Linus Health

Machine learning is the engine of the flagship scoring, and the platform around it carries other value. The Digital Clock and Recall (DCR), the objective half of the Core Cognitive Evaluation, is scored by trained models: the company's 2025 paper in Alzheimer's Research and Therapy describes random forest classifiers over nearly 2,000 drawing, pause, stylus and speech features, and the assessments page says DCTclock and the DCR use machine learning algorithms to measure what a paper clock test cannot.

Anywhere for Health Systems, the remote battery, is described as AI assessing response speed, hesitation and verbal fluency, and the Radar engine sold to health plans flags members with possible undiagnosed cognitive impairment from claims and clinical data.

Remove the models and the company still sells a tablet test administration platform, the Life and Health Questionnaire, personalized action plans and care pathways, the ePSOM outcome measure, speech and hearing screeners and EHR connected workflow. That is the model as the engine of a core module, not the whole of what is sold.

The oversight structure is described, and the handling of a wrong or discordant result is not. Provider reports include the clock drawing and recall recordings behind the score, so a clinician can inspect the raw performance; Anywhere for Health Systems gives the primary care provider a positive or negative result with a breakdown by test and domain and the option to replay the patient's session; action plans and care pathway guidance suggest next steps.

The terms of use state the platform is not intended as a stand-alone or adjunctive diagnostic device and is not intended to determine the level or type of care, and the company's 2024 paper sets the DCR impairment threshold at a score of 3 or below.

What is missing is the stopping rule: nothing published says when a result must go to full neuropsychological testing or biomarkers, what a clinician should do when an amyloid estimate and the clinical impression disagree, or how a result later shown to be wrong is handled.

The approach and the training data are published; versioning and update practice are not. The company's 2025 paper (Jannati et al., Alzheimer's Research and Therapy) describes the Digital Clock and Recall models as random forest classifiers over nearly 2,000 features from the command and copy clock tasks, immediate and delayed recall, Apple Pencil signals (altitude, azimuth, pressure), pauses and speech and voice, plus age, trained on the 930 participant Bio-Hermes-001 cohort with nested leave-one-out cross-validation and cutoffs set by the concordance probability method.

The 2024 paper sets out the DCR score and its impairment threshold of 3 or below, and the code is available from the company on request. The assessments page says DCTclock and the DCR use machine learning algorithms, and Anywhere for Health Systems is described as AI assessing response speed, hesitation and verbal fluency, with no method published for it.

No version identifier or update practice for the commercial models appears on the site, in the research library or in the terms, which reserve the right to modify features with notice only where a change affects how the platform is used, so a buyer cannot tell when a scoring model changed.

The chain is structurally short at the model layer and the rest of it is published. The models are the company's own, trained on research cohorts such as Bio-Hermes-001, and nothing describes a third party model receiving assessment data.

The subprocessor page (updated 28 March 2025) names each party with its purpose and location: Amazon Web Services and Google for hosting, DataDog for logging and security monitoring, Okta through Auth0 for authentication, RPost for secure email, Twilio for SMS, Zendesk for support and Zoom for support and training, and says each is reviewed before engagement and bound by a data processing agreement.

The gap is in the terms that govern data once it arrives: the list gives purposes, not retention or use limits per party. Radar came with the acquisition of Together Senior Health and the gait and balance tools with Kinesis Health Technologies, and both are now the company's own. The company also holds a license to use deidentified customer data for product improvement.

Peer reviewed, prospective multi site evaluation with the method readable, and the external results are moderate. In the Bio-Hermes-001 study (930 participants at US sites, organized by the Global Alzheimer's Platform Foundation and funded by the Alzheimer's Drug Discovery Foundation), the company's 2025 analysis in Alzheimer's Research and Therapy reports DCR models classifying cognitive impairment at AUC 0.83 and amyloid PET status at 0.81, against 0.70 to 0.73 for the MMSE, RAVLT and Cognivue on amyloid, with plasma p-tau217 alone at 0.89 for amyloid.

External validation is weaker: sensitivity 0.72 and specificity 0.67 (negative predictive value 0.49) for impairment in 134 memory clinic patients, and sensitivity 0.65 and specificity 0.73 against plasma p-tau217 in 373 APHELEIA participants. The 2024 Bio-Hermes paper (706 participants) reports AUC 0.70 for the DCR against 0.63 for the MMSE.

Both papers are led by company staff with the co-founders among the authors, and the underlying study data are governed by the Global Alzheimer's Platform. The research library also lists primary care comparisons with the MoCA and SLUMS. The remote battery's launch release claims 91 and 95 percent accuracy against a three hour neuropsychological evaluation with no study, denominator or method named.

Retention and deidentification are stated specifically, and so is a license to use customer data for product improvement. The privacy policy (updated 15 July 2026) says protected health information is kept only for the term of the contract with the provider, then transferred to the provider and deleted, and deleted on a verifiable individual request; that facial data, collected only for the VisMET gaze assessment, and voice recordings are not de-identified; and that PHI is encrypted and stored apart from unrelated non-PHI data.

Against that, the terms and conditions of sale grant the company a perpetual, irrevocable license to use customer data in aggregated, de-identified form "for the sole purpose of enhancing and improving the Linus Health Services", and the terms of use allow personal data to be used to test, benchmark, enhance and improve products. Training on customer assessments is therefore permitted rather than excluded. Nothing is published on safety engineering: guardrails on the models, how a model error is handled as a safety event, or how a model change is validated before release.

Business associate status is stated and the agreement is not published. The privacy policy (updated 15 July 2026) says the company handles PHI only when acting for a licensed medical provider under a business associate agreement in the United States, or a data processing agreement elsewhere, and the terms and conditions of sale say it may be deemed a business associate and will sign an agreement with the customer at contract.

The life sciences page calls the Healthcare Delivery platform HIPAA compliant. The same documents set out retention and deletion, AES-256 encryption and European standard contractual clause obligations. The agreement, the tier it applies at and its subcontractor terms are not published. The platform is also sold to health plans, where Radar reads claims and clinical data, and to trial sponsors, and which of those routes the agreement covers is unstated.

A SOC 2 Type 2 attestation is named in the company's own material without the report. The life sciences page, answering a question on data safety, calls the Healthcare Delivery platform SOC 2 Type 2 certified and HIPAA compliant. The trust center, hosted on Vanta, shows no content without a signed browser token, so the report, its period and its scope are not visible from outside.

The privacy policy describes AES-256 encryption, PHI stored in SOC 2 certified cloud environments (the hosting provider's certification, not the company's), an incident response plan and regular vulnerability monitoring, and the subprocessor page says each subprocessor's security practices and certifications are reviewed before engagement. No ISO/IEC 27001 or HITRUST certification and no penetration testing cadence is stated on the site, the privacy policy or the terms.

The position is stated and verifiable, and it is a listing, not a clearance. The FDA registration and listing database shows Linus Health, Inc. (registration 3014174644) listing the Linus Health Platform under product code PTY, computerized cognitive assessment aid, exempt, Class II, 21 CFR 882.1470; no 510(k) is held under the company's name.

The site says the company is registered with the FDA, MHRA and Health Canada; the life sciences page lists FDA listed Class I and II devices, CE and UKCA marks and registration under Health Canada's MDEL; and a December 2024 release announces EU MDR Class IIa certification for the Core Cognitive Evaluation. The terms of use and terms of sale state the platform is not intended as a stand-alone or adjunctive diagnostic device. Two scope questions remain.

The classification describes use by clinicians in a healthcare setting, while Anywhere for Health Systems is sold as an FDA listed Class II device completed at home. And the company's research presents the DCR as estimating amyloid positivity, while the listed device is an aid for assessing cognitive function.

Two company led papers on the Digital Clock and Recall publish subgroup evaluation with a stated method, and the weak spots are reported. Jannati and colleagues (Alzheimer's Research and Therapy, 2024; 706 Bio-Hermes participants) compared score bias by race (White against non-White), ethnicity (Hispanic against non-Hispanic) and education (15 years or more against fewer) using rank sum tests, Poisson regression and a bootstrapped bias comparison against the MMSE.

The DCR score showed no ethnicity effect where the MMSE did, a significant difference by race when accounting for cohort where the MMSE's did not reach significance, and a significant education effect, as the MMSE did. The 2025 paper in the same journal (930 Bio-Hermes participants) reports classification by APOE e4 carrier status: AUC 0.80 against 0.84 for amyloid PET status and 0.82 against 0.84 for cognitive impairment, lower in carriers and not significantly different.

Its authors state the cohort, more than 24 percent non-White, may not represent every population where the test is used, and its stratified analysis covers APOE status only, so no performance split by race or sex is published for the amyloid model. The research library also lists Harvard Aging Brain Study work relating amyloid and tau burden to faster DCTclock decline.

A published error rate with its method and denominator, and nothing standing behind the output. The company's own 2025 paper (Jannati et al., Alzheimer's Research and Therapy; first and corresponding author a Linus Health employee), listed on its published research page, reports the Digital Clock and Recall models on external data.

In 134 memory clinic patients at the Rowan University and Virtua New Jersey Institute for Successful Aging, 95 of them impaired on full neuropsychological evaluation, the cognitive impairment model had sensitivity 0.72 (0.62 to 0.81), specificity 0.67 (0.51 to 0.82), negative predictive value 0.49 and positive predictive value 0.84. In 373 APHELEIA participants, 121 of them positive for plasma p-tau217, the amyloid model had sensitivity 0.65 and specificity 0.73.

These are test set results rather than performance in use, and the amyloid check uses a blood marker as the reference instead of PET. The terms of use disclaim accuracy and cap liability at $10,000, and the indemnity in the terms and conditions of sale covers intellectual property claims only.

Named record systems with results flowing in one direction. The athenahealth Marketplace release (June 2023) says the platform is integrated with athenaOne so results appear in the EHR, with Healthtopia Clinics quoted on results arriving in its record, and a neurology chair at an unnamed California health system describes integration with Cerner on the health systems page.

The Anywhere for Health Systems release says assessments can be assigned from major EHRs with results returned, and the life sciences page calls the platform integrated with all major EHRs, without naming them. No integration documentation, standard such as FHIR or HL7, or production site running ordering and results both ways is published.

A single hosted service with residency committed by region. The subprocessor page (updated 28 March 2025) places hosting on Amazon Web Services in the United States, with Ireland for EU customers and London for UK customers, and authentication through Okta in the United States, Dublin and Frankfurt for EU customers and the United Kingdom for UK customers.

DataDog, Twilio, Zendesk, Zoom and RPost are listed in the United States only, so some processing for customers outside the United States crosses a border. The terms of use say information is stored in geographically diverse locations, and the privacy policy says PHI is held apart from unrelated non-PHI data; tenant isolation is not described. Assessments run on an iPad with Apple Pencil in clinic, and Anywhere for Health Systems runs at home on a laptop, tablet or phone.

No price is published, and how the product is sold is discoverable. The terms and conditions of sale set fees by order form or statement of work, payable within 30 days of invoice, with iPads and styluses either supplied by the company or bought by the customer.

The company's billing guide (July 2026) explains how a practice bills the Core Cognitive Evaluation as cognitive testing under CPT 96138 or 96136, with 96132 for evaluation, the time minimums and the medical necessity rules, and shows Medicare's 2024 non-facility rates; those are what the provider recovers from the payer, not the vendor's charge. Health plan and trial sponsor work is sold through a sales conversation.

Coverage is named across settings, with published validation behind in-clinic use and not behind the remote route. The company sells to medical practices and health systems for early detection of cognitive impairment in primary care and specialty triage, to health plans for risk identification and diagnosis capture, and to trial sponsors for prescreening and outcome measurement.

The Digital Clock and Recall is sold to detect cognitive impairment and, in the company's research, to estimate amyloid positivity. Validation covers a research cohort (Bio-Hermes-001, adults about 60 to 85), a memory clinic (Rowan University and Virtua) and primary care comparisons with the MoCA and SLUMS. Anywhere for Health Systems, completed unsupervised at home, carries an accuracy claim in its launch release with no study named, and the research library lists no evaluation of it. Radar's accuracy is described in a release citing a Milliman report rather than a published method.