Unlearn.AI

The generative model is the product and also the intellectual property. Digital Twin Generators are trained on historical control and observational data to forecast individual disease progression, and the entire commercial offering is that forecast plus the statistical methodology for using it. Published work names the model class directly, including a conditional restricted Boltzmann machine trained on harmonized data from 6,736 subjects for the Alzheimer's generator. There is no services or platform layer to fall back on.

The most rigorous answer to the autonomy question in this index, because the company solved it by design rather than by policy. The team explicitly rejected the obvious application, using digital twins as external control arms in single arm trials, on the grounds that all non randomized designs with external controls are prone to bias from unknown confounders.

Instead the twin only supplies a prognostic score adjusted for within a randomized controlled trial, which preserves randomization and bounds bias even when the model omits relevant variables. The AI never replaces a control patient or a human judgment; it improves the precision of an analysis that remains randomized. Constraining your own method to preserve statistical validity is a stronger safety posture than any oversight disclosure.

Method transparency here approaches the academic standard. The three step procedure is published and patented, the mathematical argument is public including a proof that adjusting for a prognostic score from a model trained to predict control outcomes yields the most efficient trial at a desired power, and extensions such as Bayesian PROCOVA are on arXiv with named authors. Model classes and training cohort sizes are disclosed in peer reviewed applications. A buyer or a regulator can audit the methodology from primary sources rather than a datasheet, which is why a regulator was able to qualify it.

The corpus provenance is structurally cleaner than the alternatives and the company draws the distinction itself, which is why this sits above the floor. The models are trained on historical control arm and observational data, meaning data from completed studies where participants consented to research participation under a protocol and an ethics review.

That is a fundamentally different acquisition path from assembling a corpus of routine clinical records through commercial data partnerships, and it carries a cleaner consent story by construction rather than by assurance: those participants agreed to be in research, and research is what the data is being used for. What remains open is the sponsor side.

Nothing published states how sponsor supplied datasets are handled, retained, segregated or returned, nothing describes the de identification standard applied to the harmonised datasets, and nothing addresses whether data contributed by one sponsor's historical trials can inform a generator applied to another sponsor's study.

That last one is the question to press, because the incentive runs the wrong way: the value of a generator grows with the breadth of historical control data behind it, so there is a structural pull toward pooling, and a sponsor has a structural interest in its own trial data not improving a competitor's analysis. Ask whether learning crosses sponsor boundaries, and what happens to sponsor data at engagement end.

Substantial published evidence, though it measures statistical efficiency rather than patient outcomes, which is the correct endpoint for this product. Applications include an Alzheimer's study using the AWARE trial of 453 subjects reporting positive partial correlations between twins and cognitive change scores of 0.30 to 0.39 at Week 96, consistent with validation across three independent trials at 0.30 to 0.46, and total residual variance reduced by roughly 9 to 15 percent.

A separate simulation of the TRAILBLAZER-ALZ 2 donanemab trial quantified potential sample size reduction. Partnerships in ALS are underway. Note that several key demonstrations are simulations or retrospective applications rather than prospective use as the primary analysis of a completed registrational trial.

Converted from Not Rated. The data profile is materially lower risk than most records in this index and the terms governing it are still unpublished.

What distinguishes this from the real world data platforms graded alongside it is provenance. The models are trained on historical control arm and observational data, meaning data from completed studies where participants consented to research participation under a protocol and an ethics review. That is a fundamentally different acquisition path from assembling a corpus of routine clinical records through commercial data partnerships, and it carries a cleaner consent story by construction. The company should get credit for the distinction, which its own materials draw clearly.

It does not remove the questions. Nothing published states how sponsor supplied datasets are handled, retained, segregated or returned. Nothing addresses whether data contributed by one sponsor's historical trials can inform a generator applied to another sponsor's study, which is the commercially and ethically sensitive question for a business built on pooled historical controls. And nothing describes the de identification standard applied to the harmonised datasets.

The second of those is the one to press. The value of a digital twin generator grows with the breadth of historical control data behind it, so there is a structural incentive to pool, and sponsors have a structural interest in their own trial data not improving a competitor's analysis.

Ask whether learning crosses sponsor boundaries, and what happens to sponsor data at engagement end.

Converted from Not Rated. The health privacy rule is largely not the operative regime, and what replaces it is not published.

The customer is a trial sponsor rather than a provider organisation, so there is usually no covered entity in the contracting chain and no business associate relationship to form. Contracting runs through sponsor data agreements and clinical trial data transfer terms instead. Nothing was retrieved describing those.

The research context supplies its own framework and it is a real one rather than an absence. Data from completed trials sits under the protocol and consent it was collected under, under ethics committee oversight, and under the sponsor's own obligations to participants. That framework is arguably stricter than the health privacy rule in some respects, particularly on secondary use, since a participant consented to a specified study rather than to open ended reuse.

Which is precisely why the unanswered question matters. Historical control data being reused to train generative models applied to future trials is a secondary use, and whether the original consents contemplated it, and who assessed that, is not addressed publicly. This is not a criticism unique to this vendor; it is the central governance question for the whole digital twin approach, and this company is better placed than most to answer it given its regulatory engagement.

Ask what consent and ethics basis covers reuse of historical control data, and who reviewed it.

Converted from Not Rated. No independent security attestation was located.

No SOC 2 of either type, no ISO 27001, no HITRUST, no trust centre and no penetration testing statement was retrieved.

The absence is conspicuous against the rest of this record rather than in line with it. This company holds the strongest regulatory position of any vendor in its category, having obtained a positive qualification opinion from the European regulator for its statistical method and a concurring comment from the United States regulator's drug evaluation centre. Securing that required submitting methodology to sustained external technical scrutiny, which is a posture of welcoming examination. Extending the same instinct to information security would be a small step and it has not been taken publicly.

What is held is commercially sensitive rather than clinically so. Sponsor engagements involve unpublished protocol designs, endpoint strategies and interim thinking about assets that are frequently the most material non public information a pharmaceutical company holds. A breach would not primarily expose patients, whose data is historical and consented; it would expose which trials sponsors expect to succeed.

Absence of a retrieved document is not proof none exists, and sponsors of the size this company works with run demanding vendor assessments, so review has plainly occurred privately.

Ask for the attestation and period, and for the segregation controls between concurrent sponsor engagements.

A distinct and arguably harder regulatory achievement than a device clearance, and unique in this index. In September 2022 the EMA issued a positive qualification opinion through its novel methodologies programme, qualifying PROCOVA as an acceptable statistical approach for the primary analysis of Phase 2 and 3 trials with continuous outcomes.

Unlike a 510(k), which establishes substantial equivalence to a predicate, a qualification opinion is a regulator endorsing a novel method as fit for evidence generation. FDA CDER subsequently stated it concurs with EMA and that PROCOVA does not deviate from current guidance.

Two important limits a buyer should hold: the qualification is narrow, covering continuous endpoints under specific assumptions, and FDA has issued no parallel qualification opinion, with applicability to oncology endpoints and long term survival outcomes unsettled.

Stronger than most, expressed as statistical rigor rather than a governance programme. The method is specifically constructed so that bias is controlled even in the presence of unknown confounders, and the company engaged FDA through a Critical Path Innovation Meeting to confirm that controlling bias and type I error was the operative concern for pivotal trials. Validation reports correlation consistency across independent trials, which tests generalization.

What is absent is subgroup or demographic performance analysis of the twin generators, which matters because a prognostic model trained on historical trial populations inherits the well documented underrepresentation of those populations. Independent commentary notes the validation question for digital twins is itself unsettled science.

This is the only record in the backfill whose central claim is a theorem, and that changes what verification means. The three step procedure is published and patented, and the mathematical argument is public, including a proof that adjusting for a prognostic score from a model trained to predict control outcomes yields the most efficient trial at a desired power. A proof can be checked by anyone with the training, using no data at all, and it either holds or it does not.

That is a stronger form of falsifiability than any benchmark, because a benchmark can be selected and a proof cannot. Extensions are posted publicly with named authors, and model classes and training cohort sizes are disclosed in peer reviewed applications.

The causal point is worth stating plainly: a regulator was able to qualify this method because the methodology could be audited from primary sources rather than from a datasheet, so the transparency was not a courtesy but the precondition for the qualification. Held below the top grade because nothing attaches commercially.

No warranty, indemnity or remediation commitment was located, and a correct theorem about the estimator says nothing about whether the generator's assumptions hold in a particular disease or population, which is where the method would actually fail. Ask what happens when the prognostic model transfers poorly to a new indication, and how that would be detected before a trial reads out.

Converted from Not Rated, and this is a genuine category non application rather than a gap.

The product operates on trial datasets inside sponsor statistical workflows. It does not sit in a care setting, receives no orders, returns nothing to a clinician and has no reason to connect to a provider record system. Grading it down for lacking clinical integration would penalise a business model rather than describe it, which this index has ruled against where an axis mismatches a vendor's position.

Declining to claim integration breadth it does not need is the creditable part, and it is consistent with how this company handles its other claims, which are unusually disciplined for the category.

The interoperability that would actually matter here is with the statistical environment rather than the clinical one: how a prognostic score generated for each participant reaches the analysis, in what format, and whether it can be reproduced and audited by the sponsor's own statisticians or by a regulator reviewing the submission. That last point is not a convenience question. A method accepted for primary analysis has to be reproducible by parties other than the vendor, and the mechanics of that hand off are what an interoperability axis should be measuring for this product class.

No detail on integration with clinical data capture systems or statistical computing environments was located.

Ask how the prognostic score is delivered, in what format, and how a regulator or independent statistician reproduces it.

Converted from Not Rated. No hosting, region, tenancy or residency terms were located, and the engagement model is the reason this reads differently from a software purchase.

Work appears to run as sponsor collaboration on a specific trial rather than as software deployed into a customer environment. A generator is applied to a study, prognostic scores are produced for participants, and those scores enter the sponsor's analysis. That is closer to a specialist statistical service than to a platform, and the deployment questions change accordingly: less about where a system is hosted, more about where the sponsor's trial data goes, who at the vendor touches it, and what returns.

None of that is published. No statement of whether sponsor data is transferred to the vendor or processed within the sponsor's environment, no region, no personnel access model, no retention schedule and no return or destruction terms at engagement end.

The transfer direction is the specific question. If sponsor trial data moves to the vendor, a pharmaceutical company's live study data is sitting with a third party during the period when it is most sensitive. If the generator can be applied within the sponsor's own environment, that exposure does not arise. Those are materially different arrangements and the published material does not indicate which is offered, or whether both are.

Ask whether the work is performed on sponsor infrastructure or vendor infrastructure, what data is transferred, and what is destroyed at close.

No pricing is published. The value case is nonetheless quantifiable from published evidence in a way most vendors cannot match, since reduced sample size and shortened enrollment translate directly into trial cost, and the studies report variance reduction figures a sponsor can model against their own trial economics. Named partnerships indicate a sponsor collaboration model rather than a licensed product, and terms are not disclosed.

Coverage is bounded by which Digital Twin Generators exist, and generators are disease specific, trained on historical data for a given condition. Published work centers on neurodegenerative disease, with Alzheimer's most developed and ALS partnerships underway. The EMA qualification is further limited to continuous endpoints, which excludes the binary and time to event endpoints common in oncology and cardiovascular trials. Deep where a generator exists, unavailable where one does not.