base models, what is retained after tuning, or what happens to an institution's model when the agreement ends, and no retention schedule is stated.

Business associate status is stated in practice and a privacy document addresses it. Every product page invites a prospective customer to use a synthetic-data sandbox or 'sign a BAA to upload your own data', and the demo form offers evaluation on a de-identified record, so real patient data requires a signed agreement first.

The company's privacy page links a General Privacy Policy explaining how Triomics complies with HIPAA and similar health data laws, alongside an online privacy policy, state notices and a trust center. The agreement itself is not published, so its terms and whether it binds subcontractors cannot be read before contracting. Prism screens every scheduled patient against open trials before anyone has consented, which rests on institutional review board approval or a waiver as well as the business associate agreement.

A trust center exists and no attestation is named on the public pages. The company's privacy page links a trust center 'to find information about our controls or request security documentation', and invites security researchers to report vulnerabilities to security@triomics.com, acknowledging valid reports though no paid bug bounty is offered.

No SOC 2, HITRUST or ISO 27001 certification is named on the home, product or privacy pages, and the trust center's contents are not publicly readable. The product reads complete oncology charts continuously across whole cancer center populations, so the vendor's own environment, release practice and staff access to deployed instances are what an attestation would examine.

No device claim is made and the products are scoped to research operations, documentation and registry work, with the clinician's review built in. Prism produces a cited candidate worklist a coordinator confirms, Harmony abstracts registry data for registrars to verify, and Symphony summarizes the chart, drafts pre-charting notes, answers clinicians' questions about the record and lists missing data before a visit, each output linked to the line in the chart that produced it.

Symphony is the closest to clinical decision support: a clinician relies on its account of disease status, lines of therapy and biomarkers, and the citations are what let the clinician check it. No regulatory position is stated.

Published evaluation and traceability built into the product, short of a comparison across patient groups. The company's research program includes peer-reviewed validation of trial matching, an equity-framed study reporting an increase in treatment trial accruals through automation, a study of matches that did not convert to enrollment, and benchmarks developed with academic cancer centers; every output carries a citation to the source line and audit trails are native.

The domain risk is documentation bias: trial matching and abstraction surface what the chart documents, so patients with thin or fragmented records are matched less often regardless of eligibility, and that population overlaps with those trials already under-enrol. No model card, formal governance framework, or comparison of candidates surfaced by the system against manual screening by patient group is published.

Published error rates with method, from the company's own studies. The PRISM paper in npj Digital Medicine (2024) reports 95 percent accuracy matching patients to trials against expert review; the preprint reports 94 percent accuracy across more than 100 data elements against expert-curated benchmarks and 95 percent acceptance by expert abstractors across more than 15,000 data points; and the company reported that in work with Medical College of Wisconsin researchers OncoLLM found 90 percent of eligible trial patients, which puts the missed-patient rate at about 10 percent in that study.

Each figure can be tested against the method published with it. They describe studied configurations rather than every institution-tuned deployment, and no warranty, remedy or service commitment attached to a missed patient or a wrong registry field is published. The failure that matters is the eligible patient never surfaced.

Named two-way integrations with major record systems, one verifiable in the record vendor's marketplace, running in production. The company states it is approved on the Epic App Marketplace and lives inside Epic, OncoEMR and iKnowMed, ingesting HL7, FHIR and CCDA as well as PDF and image files, and that outputs land in the chart rather than a separate portal: trial candidates arrive as a worklist or record-system message, Symphony's summaries and draft pre-charting notes open inside the record, and Harmony auto-populates cases in registry systems of record. Eligibility is re-evaluated as the chart changes. Production use at Memorial Sloan Kettering and Mount Sinai, both Epic sites, is announced by the institutions themselves.

The shape of the deployment is described; where it runs is not. The company deploys a customized model for each partner institution and delivers outputs inside the institution's record system (Epic, OncoEMR, iKnowMed) rather than a separate portal. Nothing published names a hosting provider or region, says whether inference runs inside the institution's environment or in the company's own infrastructure, describes tenancy, or states what the deployed system sends outbound.

No price is published, and the shape of the offer is discoverable. Triomics sells to cancer centers through a demo, as three products on one platform (Prism for trial matching, Symphony for visit preparation, Harmony for registry and quality work), and separately to life sciences companies for enrollment, feasibility, site identification and trial design. Project-specific Harmony tuning is described as taking a few weeks per data dictionary.

The product covers cancer registry abstraction and oncology data work: Harmony curates NAACCR data elements from unstructured records, implements NAACCR, CoC and QOPI requirements, and autopopulates cases in registry systems for registrars to verify, alongside precharting and visit preparation for clinic teams.

Coverage is deliberately confined to oncology and supported where it is named: NCI-designated comprehensive cancer centers (Memorial Sloan Kettering, Mount Sinai Tisch Cancer Center), academic medical centers in published studies, a community oncology practice case study for visit preparation, and life sciences sponsors for feasibility and prescreening. Workflows are named per team: research coordinators for trials, physicians and medical assistants for pre-charting, registrars for abstraction. Outside oncology nothing is claimed.