Novellia

The models are the engine of the data product, and the collection channel carries value of its own. The life sciences site says the company's proprietary models clean, link and enrich structured and unstructured records from patients' connected portals into longitudinal journeys, and that its AI structures more than 15 years of multimodal data per patient; the FAQ adds that analyses such as adverse event prediction and adherence drop-off risk are delivered on top.

Remove the models and the company still holds a consented, patient-contributed record network and a free personal health record app, which is the collection mechanism and a large part of what a buyer is paying for. That is the model as the engine of a core module rather than the whole product.

The records are stitched automatically, and the person the record is about can review and correct it. The patient site says users see their consolidated history, fix outdated information and add details, which puts the one person who can tell whether a record has been merged onto the wrong patient in a position to catch it.

What is missing is what happens without that review and after it: nothing says what automated checks apply to merges no patient reads, whether corrections reach data already contributed to research, or what an analyst sees when two sources conflict. The adverse event, adherence and sequencing analyses the life sciences FAQ offers carry no stated limits or review step.

The capability is described and the technology is not. The company says its proprietary AI models clean, normalize, link and enrich structured and unstructured records from many providers into longitudinal journeys without tokenization or manual stitching, and that paper records can be captured by camera.

No model, method or architecture is named, and nothing describes how entity resolution is done, how conflicts between sources are resolved, how accurate extraction from photographed documents is, or how and when the models change. For a product whose value is coherence across sources, the linkage layer is the technical proposition a buyer is purchasing.

The model layer is the company's own, and no other party is named. The life sciences site describes proprietary models doing the cleaning, linking and enrichment, and the data comes straight from patients' authorized connections rather than from brokers, vendors or clearinghouses, so there is no data supplier or tokenization party in the chain. The privacy policy says personal information may go to vendors, contractors and service providers, and names none; no hosting provider or subprocessor list is published.

Vendor reported scale and case examples, with no published validation. The life sciences site reports more than 20 years of longitudinal journeys, more than 4,500 records per patient, 100 times more data per patient than legacy platforms and more than 99 percent patient retention, without method or denominator, and the company describes registries (the GLP-1 and Metabolic Health Registry, 2026; Project EVOLVE in HER2 positive breast cancer, 2023) and blog case studies.

Its resources list carries announcements, awards and events rather than peer reviewed studies of the dataset's completeness, linkage accuracy or representativeness. Scale of use, however large, is not evidence of accuracy.

Categorical commitments are published without a retention schedule or a training use position. The patient privacy page says records are accessible only to the patient, that anything shared for research is anonymized first, and that contribution to research is a separate choice made with consent; the life sciences FAQ says all patient data is permissioned and de-identified.

The privacy policy keeps personal information for as long as needed for its purposes or as the law requires, which is not a schedule, and lists research with business partners and improving the services among its uses. Nothing says whether patient records train the company's linkage and structuring models, how de-identification is performed or tested, or whether a contribution can be withdrawn after it is shared. Nothing is published on safety engineering.

Compliance is claimed and no business associate position is taken. The patient privacy page says the company complies with the health information privacy regulations of the US Department of Health and Human Services, the life sciences FAQ calls its protections HIPAA-aligned, and the terms say account creation uses a checkbox labeled "I agree to the HIPAA Privacy Rules and Novellia's Terms of Use".

Records reach the company because patients direct their providers to send them to an app they control, which is the route under which a recipient app is usually neither a covered entity nor a business associate; the company does not say which it is. Substantive documents reach the service: a privacy policy covering the records it retrieves with patients' authorization, and a separate consumer health data privacy notice under state law.

Security is asserted with nothing independent behind it. The patient privacy page says health data is protected with top security standards and the app store listing describes robust, verified security systems; the privacy policy describes safeguards in general terms and says no security measure is perfect. No SOC 2 report, ISO/IEC 27001 or HITRUST certification, penetration test or trust center is published on the patient site, the life sciences site or in the legal documents.

Because records arrive by patient-directed disclosure, the HIPAA Security Rule may not reach them, which leaves the company's own safeguards as the protection for complete longitudinal records including images and laboratory results.

No device claim is made and the products are scoped as a personal health record and a research dataset. The terms of use say information in the app is for informational purposes only and not for treatment, and that some claims or interpretations on the service may not be validated by the medical community or FDA.

That is a general disclaimer rather than a stated regulatory position, and no page addresses how the dataset meets FDA's real world data reliability expectations (accrual, missingness, traceability to source) for the regulatory uses the company markets it toward. The adverse event prediction and burden modeling offered to life sciences customers are research analyses, not clinical tools.

Diversity is asserted and no evaluation is published. The life sciences site describes millions of patient-consented records from diverse populations across race, gender and geography, and the company announced a cross-industry initiative on disparities in breast cancer care in 2025. No cohort composition against a reference population, no evaluation of the record linkage or structuring models by group, and no governance framework is published. Two questions follow from the design.

The cohort is people who download a health app, connect their portals and consent to research, and the company says it concentrates on patients with serious and complex conditions, which is a selection mechanism a research buyer needs characterized. And record linkage has its own error modes, merging two people or splitting one, with no accuracy published.

A mechanism lets the person the record is about correct it, and nothing stands behind the output. The patient site says users can see their consolidated history, fix outdated information and add details themselves, so the data subject can challenge what the linkage produced in their own record.

Nothing published says whether a correction reaches data already contributed to research, and nothing answers to a research buyer for linkage or extraction errors: the terms of use disclaim any warranty that defects or errors will be corrected or that results will meet the stated purpose, and say some claims on the service may not be validated by the medical community or FDA. No error rate for record linkage, entity resolution or extraction from photographed documents is published.

Named record systems with read access, through the patient. The patient site says users connect records from more than 50,000 US health systems in one place, and the company names Epic, Oracle Health, athenahealth, Quest Diagnostics and the VA among the sources; records arrive through patient-authorized portal connections under federal access rules, with paper records captured by camera.

The access is read only: nothing is written back to any provider record, and the flow runs through each patient's own portal connections rather than institutional integrations, so there is no bidirectional integration running in production.

A single hosted service for United States patients, with location and separation undescribed. The patient app runs on iOS, Android and the web and connects to United States providers only; the privacy policy describes collection and use but names no hosting provider, region or retention schedule.

The design holds two estates, the identified record each patient reads and corrects and the de-identified research dataset licensed to life sciences customers, and nothing published says how they are separated, where either sits, or what a customer's access reaches. Radiological images, which the terms list among retrieved records, carry identifiers in metadata and sometimes in the image.

No price is published, and how the business charges is clear. The patient app and web platform are free, and revenue comes from de-identified, patient-consented data and insight products sold to life sciences teams: Novellia Atlas, a ready-to-use longitudinal dataset, and Novellia Precision, custom datasets built patient by patient with ePROs, virtual advisory boards and prospective tracking. No rate card, unit of charge or contract basis is published.

Coverage is stated clearly, with nothing yet published to validate it. Collection is US patients drawing records from US providers; the data is sold to biopharma teams in medical affairs, RWE and HEOR, commercial strategy, market access, clinical operations and R&D; and the dataset is said to span oncology, rare disease, cardiometabolic and neurologic conditions, with registries in GLP-1 and metabolic health and HER2 positive breast cancer.

Trial work is limited to Novellia Precision, which adds patients from a sponsor's studies or identifies patients through Novellia and follows them prospectively. No study of the dataset's depth or representativeness in any of these areas is published, and a cohort drawn from people who choose to use the app is not the same as the population a buyer studies.