Guardant Health

Learned models make the calls in the screening test; the rest of the business is a laboratory. The Shield instructions for use state that the result comes from two methylation classification models, a MAF-regression model and a logistic regression model, each with a pre-defined threshold, over thousands of methylation features in cell-free DNA; without them there is no result.

The therapy selection and recurrence tests are sequencing assays with bioinformatics, and InfinityAI, the company's branded AI, is described as technology used to develop report features and to build data products for drug developers, not a standalone product. The models are the engine of a core module inside a sequencing laboratory business.

The screening test's limits and next step are published in full; the AI-derived report features have none. Shield is for people 45 and over at average risk; a positive result should be followed by colonoscopy, which Medicare covers; it is not a replacement for diagnostic or surveillance colonoscopy; and the labeling lists who should not use it (a personal or family history of colorectal cancer or adenomas, inflammatory bowel disease, hereditary syndromes, a positive result on another screening test in the last six months). The result is a binary call from two classifiers with pre-defined thresholds, and a batch is reported only if its controls pass.

The missing part sits with the report features the company says are powered by InfinityAI: they are reported as a professional service outside FDA review, and nothing published says what they may not be used for or when a clinician should disregard them.

The screening models are named and their changes are regulated; their training is not described. The Shield instructions for use set out the workflow (methylation partitioning, targeted enrichment, sequencing on Illumina NovaSeq) and name the two classifiers that decide the result, a MAF-regression model and a logistic regression model with pre-defined thresholds. A change to the classifier requires FDA approval, and one was approved in July 2026, so a buyer can see when the model changed.

What is not published is the data the classifiers were trained on, or anything about the models behind InfinityAI, which the company describes only as purpose-built, domain-specific AI.

A short chain at the model layer: the classifiers are Guardant's own and run in its laboratories. The approved screening model cannot be changed without an FDA supplement, and nothing published suggests a patient's data goes to an outside model. The named parties sit around the sample: QIAGEN extraction instruments and Illumina sequencers in the laboratory workflow.

The infrastructure is not enumerated. The annual report mentions cloud data centers and outside infrastructure vendors without naming them, no subprocessor list is published, and the recipients of licensed data through InfinityAI are described only as biopharma partners.

A prospective multicenter screening study supports the approved test, and it shows weak detection of precancer. In ECLIPSE, summarized in the Shield instructions for use and published in the New England Journal of Medicine in 2024, the test detected 83.1 percent of colorectal cancers (54 of 65) and 13.2 percent of advanced adenomas (147 of 1,116), with 89.6 percent specificity for advanced neoplasia, against colonoscopy in average-risk adults. The American Cancer Society's 2026 guideline update added an FDA-approved blood test as a screening option.

The therapy selection tests' companion diagnostic claims each rest on the drug trial behind them, reviewed through PMA supplements, and the recurrence test holds Medicare coverage for colorectal cancer. The studies were company sponsored.

The company discloses that patient data becomes a commercial data asset. The HIPAA notice of privacy practices (last updated 9 October 2025) says Guardant keeps a database of de-identified health data from many patients, including de-identified DNA and RNA data, and may use and share it for any lawful purpose without further notice or compensation, and that its own or external researchers may access identifiable data to develop research projects. The InfinityAI page describes the product built on it: a real-world data library derived directly from Guardant's clinical testing, offered to drug developers through a self-service exploration platform, custom reports, trial matching and data licensing. The web privacy policy keeps personal data only as long as necessary and allows de-identified data to be used without restriction.

Genomic data cannot be made anonymous in the way a name can be removed, and nothing published sets retention for specimens or sequence data. Safety engineering, meaning guardrails, hallucination mitigation and how a safety event is handled, is not published either.

The company states its status as a covered entity and publishes the notice that goes with it. The HIPAA notice of privacy practices for US patients (last updated 9 October 2025) sets out how Guardant, as the laboratory receiving a prescribed test, uses and discloses protected health information, patients' rights to access, amend and receive an accounting of disclosures including those by its business associates, and its uses for research and de-identification. No contracting instrument for ordering providers, such as a data protection addendum stating how subcontractors that handle patient data are bound, is published.

Controls are described with outside testing behind them, and no attestation is claimed. The annual report's cybersecurity section says the program takes into account NIST CSF, ISO 27001 and 27002 and HIPAA, and that the company periodically performs penetration testing and engages third parties to assess its data security practices, under a chief information security officer who reports to the board twice a year. It does not claim ISO 27001 certification or a SOC 2 report, and no trust portal is published.

Each regulatory status is identifiable, and the company separates approved from unapproved. Shield was approved under PMA P230009 on 26 July 2024 for colorectal cancer screening in average-risk adults 45 and over; supplement S010, approved 22 July 2026, changed its classification model to a methylation-only workflow, removing the fragmentomic and somatic components.

Guardant360 CDx (P200010, approved August 2020) has added companion diagnostic indications through supplements, most recently in September 2026, and Guardant360 Liquid CDx was approved on 19 May 2026 (P250027). Guardant Reveal, Guardant360 Tissue and the hereditary panel are laboratory developed tests, and the company labels them as not FDA cleared or approved; the InfinityAI-derived report features are labeled as professional service outside FDA review.

Subgroup performance for the screening test is published on the company's own site, with its counts. The Shield instructions for use tabulate cancer sensitivity, advanced adenoma sensitivity and specificity by sex, age band, race and ethnicity from the ECLIPSE study. Specificity falls with age, from 95.5 percent at 45 to 49 to 80.9 percent at 70 to 79 and 75.5 percent at 80 and over, so older patients get more false positives; by race, cancer sensitivity was 90.0 percent in Black participants (9 of 10) and 75.0 percent in Asian participants (3 of 4), with specificity of 92.1 and 84.4 percent. The small cancer counts in most subgroups mean the sensitivity comparisons carry wide uncertainty, which the tables make visible.

Nothing comparable is published for the InfinityAI-derived report features or data products.

Error rates are published with their denominators, and nothing stands behind them. The Shield instructions for use give cancer sensitivity of 83.1 percent (54 of 65), specificity for advanced neoplasia of 89.6 percent (5,982 of 6,680) and advanced adenoma sensitivity of 13.2 percent (147 of 1,116), and the product page states that Shield detects 55 to 65 percent of stage I cancers, misses 87 percent of precancerous lesions, and that one in ten people with a negative result may have a precancer a colonoscopy would find. A buyer and a patient can hold the test to those figures.

No guarantee, indemnity or remediation attaches to a result, and the report features built with InfinityAI carry no published error rate.

Named two-way integrations with major oncology record systems, running in production. The EMR integration page describes ordering Guardant tests and receiving results directly in the patient chart, EMR-integrated order sets, reflex testing triggered by clinical events and Epic MyChart messaging, and cites a 2022 partnership with Epic (the six-week go-live figure assumes the site already uses Epic Aura) and a 2023 integration into Flatiron's OncoEMR. It reports more than 1,200 integration partners and links a US Oncology Network case study of automatic EMR-based result delivery. Guardant also has a listing in Epic's Showroom.

A laboratory service in named US laboratories, with the data platform unnamed. Specimens are shipped to Guardant's CLIA-certified, CAP-accredited laboratory in Redwood City, California, with other laboratories in Palo Alto, San Diego and Long Island City, New York, and results return through the Guardant portal or the ordering EHR. The annual report says applications and data run on a mix of on-site systems and cloud data centers, with outside vendors managing parts of the data centers, none of them named.

Where sequence data and the de-identified data library are stored, and whether any processing leaves the United States, is not stated.

One test is priced; the rest reach buyers through payer contracts. The Shield site states a cash price of $1,495 per test where insurance does not fully cover it, zero out of pocket for eligible Medicare fee-for-service patients, and coverage by UnitedHealthcare, VA Community Care and TRICARE, and it offers cost estimates before testing. No price is published for Guardant360, Guardant Reveal, the hereditary panel or the InfinityAI data products, which are sold through payer contracts and biopharma agreements.

The $1,495 Medicare rate that Shield carries under its advanced diagnostic laboratory test status is payer context rather than a price, though here it matches the published cash price.

Coverage is named for each product, with regulatory validation behind part of it. Shield is for colorectal cancer screening in average-risk adults 45 and over, validated in ECLIPSE; Guardant360 CDx and Liquid CDx are for advanced solid tumors, with each companion diagnostic claim validated for its drug; Guardant Reveal monitors recurrence after treatment for colorectal, breast and lung cancer and response in late-stage disease, as a laboratory developed test with Medicare coverage in colorectal cancer. The InfinityAI trial matching and data products serve drug developers without published validation.