PrecisePK

The dose is computed by Bayesian updating of a published population pharmacokinetic model against a patient's measured serum concentrations, which is statistical inference rather than machine learning. The company's technology page does state that the product uses machine learning alongside Bayesian principles, and the mechanism it describes for that is assigning a patient into a cluster with corresponding pharmacokinetic parameters to select the population model, which is a modest and legitimate use of the term rather than a claim about a learned dosing model.

A sibling brand, PreciseRx, is described as uniting clinical research, machine learning and cloud computing, which suggests the learned component is a direction rather than the current engine. Graded C on the mechanism with the accuracy of the description credited: this is a rigorous pharmacometric tool that does not need to be an artificial intelligence product to be a good one, and the grade describes what the software does rather than how well it does it.

A clinical pharmacist runs the calculation, reviews the output and makes the recommendation, so the software never touches an order. Two design choices strengthen the oversight position. The user can either have the software select the pharmacokinetic model that best fits the patient or select it manually, which keeps the single most consequential modelling decision in human hands rather than automating it silently, and that is a meaningful contrast with competitors that set defaults centrally.

The product also carries a toxicity alerting layer that warns when area under the curve falls outside the recommended range or when a patient approaches an ototoxic exposure, so the software flags risk in the direction of caution rather than only proposing a dose. Held at B because no acceptance rate, override rate or confidence bound is published, which is the metric this whole category still lacks.

The technology page walks a reader through the actual computation in sequence: a population pharmacokinetic model drawn from published literature and research, patient specific covariates such as age, sex and weight forming a patient profile, assignment into a cluster with corresponding pharmacokinetic parameters, and then additional drug factors. That is more mechanism than most vendors in this index publish and it is written for a pharmacist rather than a procurement committee.

Independent analysis attributes the product's accuracy to the comprehensiveness of its model library and its use of two compartment kinetics. Held at B rather than higher because the specific population models in the library are not named the way a direct competitor names its Thomson, Goti and Carreno models, no thresholds or calibration are published, and the machine learning element referenced on the same page has no published method behind it.

This vendor holds the best independent comparative result in the category and it should be stated with its limits attached, because the company states it without them. Turner and colleagues, publishing in Pharmacotherapy in 2018, compared five Bayesian dose optimising programs and two first order equations for estimating vancomycin area under the curve in critically ill patients, and found PrecisePK the least biased of the group at a median of 5.1 percent.

The same study found this product among the more difficult of the five to use, and found that adding a second serum level improved accuracy for two competitors but not for this one. It included nineteen patients. The company's own material describes the product as validated as the most accurate and least biased Bayesian guided software without qualifying the sample size, which is a superlative resting on a very small study.

Set against that, the clinical lineage is real and long: development inside a university applied pharmacokinetics service, beta testing at an academic medical centre and a children's hospital, and more than thirty years of continuous use at UC San Diego Health across nine University of California sites.

The data footprint is narrow by design, since a dosing calculation needs demographics, renal function, dosing history and drug concentrations rather than a full chart, and integrated deployments import exactly those fields. Retrieval located no retention schedule, no de identification statement and no description of what happens to patient data after a case is saved, and the product explicitly supports saving patient cases and sharing them between colleagues for handoff, which is a clinically sensible feature that raises an access control question nothing published answers.

The local population model capability, where a hospital's own patient data is used to refine models for that institution, is the point at which customer data becomes model input, and a buyer should establish whether those locally derived models remain the institution's or become part of the vendor's library.

Two retrieval passes, one general and one targeted specifically at compliance material, surfaced no HIPAA statement, no business associate agreement terms, no execution path and no compliance page belonging to this vendor. The grade describes published posture rather than contractual reality, since a business associate agreement necessarily governs deployment inside University of California health system sites, which run some of the more demanding vendor security reviews in American academic medicine.

The gap is notable precisely because the customer list implies the paperwork exists and has been scrutinised by sophisticated buyers, so publishing it would cost the company very little and would answer a question every subsequent prospect has to ask from scratch.

Retrieval located no SOC 2 attestation, no HITRUST certification, no ISO 27001, no trust centre, no penetration testing cadence and no vulnerability disclosure policy. Two genuine credits belong on the record even so. A third party review directory reports single sign on integration with named enterprise identity providers including Okta, Duo, InCommon and SecureAuth, which is a real enterprise security capability rather than a badge, and InCommon in particular reflects the academic federation its university customers actually use.

Listing in Epic's integration marketplace also implies that vendor's own technical review was passed. Neither is an independent attestation a security team can read, and for a product used at academic medical centres for three decades the absence of one is the most surprising finding on this record.

No clearance was located in any market and none is claimed. That is the long standing position for therapeutic drug monitoring software, which has been used under a clinical pharmacist's professional judgement since well before software regulation contemplated it, and the pharmacist mediated workflow supports the argument.

The tension worth putting to the vendor is that this product computes a dose, and computing a dose is precisely the function that led one direct competitor to obtain a European CE mark as a medical device and led insulin dosing software in this same category to be cleared as Class II software in the United States.

So the same act is treated as a regulated device function by some vendors and as unregulated clinical pharmacy software by this one, with no stated reasoning published either way. A buyer should establish the vendor's regulatory position explicitly rather than infer it from thirty years of uneventful use.

Retrieval located no performance reporting by age, sex, race, ethnicity, renal function band or body habitus, no bias assessment and no post deployment monitoring statement. The exposure is inherent to the method rather than to this vendor's conduct: a Bayesian dose depends on prior distributions taken from a published population, so whoever was enrolled in the derivation cohort determines whose physiology the prior describes, and a patient unlike that cohort receives a worse starting estimate.

One architectural feature deserves credit as the most interesting partial answer in this lane. The platform supports building population models from a hospital's own patient data, which addresses the generalisation problem structurally by letting an institution replace a foreign prior with a local one. That is a real mitigation and it is undercut by the absence of any published evidence about how much it improves accuracy or for whom.

Integration is real and narrower than the leaders in this category. A third party review directory reports full Epic integration through that vendor's app marketplace with automatic import of medication orders, prior doses, drug levels and serum creatinine, which are precisely the four fields the calculation needs, plus single sign on support and an integration with the Sanford Guide that puts infectious disease guidance beside the dosing decision.

The vendor's own material references electronic health record integration in general terms and treats it as a purchasing consideration alongside clinical surveillance tools and reporting compliance. Held at B because only one electronic health record is named, no FHIR conformance statement or public API documentation was located, and the specifics come from a secondary source rather than from the vendor's own documentation.

The product moved from desktop software to a web based platform, which is the current delivery model, and the company describes revamped integration features accompanying that move. Retrieval located no named hosting provider, no cloud region, no data residency commitment and no statement of whether an on premise or institution hosted option survives from the desktop era, which some academic customers may still want.

The desktop lineage matters commercially as well as technically, because a buyer evaluating a thirty year old product should establish which version their peers are actually running and whether the published evidence and the current platform are the same software.

This vendor does better than the rest of the lane without publishing a price. The site offers a sixty day free trial of the software prominently and repeatedly, which is a concrete commercial term a buyer can act on immediately, and it materially changes evaluation economics for a pharmacy department that wants to test accuracy against its own patients before committing.

Unlimited one to one onboarding training and ongoing technical and clinical support are described as included rather than as separately quoted professional services. What remains undisclosed is the price itself, the unit, the contract term and whether the trial converts automatically.

Graded C rather than higher for that gap, and higher than the rest of the category because a published trial period is a real disclosure and every competitor graded so far offers nothing a buyer can evaluate without a sales conversation.

Coverage is deep within a narrow function. The product supports therapeutic drug monitoring across the drug classes where measured concentrations guide dosing, with vancomycin dominating current practice and the aminoglycosides and other narrow therapeutic index agents alongside it, and the company states use in all types of patients. Paediatric capability is credible rather than asserted, given beta testing at a children's hospital in the product's development history.

Institutional coverage is concentrated in American academic medicine, with University of California sites named and a thirty year relationship at one of them. Held at B because the footprint outside academic centres is not described, there is no evidence of use outside the United States, and the function is confined to drugs with routine concentration monitoring rather than to prescribing generally.