Lifen DataLab
Lifen DataLab is the multicenter research product from Paris based Lifen, the document platform used by more than 800 French healthcare facilities. It turns medical reports into structured data for real world evidence studies, cohorts and registries. Each project starts with a scientific committee that validates the variables and processing rules. Patients join a cohort on an ongoing basis, and their data updates as new documents enter the patient file with no manual step.
Language models adapted by Lifen extract the variables from pseudonymized reports, research assistants check low confidence results in a hybrid workflow, and site medical teams trace every value in a dedicated interface. Researchers can add a variable after launch and rescan earlier reports, then follow inclusion on dashboards and export to analysis tools.
Running cohorts include the LUCC lung cancer cohort at Gustave Roussy, CUB Trajectory for chronic bronchial disease with Hôpital Foch, the HARPE hereditary angioedema registry and France PAD, a national network for lower limb arterial disease. In a ten center study published in Annals of Oncology, model extraction erred on 7.0% of data points against 14.2% for manual abstraction, and on 4.4% with targeted review, while structuring time fell from 17.5 minutes to 4.6 with the hybrid method. DataLab runs on dedicated servers in a certified French health data hosting cloud, and sites join through a demo without heavy upfront investment.
Capability Axes
An AI Health Index grade measures what a buyer can verify from public sources on the date shown. It is not a rating of how good the product is. A vendor can build an excellent system and grade low on an axis because it publishes nothing an outsider can check. How grades read
Language models do the product's central job, reading pseudonymized letters and reports and extracting the predefined variables each cohort needs. Lifen adapts open source models on annotated health data to classify documents and locate the relevant facts, and a project can add a treatment history or biomarker variable after launch and rerun extraction over earlier records. Dashboards, exports and the trace interface all sit on that structured output. Take the models away and participating centers would be back to abstracting research fields by hand, the 17.5 minute manual method the hybrid workflow replaced, before any of the analytics could run.
A scientific committee approves each study's variables and rules, the software then extracts fields and updates the cohort as reports arrive, and research assistants validate or correct predictions where the model signals low confidence, ambiguity or unusual context. Site medical teams can trace every resulting value. In the LUCC evaluation, reviewing the 30% of answers with the lowest confidence brought errors down to 4.4%, a study configuration rather than a fixed rule.
The structure is clear about who reviews what. What it leaves out is the confidence threshold each cohort runs at in production, who settles a disputed value after review, and how an eligible patient the model failed to include would come to light.
Lifen uses pretrained open source language models adapted on annotated health data for classification and extraction, and research staff corrections feed back into improving them. The ten center lung cancer study names the pipeline behind its results, lucc-ai-1, a fine tuned model that took pseudonymized PDF reports and extracted 31 variables, both single answer and multiple choice.
Adding a variable after launch and rescanning earlier reports changes the project's dataset, not the production model. What goes unnamed is the base model family and version now running across DataLab projects, the makeup of the annotated training data, and how a model update is versioned and announced to studies whose earlier extractions came from the previous release.
Extraction runs on open source models that Lifen adapts on annotated health data, so no commercial model provider sits in the chain. Lifen's May 2026 subprocessor list for DataLab names S3NS for hosting sensitive reports in France and Azure services in France for detection and machine learning over those reports, alongside providers for authentication, postal notices, rights handling and other operations. Naming Azure on the machine learning path goes further than a hosting mention.
The list stops short of saying which model packages run on Azure, whether prompts or intermediate outputs persist there, and what retention and training terms govern that traffic. Should Lifen move hosting outside France, the charter commits it to let the institution choose the country.
A peer reviewed study in Annals of Oncology compared the Lifen supported LUCC pipeline with manual chart abstraction across ten French centers, with a test cohort of 311 patients. Across 8,708 data points for 28 single answer variables, model extraction erred on 7.0% against 14.2% for manual abstraction, and reviewing a targeted 30% of lower confidence answers brought errors to 4.4%. A blinded expert adjudicated discrepancies in the reference values.
Lifen also reports 91.9% concordance with manually entered data across 1,057 thoracic cancer patients and structuring time of 4.6 minutes with the hybrid method against 17.5 manually. Lifen supported the study, and its results hold for that pipeline and lung cancer cohort.
Reports are pseudonymized before research use, with direct identifiers replaced by consistent aliases so follow up continues across documents. Lifen's privacy charter sets retention for the health data it processes as a processor: patient data used to apply AI predictions stays three rolling months in the active base, then a rolling year in intermediate archive before deletion, and document and appointment data keeps a rolling year in each.
At the end of those periods the institution chooses deletion or return. Training use is the gap. Lifen tunes its models on annotated health data and uses research staff corrections to improve them, but the terms do not say which inputs or corrections are eligible, pseudonymized research datasets carry no separate period, and no safety incident process is described.
Lifen acts as a processor for the participating institution when DataLab structures its documents, extracting set variables from pseudonymized reports only for the research or care purposes the controller defines. Sites keep their duties as controllers, and patients exercise their right to object through each study's own arrangements.
The charter binds Lifen's onward processors by contract to limit their access to the service they provide and to keep equivalent security, and the DataLab subprocessor list names them. GDPR and the HDS hosting framework are the governing regimes, documented to the depth a business associate arrangement would be in the United States. No US business associate agreement is offered, which fits a product sold to French research hospitals.
Lifen is certified as a health data host under France's HDS framework, holds ISO 27001 for its information security management and is an approved operator of the MSSanté secure messaging service. DataLab's heavy processing runs on dedicated servers inside that certified hosting cloud, with S3NS named for sensitive report storage.
The privacy charter commits Lifen to give a controller proof of its certifications on request, notify it within 30 days of any change of certification body and share the latest certification audit report. Those artifacts come through a request rather than a self service trust portal, and the certificates are company level, so their scope for the research service is set out in that exchange rather than on the product page.
DataLab is built for research: multicenter cohorts, registries and real world evidence studies, with extracted variables feeding study datasets rather than decisions about a patient in front of a clinician. Lifen designed it to the CNIL reference framework for health research, informs patients and honors their right to object under the GDPR, and Lifen Research holds a CNIL authorization to run a health data warehouse for partner institutions.
That authorization covers the research data infrastructure and its contributing cohorts. No diagnostic or treatment claim is made, and the product on offer matches the research purpose and setting the company describes, under a named European data regime rather than device regulation.
Each project has a scientific committee that validates the variables and processing rules before extraction runs across participating centers, and site medical teams trace and correct results in the same interface, a named governance process at the study level. The ten center lung cancer evaluation also compared error by institution and found less variance between sites with model extraction than with manual abstraction.
That shows consistency across contributing hospitals within one cancer cohort. It does not show how extraction performs by patient age, sex, disease, document type or the writing style of different specialties, and no fairness method or outside audit of the models is offered for the other cohorts and registries.
Every extracted value can be traced to its source report in the review interface, and research assistants correct predictions before a dataset goes to analysis. The LUCC study in Annals of Oncology gives the error rate behind that workflow: 7.0% for model only extraction across 8,708 single answer data points and 4.4% after targeted review, with discrepancies in the reference values settled by a blinded expert. A research team can test its own cohort against those figures.
Data subjects' rights requests are routed to the controlling institution. Beyond correction, nothing stands behind a wrong variable, since no accuracy warranty for cohort data or remedy for lost study work appears in the published terms.
Patients enter a cohort on an ongoing basis, and their data updates as new documents reach the patient file, with no manual step at the site. The service is presented as deployable whatever information system the contributing hospital runs, and the ten center lung cancer study took pseudonymized PDF reports as input. Data leaves as structured variables in exports to analysis tools.
No electronic health record is named as a source connector, no standard such as HL7 or FHIR is described for the research feed, and nothing is written back into a patient's chart. Lifen's document integration product, which files reports into hospital records, is sold separately and is not described as part of DataLab.
Dedicated servers in a certified French health data hosting cloud run DataLab's processing, so participating sites contribute documents without operating the heavy compute themselves. The DataLab subprocessor list places sensitive report hosting with S3NS in France and machine learning over those reports on Azure services in France, and the charter commits Lifen to let the institution choose the country if it ever contracts a host outside France.
Authentication, notification and rights tools come from other named suppliers. Storage and model processing are therefore placed in France. How one project's or one hospital's data is kept apart from another's inside the shared research cloud, and which transfers outside those functions are permitted, are left open.
Research groups and hospitals start with a demo and a scoping conversation, and Lifen says sites can join without heavy upfront investment. Projects take two shapes, a study a single center initiates or a collaborative cohort with other institutions, and the cost follows the number of centers, the inclusion rules, the variables extracted, how often the cohort refreshes and the analysis access each group needs. A downloadable brochure describes the service in more depth. No price, rate card or unit of charge appears, so whether site count, cohort size, extraction volume or workspace access drives the fee is settled project by project.
Hospital research departments, teaching hospitals, cancer centers and multicenter consortia use DataLab, within a Lifen research network of more than 30 establishments in France. Gustave Roussy runs lung, pediatric and prostate cancer cohorts, Hôpital Foch runs CUB Trajectory for chronic bronchial disease, and the HARPE hereditary angioedema registry and the France PAD network for lower limb arterial disease also run on it.
Prof. Benjamin Besse and Dr. Mihaela Aldea of Gustave Roussy speak for the product. Validation covers part of that range: the ten center study and the 1,057 patient concordance figure both come from thoracic oncology, and no extraction results are given yet for the respiratory, vascular or rare disease cohorts.
Alternatives to Lifen DataLab
Other products that do the same kind of work as Lifen DataLab, and what each one does.
HealthKey helps research practices find potential clinical trial participants in their existing patient records. Teams supply a protocol's inclusion and exclusion criteria, and the software applies those criteria across structured EHR fields, clinical notes, and documents such as faxed PDFs.
ConcertAI's Accelerated Clinical Trials platform helps pharmaceutical sponsors and contract research organizations design, plan and run oncology studies.
Mendel, based in San Jose and backed by Oak HC/FT and DCM, sells Hypercube, a platform that reads unstructured patient records for chart review, cohort building and clinical trial prescreening.
CluePoints provides clinical trial data quality and oversight software for pharmaceutical sponsors and contract research organizations.
Frekil builds AI agents that run observational real world evidence studies for pharmaceutical, biotech and contract research teams on data they already hold or license, from protocol through final report.
CancerLinQ is ConcertAI’s oncology software for practices, cancer centers and health system oncology programs. It turns longitudinal records into care quality measures, treatment context and clinical trial matches.
Pricing
Vendor-published figures are labeled as such. Figures labeled “Estimated” are derived from third-party sources and have not been confirmed by the vendor.
| Entry Price | Pricing Basis | BAA Tier | Implementation | Source |
|---|---|---|---|---|
|
Project quote
|
Lifen scopes DataLab by project, either a study one center initiates or a collaborative cohort across institutions. The variables, inclusion rules, update frequency, participating sites and analysis access define each engagement. | Lifen acts as a processor under the GDPR and the French HDS hosting framework, with the study agreement setting responsibilities between sites. No US business associate agreement applies. | Heavy processing runs on Lifen's dedicated hosted servers, and Lifen says sites join without heavy upfront investment. No setup, cohort configuration or per site fee is listed. | Vendor Published |
No rate card, per cohort fee, per site minimum or separate price for exports appears. A demo and project scoping are the way in.