Medidata

The AI layer is now sold as a tier of its own, on a trial platform whose value stands without it. Medidata Plus, launched in July 2026, is described as a clinical AI layer that unlocks the platform's AI capabilities, including the Dot assistant, AI study build that generates forms, edit checks and synthetic test patients, no-code data transformation, embedded risk and safety signal analytics, and an agent for querying audit history. Those capabilities do not exist without the models.

What most sponsors license is the platform around them, anchored by Rave data capture, Clinical Data Studio's data integration and review, imaging, trial management and payments, which work without the AI layer.

Human in the loop is stated throughout; one capability shows the control. Health Record Connect submits nothing to Rave until the site saves the form, so a person confirms every record value used. For the rest the company says every AI capability has a human in the loop, without describing the checkpoint: whether generated forms, edit checks and synthetic test data from AI study build reach a live study without approval, whether redacted images are sampled or verified, and what the audit-history agent in Medidata Plus may do. Redaction is where a missing step matters most, because a reviewer looking at a redacted image cannot see what was missed.

The approach and training data are described; versions and base models are not. The company describes proprietary machine learning, patented algorithms, generative AI and predictive modeling, trained on an anonymized dataset of more than 42,000 trials and 13 million patients, developed under a formal AI Validation Framework and monitored after launch.

Company papers show methods behind specific tasks: Siamese network embeddings for mapping capture fields to SDTM and random forests for cardiovascular event detection. The imaging redaction uses automated text detection. The base models behind the generative features, including Dot, are not named, and no version history or update notice is published for a validated environment.

The model layer is described as the company's own; no model provider is named publicly. The company calls its AI proprietary and trains it on its own anonymized trial corpus, and the ISO/IEC 27001 certificate places hosting in its Houston data center. The generative features, including the Dot assistant, imply a language model whose origin is not stated. A sub-processor list exists on the trust page but sits behind a customer login, and the public data processing exhibit binds sub-processors without naming them.

Named customers and company research, with no method behind the product figures. Clinical Data Studio customers include Eisai and Worldwide Clinical Trials; the product page reports up to 80 percent shorter data review cycles, 90 percent less time to build listings, 50 percent faster patient profile review and 83 percent faster signal to action, and Health Record Connect reports 60 percent fewer queries and 70 percent fewer data changes, none with a study, denominator or method.

Company researchers have published peer reviewed methods work: SDTM field mapping with Siamese network embeddings above 90 percent accuracy (PLoS One, 2024) and random forest detection of cardiovascular events against adjudication panels (Journal of Biopharmaceutical Statistics, 2024). The company says its AI has supported more than 500 studies; that describes use, not benefit.

A categorical commitment on secondary use is published, and the training corpus is described; retention is not. The trust and transparency page says the company does not use data for secondary purposes without the customer's explicit written permission, reviews every internal data use request, and runs a Data Collaboration Program in which aggregated, anonymized data sits in a separate environment.

The AI pages say Dot is built on an anonymized dataset of more than 42,000 trials and 13 million patients and gets smarter with every decision, and company research trained models on anonymized trial data from its enterprise data store. A public data processing exhibit with standard contractual clauses, ISO/IEC 27701 privacy certification and a SOC 2 audit covering privacy stand behind it.

Not stated: the retention schedule, the anonymization method, and whether a sponsor outside the collaboration program contributes to the corpus. Nothing is published on safety engineering for the generative features: guardrails, hallucination handling or how a safety event is handled.

A substantive processing document reaches the service; business associate status is stated nowhere. The company publishes a data processing exhibit with standard contractual clauses and holds ISO/IEC 27701 privacy certification and a SOC 2 audit covering privacy, which is the right paper for study data processed for sponsors under GDPR. No page states whether it signs business associate agreements or takes a position on where HIPAA applies.

That matters for Health Record Connect, which pulls identifiable records from provider systems and health information exchanges into trial data capture; in the United States a business associate relationship with the providers is the ordinary consequence, so the gap matters to the provider more than to the sponsor.

Certifications are named with version and scope and published as documents anyone can open. The trust and transparency page links the ISO/IEC 27001:2022 certificate (DQS, registration 10014794, covering the Medidata Clinical Cloud and its operations and development, certified 27 Oct 2025, valid to 29 Oct 2028, with the New York, Houston, London and Iselin sites in the annex), ISO/IEC 27017 and 27018 statements of conformity, the ISO/IEC 27701 privacy certificate, a Cloud Security Alliance STAR registry entry, an information security white paper in four languages, a program fact sheet, a statement on security incidents at study sites and Dassault Systemes' vulnerability disclosure program.

The company also names a SOC 2 Type II audit with privacy criteria run twice a year. Penetration test results and vulnerability scan summaries are listed but sit behind a customer login, as does the SOC 2 report.

No device claim is made for the AI capabilities, and they are scoped as trial operations, with an unusually full account of the rules that do apply. The company's 2025 data integrity white paper, public on the trust page, sets out how the platform meets 21 CFR Part 11 and FDA guidance on computerized systems and electronic source data in clinical investigations, MHRA GxP data integrity guidance, the EMA guideline on computerised systems and electronic data in clinical trials, EU GMP Annex 1, and Japanese (MHLW) and Chinese (NMPA) requirements; per-instrument position statements sit in a compliance library in the company's knowledge hub.

Separately, the FDA registration and listing database shows Rave Imaging listed by Medidata as a Class I medical image communications system (21 CFR 892.2020, product code LMD); the automated redaction is not described as part of a device. These are the company's own readings, not a regulator's finding.

A governance framework with named process is published; no evaluation of how the models behave across groups is. The trust and transparency page says AI principles are grounded in the NIST AI Risk Management Framework and ICH good clinical practice, that every AI capability passes a formal AI Validation Framework review before deployment and is monitored after launch, that internal data use requests are reviewed by the privacy office, compliance and information security, and that humans stay in the loop.

Nothing published shows results by site, region, therapeutic area, source system or patient group. For data review the material question is whether flagging is even: a detector can be accurate overall while over-querying some sites or sources. For study build, a model trained on prior protocols will reproduce their eligibility conventions. For image redaction, detection may vary by document type or language.

Outputs can be checked before they land; nothing stands behind them. Health Record Connect presents matching record data to the site as it completes Rave forms and submits nothing until the site saves the form; AI-enabled audit trail review sits in Clinical Data Studio; and every AI capability is described as validated before deployment and monitored after.

No error rate is published for any indexed capability: the figures are speed and volume (up to 80 percent shorter data review cycles, 90 percent faster listing builds, 32 percent faster redaction, 60 percent fewer queries). Company researchers have published an SDTM field mapping method above 90 percent accuracy (PLoS One, 2024) and a MACE detection model with recall 0.72 and precision 0.68 (Journal of Biopharmaceutical Statistics, 2024), neither tied to a named product version.

Automated redaction fails silently: a missed identifier travels onward as if redaction succeeded, and no recall figure or verification step is published. No warranty, indemnity or remediation commitment is published.

Record data is read into trial capture through standards and exchanges, with one named partner system. Health Record Connect reads patient records through national and local health information exchanges and FHIR, C-CDA, DICOM and HL7, and presents matching values to sites as they complete Rave forms; it combines record and eSource data through an integration with CRIO. Clinical Data Studio brings in non-Medidata sources, including third party laboratories and other data capture systems.

Data flows one way, from the record into the trial; no named record system, marketplace listing or integration documentation shows a production connection to a specific EHR, and the exchanges are not named. The company reports 60 percent fewer queries and 70 percent fewer data changes without a method.

Hosting sites and transfer terms are published; region election and tenancy are not. The ISO/IEC 27001 certificate annex names the data center operation in Houston, Texas, within the Medidata Clinical Cloud scope. The trust page publishes French health data hosting (HDS) information, China data processing documentation, a data processing exhibit with standard contractual clauses, a transfer impact white paper, and ISO/IEC 27017 and 27018 statements of conformity for cloud controls and cloud personal data. The sub-processor list sits behind a customer login. Whether a sponsor can elect a region, and how competing sponsors are isolated on the shared platform, are not stated publicly.

No price is published, and how the AI is sold is now discoverable. Enterprise agreements with sponsors, contract research organizations and medtech companies are scoped by study count and module. Medidata Plus, launched in July 2026, packages the AI capabilities as one layer the company describes as a predictable investment rather than per-tool costs.

Coverage is named and validated for part of it. The platform serves biopharmaceutical and medtech sponsors and contract research organizations across study design, build, execution and closeout, with therapeutic area solutions in oncology, vaccines and CNS; Health Record Connect also feeds Medidata Surgical Planning and real world data work that reaches beyond trials. The AI capabilities have named customers for data review and company research on SDTM mapping and event detection; no published results cover AI study build, redaction or the Dot assistant.