ClosedLoop

There is nothing here but the models. The company describes itself as healthcare's data science platform and the product is exactly that: a model library, a training and deployment environment, and individual level predictions delivered to the organisations that act on them.

No workflow application, no clinical documentation layer, no hardware, no staffed service and no patient facing component exists to carry value independently. Strip the machine learning out and what remains is an empty pipeline.

The automation of data preparation reinforces rather than dilutes this. Preparing healthcare claims and clinical data for modelling is where most of the effort in this discipline is consumed, and productising that step is machine learning infrastructure serving machine learning rather than a separate business.

One detail confirms the centrality externally rather than by assertion. The federal competition the company won judged predictive performance and explainability, which means the thing assessed by an independent body was the algorithms themselves, not an application wrapped around them.

Graded A without qualification. Every question worth asking about this vendor is a question about its models, and the record's weaknesses on disclosure axes reflect a company that has published its results and very little else.

Explainability is the product's founding requirement rather than a feature added to it, and an independent federal body adjudicated it against 300 rivals.

The competition brief is the reason this earns the top grade. The challenge did not ask for the most accurate predictor; it asked for explainable artificial intelligence solutions that front line clinicians could understand and trust. Explainability was a judging criterion in a multi stage contest run over roughly two years by a federal agency's innovation centre alongside a physician professional body. Winning that is external, adversarial validation of exactly the property this axis exists to measure, and no other record in this index carries anything comparable.

The design reflects it concretely. Each Patient Health Forecast surfaces the specific variables driving an individual's risk alongside the prediction, and links to interventions a clinical team can act on. That gives a clinician three things rather than one: what the model concluded, why it concluded it, and what could be done. A contributing factor breakdown at individual level is inspectable in a way a population risk score is not, and the intervention link closes the gap between a prediction and a decision.

The operating model keeps humans throughout. Predictions inform outreach prioritisation and care management decisions made by clinicians and care teams; nothing acts autonomously.

What is missing is calibration. No operating point, threshold, precision at rank, or guidance on how many patients to action is published, and in risk stratification the cut off determines who receives scarce intervention capacity.

Ask for model performance at the deployed threshold and how customers are advised to set it.

A published open source model and adjudicated explainability, without performance figures for the commercial products.

The strongest disclosure is the open sourced vulnerability index released during the pandemic and documented in a medical artificial intelligence journal. Releasing a working model publicly, with methodology in the literature, is transparency that costs something, and it allowed independent inspection of the company's approach rather than requiring trust in it.

The platform is also described at a useful architectural level: a library of off the shelf models for common healthcare use cases, automated data preparation to accelerate development, and individual level forecasts carrying contributing factors. A technical buyer understands what they are adopting.

Explainability is not merely claimed but was assessed by a federal competition specifically judging it, which substitutes external scrutiny for a self report.

What is missing is performance. No discrimination or calibration figure is published for any commercial model, no model card exists for the library, no architecture is described beyond the platform framing, and the number of models available is stated only as hundreds built historically.

That gap is awkward for a company whose credibility rests on a prediction contest, since the competitive result implies performance data exists and was submitted, and none of it is assembled where a buyer would find it.

One pre emptive note: further awards cannot move this grade. Only published model performance will.

Ask for discrimination and calibration on the flagship admission and utilisation models, and a model card for the shared library.

A dedicated pass located no cloud or infrastructure provider, no sub processor register, no machine learning framework or third party component, and no training data provenance for the commercial model library.

One genuine exception exists and deserves recording. The vulnerability index released during the pandemic was open sourced and documented in a journal, which means at least one model's construction is fully inspectable. That is more than most vendors here offer for any model, and it is a single artefact rather than a description of the supply chain behind the commercial products.

The library is where the question concentrates. Off the shelf models for common healthcare use cases are a headline capability, and a pre built model necessarily encodes a training population. Whose data, from which organisations, over what period, under what agreements, and how well that population matches a new customer's members are all unstated. A model trained predominantly on commercially insured members and deployed against a Medicaid population would behave differently, and nothing lets a buyer assess that risk.

The federal work adds a further unknown, since Medicare beneficiary data used in a challenge context carries data use terms and nothing describes whether that data informed anything commercial.

Infrastructure is entirely undisclosed for a platform processing population scale identified records.

Ask for the training population behind the shared model library, the hosting provider and sub processor register, and whether customer data enters the library.

Competitively adjudicated validation by a federal body, sustained independent ranking, and published work, which together make this among the strongest evidence positions in this index.

The competition result is the centrepiece and is different in kind from the evidence most vendors offer. The Centers for Medicare and Medicaid Services ran a multi stage artificial intelligence challenge from 2019, through its innovation centre with a physician professional body and a philanthropic funder, on predicting unplanned admissions and adverse events among Medicare beneficiaries. More than 300 entrants competed, including large technology firms, academic health systems and pharmaceutical companies, with a major integrated health system taking second place. A vendor's own study is evidence it commissioned; a placing in an adjudicated contest against named rivals on a common task is evidence it could not control.

Independent recognition continued, with the top position in a major analyst firm's healthcare artificial intelligence data science category across 2022, 2023 and 2024. That ranking derives from customer surveys, so it measures delivered satisfaction over three years rather than a single moment.

Publication exists on two fronts: an open sourced vulnerability index documented in a medical artificial intelligence journal and stated to have reached more than 10 million lives, and collaborative work with a value based primary care provider published in a clinical management journal.

Two qualifications belong on the record. A named customer also invested in the Series B, so references from it carry a commercial interest. And no evidence, publication or customer announcement was located after early 2024.

Ask for outcome data from a current customer, and what has been published since 2024.

A dedicated pass located no encryption statement, no retention schedule, no access control model, no audit logging description, no deletion process and no data ownership position.

The data concentration here is among the largest in this index and that is what makes the silence weigh. Person level claims and clinical histories for entire health plan memberships, combined across sources and retained long enough to train longitudinal models, constitutes a population scale identified dataset. Predicting chronic disease onset in particular requires years of history per individual, so retention is not incidental to the product but required by it.

The training question follows directly and is unanswered. The commercial proposition includes off the shelf models for common healthcare use cases, and those models were built from data. Whether that data came from customers, whether current customers contribute to it, whether an organisation can decline, and whether one customer's population improves models sold to a competitor are all unaddressed, and in a market where health plans compete on risk management that last point is commercially as well as legally significant.

The federal relationship adds a further dimension, since work involving Medicare beneficiary data operates under its own data use agreements and nothing describes how that is segregated.

One pre emptive note: further awards or publications cannot move this grade. Only a published retention position, a data ownership statement and a training data policy will.

Ask what is retained and for how long, who owns derived models, and whether customer data trains the shared library.

A dedicated pass located no health privacy position of any kind: no compliance statement, no business associate agreement template, no execution requirement and no processing terms.

The omission is consequential because of what this platform necessarily holds. Predicting individual level risk across a health plan's membership requires ingesting claims, eligibility, clinical and often pharmacy and laboratory data at person level for entire populations. A national health plan customer means millions of identified records, and an accountable care organisation means the complete longitudinal picture for its attributed patients. This is not a product that touches a patient at a moment of care; it holds the whole file.

Every named customer is a covered entity, so agreements demonstrably exist across the base and are simply unpublished, which leaves a prospective buyer unable to begin diligence without a sales conversation.

One question is specific to this architecture and unaddressed. A platform supplying pre built models for common use cases must have developed those models on data from somewhere, and whether customer data contributes to the shared model library, whether models trained on one organisation's population are reused for another, and what contractual protection exists against that are the questions a health plan's counsel would raise first.

Ask for the agreement template, whether customer data informs the shared model library, and what governs data used in federal contract work.

A dedicated pass located no security page, no external attestation, no trust centre, no penetration testing statement and no vulnerability disclosure policy. No security credential of any kind was found.

The absence is striking against the customer base rather than merely regrettable. A large nonprofit health plan, the largest Medicaid accountable care organisation and a federal contractor all conduct demanding vendor security reviews before granting access to member data, so a posture certainly exists and has been examined privately several times over. None of it is published where a prospective buyer could begin diligence, which is a straightforward gap between what the company can evidently satisfy and what it chooses to show.

The asset concentration justifies the scrutiny. This platform holds person level claims and clinical histories for entire covered populations, which is among the largest identified datasets any vendor in this index assembles, and it is exactly the profile that attracts organised intrusion.

The three consecutive top analyst rankings are customer satisfaction measures and say nothing about security controls, so they do not substitute.

One pre emptive note: further awards, publications or customer names cannot move this grade. Only an external attestation, or security documentation available under agreement, will.

Ask whether an information security attestation is held, what documentation is available under agreement, and what authorisation covers federal work.

Device regulation is genuinely not the applicable regime, and no determination is published to say so.

Population risk stratification for care management sits outside device regulation in most readings. The output prioritises outreach and resource allocation rather than diagnosing a condition or directing a treatment, the consumer is a care management team rather than a clinician at the point of care, and the underlying data is largely administrative. No clearance is claimed and none is evidently required.

Two capabilities sit closer to the boundary than the rest and neither is addressed. Predicting the onset or progression of chronic disease in a named individual is a clinical prediction about a person rather than a resource allocation signal, and the more the output is presented to a clinician alongside recommended interventions, the more it resembles decision support. The intervention linkage that earns this vendor credit on oversight is the same feature that moves it toward that line.

The federal relationship is a different regulatory dimension worth recording. Winning an agency challenge and working with a federal contractor implies exposure to government data use and contracting requirements, and nothing describes that posture.

One regime is emerging and unmentioned. Health plans deploying algorithms that influence coverage related decisions face growing regulatory attention on algorithmic fairness and transparency, and a vendor supplying risk models to plans will be drawn into it.

Ask for the written device determination covering disease onset prediction, and the compliance position on payer algorithmic oversight.

Genuine transparency practice, and no bias evidence in the one domain where the field's most famous failure occurred.

What supports the grade is real. Explainability is designed in and externally adjudicated, and per patient contributing factor disclosure is itself a governance mechanism, because a reviewer can see which variables drove a prediction and challenge them. The company also open sourced a vulnerability index and documented it in a journal, which is a substantive transparency act rather than a statement of principle.

What is absent is any fairness analysis, subgroup performance, calibration data, drift monitoring or external audit.

That gap is more serious here than the same gap elsewhere, because this is precisely the product category in which algorithmic bias was most consequentially demonstrated. A widely deployed population health risk algorithm was shown to systematically understate need for Black patients because it used healthcare cost as a proxy for illness, and cost tracks access rather than sickness. This company's own use cases include predicting total utilisation and total risk, which are cost adjacent targets, and its customers include a Medicaid accountable care organisation serving a population where that failure mode bites hardest.

Explainability helps and does not resolve it. Seeing which variables drove a prediction does not reveal that the target variable itself encodes unequal access.

Ask what outcome variables the risk models predict, whether cost proxies are used, and for performance and calibration by race, ethnicity and payer type.

Nothing allocates responsibility, for predictions that determine who receives scarce care resources.

A dedicated pass located no service level agreement, no accuracy warranty, no performance guarantee, no indemnity and no remediation position.

The harm profile here differs from the clinical products elsewhere in this index and is easier to overlook. A missed prediction does not produce an immediate event; it produces an omission. A patient who should have been prioritised for care management is not contacted, deteriorates over months, and is admitted, and nothing in the record connects that outcome to the model that failed to rank them. Errors of this kind are invisible by construction, which is exactly why a published performance commitment would matter and why its absence is difficult for a customer to detect.

The allocation dimension makes it sharper. Risk stratification exists to direct limited intervention capacity, so a model that systematically under ranks a subgroup withholds care from that group at scale while appearing to function normally.

No performance figure is published for any commercial model, so a customer cannot characterise the residual risk it accepts, and no contractual term addresses what happens if the models underperform against expectation.

One pre emptive note: further competition results or analyst rankings cannot move this grade. Only contractual terms, or published model performance with a stated error profile, will.

Ask what the agreement warrants on model performance, what recourse exists if predictions underperform, and how model drift is detected and remedied.

A dedicated pass located no electronic health record integration, no interface standard, no marketplace or validated listing, no published application programming interface and no named source system.

The platform necessarily ingests data at scale, since individual level prediction across a health plan membership requires claims, eligibility, clinical and frequently pharmacy and laboratory feeds, and automated data preparation is marketed as a core capability. Something is therefore connecting to source systems, and nothing describes what.

The outbound direction is the more consequential omission. A prediction that a specific patient is at risk of admission is worthless unless it reaches whoever can act, and nothing states whether forecasts post to a record system, a care management platform, an outreach tool or only to the company's own interface. The product's own strength on oversight, which is linking predictions to interventions, depends entirely on that delivery path.

Customers span health plans, accountable care organisations and physician groups, which run entirely different downstream systems, so a single unspecified integration story is unlikely to cover all of them.

One related question follows from the automated data preparation claim. Whether that automation implies standardised connectors to named systems, or a services led mapping exercise per customer, materially changes implementation effort and is unstated.

Ask which source systems are connected in production, through what standards, and how forecasts reach the teams that act on them.

A dedicated pass located no hosting provider, no region, no residency commitment, no tenancy model and no continuity position.

The delivery model is characterised in third party material as software as a service, which implies a hosted multi tenant architecture, and the company's own material does not state it. Nothing indicates whether a customer can deploy within its own environment, which matters more here than for most products because the customers are health plans and accountable care organisations holding population scale identified data, and some will have policies against exporting it.

The tenancy question is unusually pointed for this vendor. Health plans compete with one another, and a platform holding several plans' complete membership data has an obvious segregation obligation that no published material addresses. Whether models, features or derived data are isolated per tenant is precisely what a plan's security review would examine.

The federal dimension adds another. Work associated with Medicare beneficiary data and a federal contractor ordinarily brings environment and authorisation requirements, and no authorisation status is claimed anywhere.

Continuity is less acute than for a monitoring product, since prediction runs are typically batch rather than real time, and no availability position is published even so.

Ask which provider hosts the platform and where, how tenants are segregated, whether customer environment deployment is available, and what authorisation covers federal work.

Cost is absent from every published surface. A dedicated pass located no pricing page, no unit of charge, no range, no implementation fee position and no minimum commitment on any company controlled surface.

A third party company profile states that the platform is available as software as a service with pricing tiers for organisations of all sizes from small regional provider networks to large national health plans. That is a description of a tiering approach rather than a disclosure, it carries no figures and no tier boundaries, and it was not located in the company's own material, so this record does not treat it as published pricing.

The unit question is genuinely open for a platform of this shape. Charging could plausibly follow covered lives, models deployed, predictions generated, data volume ingested, seats for data scientists, or an enterprise licence, and each produces a different total for the same customer. A health plan with several million members and an accountable care organisation with fifty thousand attributed patients are both named buyer types and would experience those models very differently.

No return proxy is published either, which is a notable omission given the domain. Value based care economics are explicitly financial, the company's own use cases include predicting avoidable admissions and emergency department over utilisation, and the cost of an avoided admission is a figure every customer already holds. No calculation is assembled anywhere.

Ask for the unit of charge, the tier boundaries, how model count affects price, and the contract term.

Broad across buyer types and use cases, evidenced by a named customer list rather than a claim.

The buyer range covers payers, providers, accountable care organisations and pharmaceutical companies, and the named customers occupy genuinely different positions in the system: a large nonprofit commercial health plan, the largest Medicaid accountable care organisation, a long term care accountable care organisation, physician groups, a value based primary care provider and a federal contractor. Those organisations hold different data, face different incentives and manage different populations, so a platform serving all of them is demonstrating adaptability rather than asserting it.

Use case coverage is enumerated specifically rather than gestured at, spanning admissions and readmissions, total utilisation and total risk, emergency department over utilisation, out of network utilisation, appointment non attendance, chronic disease onset and progression, and clinical documentation and reimbursement, with the company noting dozens more.

That breadth follows naturally from the architecture, since a platform supplying a model library and automated data preparation is not tied to one clinical question the way a purpose built application is.

What is not evidenced is depth. Nothing states how many models a typical customer runs in production, which use cases carry real adoption against which are available, or the size of the covered population under management.

The Medicaid and long term care customers are worth noting as populations that commercial risk models frequently serve poorly.

Ask for models in production per customer, covered lives under management, and which use cases have the deepest adoption.