DrFirst

The models are the differentiator on top of a large network business, and the company draws that distinction more honestly than most.

The base is infrastructure. Electronic prescribing, controlled substance prescribing, secure clinical messaging, benefits checking and a national medication history network are transaction routing and connectivity, built over twenty five years on industry messaging standards. That business would exist and function with no model in it, and the chief executive describes joining what he would then have called an electronic prescribing company.

The models are what the company now sells on and they do specific, difficult work. SmartSig parses unstructured free text sig instructions into discrete components, resolves synonym variation into a facility's own terminology, infers missing elements, and supplies alternative drug identifiers where a medication cannot be matched. Natural language processing and machine learning are both named, the technology is patented, and the output is measurable: automatic sig mapping rising from 26 percent to 86 percent at one health system across 300,000 medications is a direct measure of what the model contributes.

The company's own line on this is worth recording because it is unusually blunt for a vendor and it is a claim a buyer can test. Asked about machine learning, the chief executive said they do not believe that slapping a chatbot on top of existing products counts as artificial intelligence. A company that names the cheap version of the claim and disowns it is inviting scrutiny of its own.

Graded B rather than A because a substantial share of what customers buy is the network rather than the intelligence, and higher than the data platform records in this index because here the models carry the product's current commercial argument.

The company argues publicly for human oversight of its own product, and one question that follows from the design goes unanswered.

The framing is explicit and it is published under a title that states the position: keeping a human in the loop, because artificial intelligence for medication management requires clinical safety checks. A vendor writing that about its own technology, rather than about the category in general, is setting a standard it can be held to. The chief executive's rejection of superficial artificial intelligence claims is consistent with it.

The architecture supports the framing in an important structural way. SmartSig does not make a clinical decision. It converts unusable data into usable data, and the usable data then triggers the record system's own interaction and allergy checking. So the model's output feeds the existing safety machinery rather than substituting for it, and the clinician performing medication reconciliation reviews the result as they always would. Compared with a model that issues a clinical determination, the autonomy here is a data transformation with a human reviewer downstream.

The unanswered question is provenance at the point of review. When SmartSig infers a missing sig element or resolves an ambiguous drug match, nothing published states whether the clinician reconciling that medication can see which fields were inferred by the model and which came from the source record. That distinction is the whole of the oversight in practice: a pharmacist who knows a frequency was inferred will check it, and one who assumes it came from the pharmacy will not.

Graded B.

The mechanism is described concretely and the model behind it is not, with patents providing partial public provenance.

The functional description is better than most in this lane. SmartSig converts free text sig data into programmable discrete components, codifies terminology into each facility's standard, supplies alternative drug identifiers for best case matching, and provides details for incomplete or uncommon sigs. Natural language processing and machine learning are both named as the underlying methods. The problem being solved is stated with a figure, 66 percent of records in the largest medication history database missing essential sig information, and one performance number is published, 93 percent of imported prescription information accurately translated. A reader finishes with an accurate picture of what the software does and roughly how well.

The patents are genuine provenance. A patented and patent pending technology is described in public documents that a technically literate reader can retrieve, which is more verifiable disclosure than a marketing page, and the company cites the patent status repeatedly.

What is absent is everything below that. No model architecture, no training corpus description, no evaluation methodology behind the 93 percent, no error analysis, no versioning and no update cadence. The phrase clinical grade artificial intelligence appears throughout and is a positioning term rather than a technical one, and the company never defines what makes an engine clinical grade.

Versioning matters here more than usual, because a parser that changes behaviour alters what enters patient charts, and nothing describes how a customer learns that it has.

Graded C.

The technology's own provenance is documented through patents and acquisitions, and the model's inputs and components are not.

What is traceable is corporate and legal. The sig technology is described as patented and patent pending across multiple public announcements, so the method is documented in filings a reader can retrieve and attribute, which is a stronger form of provenance than a marketing description. The Myndshft acquisition in 2024 is documented with the acquired platform's own capabilities stated, so a buyer knows the medical benefit authorisation capability arrived from outside and can assess it separately. Twenty five years of continuous operation under a single founder establishes that the core network is in house rather than assembled.

The data supply chain is visible by category if not by name. Medication history is sourced from pharmacies, payers and the national prescription network, which is inherent to the product and understood by any buyer in this market.

What is absent is the model itself. No framework, no third party component, no pretrained model and no external natural language processing service is named or ruled out, and no bill of materials exists. Nothing states whether the parsing models are built entirely in house or incorporate licensed components, which matters because it determines whether prescription text reaches a party other than DrFirst.

The training corpus is the larger gap and is recorded on the stewardship axis: a sig parser learns from real prescriptions, and nothing describes whose.

Graded C.

An unusually strong body of named, quantified deployment evidence, with no peer reviewed or independent validation located.

The deployment side is close to best in class for this index. Five health systems are named with before and after figures rather than testimonials. Hackensack Meridian Health reports automatic sig mapping rising from 26 percent to 86 percent across a stated sample of 300,000 medications, and publishes that on its own newsroom rather than leaving it to the vendor, which makes it customer attested. Scripps reports medication history available for 96 percent of patients over 65 with 85 percent of home medication sigs enhanced. Monument Health reports 93 and 82 percent. Covenant HealthCare reported recapturing 15 percent productivity per shift in the first month, valued at roughly $650,000 monthly. Emory Healthcare reports a 13 percent improvement in best possible medication history. An academic medical centre reported a 20 percent improvement in prescription first fill rates using the population health capability.

Those are operational measures with denominators in at least one case, drawn from different institutions on different record systems, and they point consistently the same way.

What is missing keeps this from an A. No peer reviewed publication and no independent evaluation were located. The model accuracy figure that does exist, 93 percent of imported prescription information accurately translated, is vendor stated with no method, sample or error analysis. And no study connects improved sig completeness to reduced adverse drug events, which is the clinical outcome the whole argument rests on, even though medication reconciliation is a national patient safety goal and the causal chain is plausible.

Graded B.

The company's own published statistic reveals the scale of what it holds, and nothing describes the terms on which it holds it.

That statistic is the tell. DrFirst states that an estimated 66 percent of records in the nation's largest medication history database are missing essential sig information. Producing that figure requires analysing the national database at population scale, which establishes both the reach of the company's data position and its willingness to study the corpus. It is a useful and honest disclosure about the state of the data, and it is silent about the governance around it.

Nothing published states whether prescription and medication history records inform model training, whether learning is confined per customer or pooled across the network, what de identification standard applies, what retention governs medication history, or whether a health system can decline secondary use. For a model whose entire function is learning how prescribers and pharmacies write sig instructions, the training corpus is necessarily real prescriptions written for real patients, and the provenance of that corpus is unaddressed.

The patent position implies documented methodology exists, since patents describe technique, and it says nothing about data rights.

The HITRUST i1 certification covering all medication management solutions is genuine assurance about controls over the data, which is the handling half of this axis. It does not address whether the data should be repurposed, which is the stewardship half.

Graded C: strong certified handling, a data position disclosed by implication rather than described, and no published position on training use.

A validated certification covering the whole product line, announced with its scope enumerated, and no published contractual layer.

The certification is HITRUST i1, achieved in March 2025, and the announcement states that all medication management solutions are covered rather than leaving scope to be discovered. Scope enumeration is the detail that matters most on a multi product catalogue, and it is precisely what several other records in this index leave open. HITRUST is the framework health organisations ask for by name and an i1 validated assessment is externally performed rather than self attested.

The honest qualification is the tier. The i1 assessment is the moderate assurance level covering a curated control set aimed at current and emerging threats, sitting below the r2 assessment in rigour and duration. It is a real credential and it is not the strongest one available.

The data flows are extensive and unambiguous. The company operates a national medication history network, ingests prescription records from pharmacies and payers, writes structured medication data into hospital records, carries clinical messages between providers, and now handles benefit and authorisation data. It is a business associate to essentially every customer it has.

What is absent is the contract. No business associate agreement is offered or described, no protected data handling summary exists, no retention position is stated for medication history records, and nothing describes what governs data acquired through the Myndshft platform.

Graded B.

A validated external certification with enumerated scope, announced publicly, and none of the apparatus that would let a buyer examine it.

The certification is HITRUST i1, achieved in March 2025 through a validated assessment rather than self attestation, and the announcement states that all medication management solutions are covered and lists them. Scope enumeration is the detail that separates a usable credential from a decorative one on a multi product catalogue, and this company provides it where several larger vendors in this index do not. The chief executive attached a public statement to it, which is accountability of a minor but real kind.

The honest qualification is tier. The i1 assessment is HITRUST's moderate assurance level, built around a curated control set addressing current and emerging threats, and it sits below the r2 assessment in depth, control count and duration. A buyer comparing this against a vendor holding r2, or against the ISO 27001 certification held by one record in this index, is comparing different levels of rigour.

What is absent is the apparatus. No trust centre, no described process for requesting the certification report or supporting documentation, no service organisation control report, no penetration testing statement, no vulnerability disclosure policy, no subprocessor list and no incident notification commitment were located.

That gap is worth weighing against what this company holds. A national prescribing network carrying controlled substance transactions and medication histories at population scale is critical infrastructure, and the security assurance published about it is one certification announcement.

Graded B.

No device clearance is required and the company operates at the centre of a dense regulatory regime it helped build.

On devices the position is clear. Electronic prescribing, medication history, sig normalisation and benefits checking carry no diagnostic claim, and the sig work in particular converts data into a form the record system can act on rather than issuing a clinical determination. Nothing here approaches a device question.

The regime that governs this company is prescribing regulation and it is substantial. Controlled substance prescribing operates under federal requirements with identity proofing, two factor authentication and audit obligations. Prescription messaging runs on industry standards, and the chief executive worked on building what became the national prescribing network and on the standards themselves before joining, which is unusual regulatory lineage. State prescription monitoring programme mandates, formulary and benefit transparency rules and prior authorisation regulation all bear on the product set.

The Myndshft acquisition extends this into medical benefit prior authorisation, an area under active federal rulemaking on turnaround times and electronic processing, so the company has moved toward regulation that is tightening rather than settled.

What is not published is any analysis of where sig inference sits relative to clinical decision support regulation. Inferring a missing dose or frequency and writing it into a chart that then drives interaction checking is closer to that boundary than transaction routing is, and the company does not address it.

Graded B.

Nothing was located. No model card, no training data description, no accuracy breakdown, no subgroup analysis and no bias statement.

The mechanism here is not demographic in the usual sense and it produces an unusually direct clinical harm, so it is worth naming precisely.

Sig parsing performance will vary by source. Prescriptions arriving from large chain pharmacy systems are formatted consistently and at volume; those from independent pharmacies, small practices, compounding pharmacies and long term care systems are more varied and less represented. A model trained predominantly on the common cases will resolve them well and fail more often on the uncommon ones, and the published figure of 93 percent accurate translation is an aggregate that conceals where the remaining 7 percent falls.

The patients affected are identifiable. Complex sig instructions, tapering schedules, split doses, as needed regimens and uncommon formulations belong disproportionately to elderly patients on polypharmacy, patients on psychiatric medications and patients receiving specialty therapy. Those are precisely the patients for whom medication reconciliation errors carry the most risk, and precisely the instructions a parser finds hardest.

Non English instructions and patients whose medications were dispensed outside the mainstream retail network are a further unexamined population.

The company has the data to answer all of this, having published population scale analysis of the national medication history database, and no breakdown by drug class, sig complexity, source system or patient age was located.

Graded D.

Nothing published addresses responsibility when the model is wrong, and the failure pathway here is more direct than on most records in this index.

The pathway is short. SmartSig infers a missing or ambiguous sig element and writes it into the medication history in the chart. A clinician performing reconciliation reviews a record that now looks complete. The record system runs its interaction and allergy checks against that data. If an inferred frequency, dose or route is wrong, the error is not a suggestion a clinician can weigh against the evidence, because the evidence is the field the model populated, and the safety checking downstream operates on the same wrong value. An incorrectly resolved drug match compounds it by checking interactions against the wrong medication entirely.

That is a mechanism by which an automated data transformation becomes a medication error, and the company's own framing acknowledges the stakes by grounding the product in a national patient safety goal.

Nothing published states accuracy commitments, indemnity, limitation of liability, error reporting routes, or how responsibility divides between the vendor that inferred the value, the source pharmacy whose record was incomplete, and the clinician who reconciled the chart. The published 93 percent translation accuracy is a marketing figure rather than a committed threshold, and no error analysis describes what the residual looks like clinically.

One genuine mitigation belongs on the record. The company publishes its own argument that artificial intelligence in medication management requires human clinical safety checks, which is a stated position on how the product should be used even though it is not a position on who bears the consequence when it is not.

Graded D.

Interoperability is not a feature of this company, it is the company, and the depth is evidenced by customers rather than asserted.

The integration is embedded rather than adjacent. Medication history and sig enhancement operate inside the record system's own medication reconciliation workflow, populating discrete fields in the chart so the record system's interaction and allergy checking can fire. That is the deepest form of integration available: the vendor's output becomes the customer's data rather than sitting in a parallel view, and the named health systems describe it working inside Epic specifically.

Three major record systems are named across the material, Epic, Cerner and MEDITECH, with a dedicated product line built for the third. Prescription messaging runs on the industry standards the company helped establish, so the outbound side is standards based rather than proprietary.

The network dimension extends interoperability beyond the hospital entirely. A national medication history network reaching pharmacies, payers and prescribers is an interoperability asset that took twenty five years to assemble and cannot be replicated quickly, and the Myndshft acquisition added connectivity to more than 600 payers for benefit and authorisation data.

The customer evidence confirms it in operational terms. Hackensack Meridian's automatic sig mapping rising from 26 to 86 percent is a measure of how much data successfully crossed the boundary into structured chart fields, which is interoperability measured as an outcome rather than claimed as a capability.

Graded A.

Nothing published on hosting, region or residency, for a company operating national scale clinical data infrastructure.

The architecture is inferable and nowhere described. Medication history is queried in real time during admission, prescriptions route through a national network, and sig enhancement happens in the path between the source record and the hospital chart. All of that is vendor operated service infrastructure rather than software installed in a hospital, so the company processes and holds prescription and medication history data for a very large share of United States patients.

No hosting provider, region, residency option, subprocessor list, retention position or export term was located.

Availability deserves particular mention on this record and is also unaddressed. Medication history is queried at the moment of admission, and sig enhancement sits between the source data and the chart, so an outage does not merely inconvenience a workflow: it returns the hospital to the manual reconciliation process the product replaced, at a moment when clinicians have already restructured staffing around not doing it. No uptime commitment, status history or continuity position is published.

The Myndshft acquisition adds an undescribed second infrastructure, since benefit and authorisation checking against more than 600 payers runs on whatever platform that company built, and nothing states whether it has been consolidated.

The HITRUST i1 certification implies a documented and assessed control environment behind all of this, without disclosing where any of it runs.

Graded D on the absence rather than on any evidence of a problem.

Nothing is published. No price, no unit of charge, no module structure, no implementation fee, no contract term and no minimum was located in vendor or third party material.

The catalogue makes that absence consequential. The company sells electronic prescribing, controlled substance prescribing, medication history, sig enhancement, secure messaging, benefits checking, patient adherence, and now medical benefit prior authorisation through the Myndshft acquisition, with several of those bundled into the Fuzion platform since 2023. A buyer cannot determine which capabilities sit inside a Fuzion licence and which are separate, whether medication history is priced per transaction, per bed, per prescriber or per site, or how the newly acquired prior authorisation capability is licensed against the rest.

The unit question matters more than usual here because the volumes are enormous. Medication history is queried at every admission and transition of care, and sig enhancement runs on every imported medication, so a per transaction basis and a site licence produce very different annual costs for the same hospital. Nothing indicates which applies.

The published outcome figures give a buyer one side of a business case in unusual detail. Covenant HealthCare's recaptured 15 percent productivity per shift, valued at roughly $650,000 monthly, is exactly the kind of number a finance office can model against a price, and no price is available to model it against.

Graded D.

Coverage spans the full medication pathway across both care settings, and the breadth is a consequence of twenty five years of network building rather than a marketing claim.

Setting coverage runs from ambulatory prescribing through hospital admission, transfer and discharge. Electronic prescribing and controlled substance prescribing serve the outpatient prescriber, medication history and sig enhancement serve the hospital at every transition of care, and discharge prescribing and adherence work extend past the encounter. Medication reconciliation is recommended at every transition, so the product touches the same patient repeatedly across settings.

Specialty coverage is inherent rather than configured. Every prescriber writes prescriptions and every admitted patient has a medication history, so the addressable population is not a specialty at all. The named customers reflect that spread, running from a large integrated network and two academic systems to a community hospital and a rural medical centre in Mississippi.

Record system coverage is the practical enabler and is unusually wide, with Epic, Cerner and MEDITECH all named and integrations built to industry prescribing standards rather than per customer.

The Myndshft acquisition extended coverage across a boundary most vendors never cross, from pharmacy benefit into medical benefit, reaching eligibility for 95 percent of insured patients and authorisation requirements for more than 600 payers. Expensive infused and specialty drugs sit on the medical benefit and are exactly the therapies where authorisation delay causes harm.

Graded A.