Regard

The diagnostic engine is a rule system, and the learned parts are features around it. Regard's technology page describes diagnostic logic written and reviewed by physicians, published in numbered versions, and deterministic: the same chart always produces the same assessment. Its worked example, thrombocytopenia in an adult inpatient, is a threshold on the lowest platelet count with severity bands and exclusions for heparin and ITP treatment.

The learned components are the language features: Max, a chat agent in the chart, and note drafting, both built on OpenAI's GPT-4 according to the company's 2023 release, plus AI that helps the clinical team write new rules. How the note parser detects negation and timing is not stated. The company markets AI-powered diagnosis, but the recommendations it counts and the revenue results it reports come from the rules.

The marketing rests on an absolute promise. The technology page opens with "The technology that reads every chart, so no diagnosis is missed", the home page with "Never miss a diagnosis", and a 2024 release says the platform ensures "no critical information is missed". Taken at face value, that promise invites a hospital to skip the review for omissions this kind of product needs.

The controls underneath are real: the clinician accepts or declines each recommended diagnosis, every recommendation links to the labs, vitals, medications and notes behind it, and the rules report insufficient evidence separately from a condition ruled out. The company counts 12,993,284 accepted recommendations on the home page and 15M+ on the Clinical Notes page, with no count of recommendations made, so the decline rate cannot be derived.

The approach and a supplier are named, with version discipline published for the rules and not for the language model. The technology page describes the pipeline: chart ingestion, an in-process clinical language parser that marks negation, subject and timing, a concept library that maps each hospital's codes to one vocabulary, and a diagnostic rule library written by physicians, reviewed by two clinicians and published in numbered versions with every prior version archived.

The company's 2023 release names OpenAI's GPT-4 for Max and note drafting. Three things go unpublished: which language model and version run today, how the parser was built or tested, and the training data for any learned part.

A model provider is named and the rest of the chain is not. The company's August 2023 release names OpenAI's GPT-4 as the model behind Max and the note drafting features, both still sold; the diagnostic rules are built in house. No host, subprocessor list, or retention and training terms for data sent to OpenAI are published, and the product pages' statement that all data stays inside the customer's environment leaves open how the language model features reach OpenAI.

Named deployments with outcome figures, all vendor published. Sentara Health's pilot case study reports a 17 percent increase in CC and MCC capture, up to 4x return per user and 74 percent voluntary adoption among hospitalists; Penn Highlands reports appeal wins rising from 20 to 80 percent; the home page cites $50 million or more in revenue at an Arizona system, $9.3 million in denials prevented in North Carolina and a 20 percent reduction in queries in Pennsylvania; the Clinical Notes page cites two hours saved per clinician per day at a multi-site academic medical center.

The diagnostic headline has less behind it: "90%+ diagnoses accuracy validated against clinician review" carries no method or sample, and PubMed lists no study of Regard. The outcomes reported are financial and throughput, not whether patients were diagnosed correctly. The company lists more than 150 hospitals.

General assurances. Product pages say all data stays inside the customer's environment with end-to-end encryption, full audit logging and role-based access. The privacy policy (updated June 2026) covers website comments, cookies and embedded media only and says nothing about patient data. Nothing states how long chart data is retained, whether customer data trains or tunes any model, or what reaches OpenAI when the GPT-4 based features (Max, note drafting) run. No safety engineering is published, meaning how a language model output is checked and how a safety event is handled.

Compliance is claimed without the document behind it. The Clinical Notes page says Regard is "HIPAA-compliant"; no business associate status is stated, and the privacy policy covers only the website. The ONC certification is a separate framework: it tests the software's export and authentication functions and says nothing about business associate terms. The product reads full charts across a population for HCC and screening work, beyond the patient in front of a clinician, so the agreement's scope matters.

Recognized certifications are named in the company's own material without the artifacts. The Mid-Revenue Cycle and HCC Capture pages say Regard is "HITRUST-certified and SOC 2 Type II compliant", and the Clinical Notes page says SOC 2 Type II certified; the HITRUST assessment type, the report periods and the scope are not stated, and no trust center is published.

The ONC Health IT Module certification (15.04.04.3192.Rega.01.00.0.240502, 2 May 2024) is an outside check of the software's security features, and for multi-factor authentication it relies on the EHR's own login through SMART on FHIR. The quality system is ISO 9001.

No device claim is made, and no regulatory position is stated for a product that recommends diagnoses. The ONC certification (2 May 2024, criteria b10, d1, d5, d12, d13, g4 and g5) is an interoperability and security certification and does not include the decision support interventions criterion; it is not FDA clearance, and FDA's 510(k) database lists none.

The design points toward the non-device decision support criteria, since each recommendation shows the evidence and logic behind it for the clinician to review, and the recommendations are not time critical alarms. The marketing runs further ("Never miss a diagnosis", "AI-powered diagnostic support").

A named review process governs every piece of diagnostic logic, and no evaluation across patient groups is published. Physicians write each condition rule, two clinicians review it, only approved versions run, and earlier versions are archived for audit; the company says the logic is updated against guidelines, peer review and real-world performance data. The process is set out in enough detail to be examined.

What is missing is any result: no performance by age, sex, race, language or care setting, and no measure of how often rules fire or miss in each group. The same recommendations feed the Mid-Revenue Cycle and HCC Capture modules, and nothing says how clinical accuracy is monitored apart from coding yield.

Mechanisms exist and nothing stands behind the output. Every recommended diagnosis carries the chart evidence behind it, a clinician can reject it on that evidence, and the rules separate insufficient evidence from a condition ruled out. The Clinical Notes page reports "90%+ diagnoses accuracy validated against clinician review" with no method, sample or denominator, and no omission rate is published, which is the error that leaves no trace in the chart. No warranty, indemnity or remediation commitment is published.

Named record systems with read access documented, and the write path not described. The technology page names native support for Epic and Cerner, plus generic FHIR servers, and shows chart loads over Epic FHIR R4 triggered when a clinician opens the chart; the app embeds in the EHR through SMART on FHIR, and the company announced an integration with Microsoft Dragon Copilot in 2025.

Recommendations, evidence and the drafted note appear inside the EHR workflow, but how the note and problem list reach the record is not described, and the pages name no marketplace listing. Named production sites include Sentara Health, Cedars-Sinai and Banner Health. The ONC certification covers electronic health information export (FHIR single patient and bulk), not the EHR integration.

Residency is stated and the processing path is left open. The product pages say all data stays inside the customer's environment and that Regard deploys inside the existing security perimeter, while the ONC certification lists the product as Regard SaaS Solution v.1. The technology page describes per-customer configuration and chart loads triggered when a clinician opens the chart. No hosting provider, region or tenant model is named, and nothing says where the GPT-4 based features process data.

Pricing is private, but part of the commercial posture is public. The company frames value as return per user (4x at Sentara), offers a free HCC gap analysis to size the revenue opportunity before contract, and states go-lives of six to ten weeks with a dedicated team. The unit of charge, tiers and implementation fees are not published. Several modules are sold on coding and denial outcomes, yet the company does not say whether any part of the fee varies with captured revenue, severity or collections.

Inpatient first, with coverage named and validation behind part of it. The product began with hospitalists and leads with inpatient measures (CC and MCC capture, DRG movement). The company says 250+ conditions are detected across specialties and announced expansion to new specialties toward all admitted patients in March 2026, and HCC Capture and Screening reach ambulatory populations. The example rule is scoped to adult inpatients, but no age, population or specialty exclusions are published for the library.