Synapse Medicine

The machine learning here sits upstream of the clinical decision rather than inside it. Natural language processing reads manufacturer product documentation, health authority recommendations and research papers, and turns that unstructured text into the structured medication knowledge the platform then reasons over. So the model builds the corpus; conventional pharmacological logic makes the recommendation.

That is a different risk profile from a learned recommender, and arguably a harder one to audit, because an extraction error does not announce itself as a wrong answer once but propagates silently into every recommendation touching that drug. A second product line applies the same approach to pharmacovigilance, automatically coding and prioritizing adverse drug reaction reports, which is a clearer case of the model doing the work. Without the extraction layer the company would be maintaining a database by hand, which is precisely the incumbent product it was founded to replace.

The clinical products are advisory and reach a prescriber or a pharmacist who decides, with medication reconciliation producing a list for a clinician to confirm rather than an order, and the product privacy policy states that Synapse makes no automated individual decisions with legal effect.

The pharmacovigilance product occupies a different position that deserves separate attention: automating the coding and prioritization of adverse drug reaction reports means the model influences which safety signals a pharmacovigilance center examines first, and a report deprioritised is a report that may wait. That is oversight of the safety system itself rather than of an individual prescription.

Synapse publishes no acceptance rate, no override figure, no confidence threshold and no description of how a miscoded report is caught, and for the pharmacovigilance line that last question is the one that matters most.

The method and the sources are named: the company states that it uses natural language processing to analyze and classify medication information, identifies the three source types it ingests, and names the Theriaque drug database under its certified prescribing software. Its newer products, the Copilote suite and the MedGPT assistant for health professionals, add generative AI whose models are not named. What is absent is the number that matters for a system of this design.

An extraction pipeline turning regulatory text into structured drug knowledge has an accuracy, a recall and a failure mode, and none of those is published, nor is any description of how extracted content is validated before it reaches a clinician, whether a pharmacist reviews it, how often the corpus is refreshed against updated labeling, or how model versions are tracked.

For a product whose architecture rests on reading documents correctly, publishing the extraction accuracy would be the single most informative disclosure available and it is the specific ask to put to this vendor.

The data sources are named and the model parties are not. On sources the disclosure is real: the company states it uses natural language processing to analyze and classify medication information from manufacturer documentation, health authority recommendations and published research, and its certified prescribing software runs on the Theriaque drug database.

On parties, the product privacy policy (July 2026) names the host, Microsoft Azure's certified health data hosting in the European Union, and lists AI model providers among the recipients of personal data, without naming any of them or stating the terms that govern what they receive, including whether they may retain or train on it. The host is named, but the model layer is acknowledged and left unnamed.

The institutional signal is strong and the published evidence of benefit is thin. Deployments include Assistance Publique Hopitaux de Paris, the Hospices Civils de Lyon, the Bordeaux teaching hospital and more than a hundred hospitals in total, plus dozens of telemedicine companies, and the French national medicines agency and the national digital health agency are named as users.

A research partnership with the national health research institute on detecting drug interactions is on record, and the Bordeaux and Poitiers teaching hospitals are reported by the company to have extended its MedGPT assistant to all their teams after pilots. Its prescribing software is certified by type test against the French health authority's functional reference, which shows conformance to a specification rather than clinical effect. None of that is evidence of benefit: no outcome study, accuracy evaluation or peer reviewed publication on the product itself is published by the company.

Synapse Medicine's product privacy policy (updated July 2026) states retention, training use and de identification specifically enough to be contradicted. Patient data processed for care, including age, weight, sex, pregnancy, organ insufficiency, conditions and prescriptions, is kept 20 years from the last consultation in line with the medical council's recommendation; device vigilance data 10 years; research and statistics 3 years; product improvement 12 months.

Health data is used to evaluate, improve and develop the products only with explicit consent unless the processing rests on public interest research, usage data is pseudonymized under a user identifier, and statistics shared outside the company are aggregated and anonymized. Data is hosted on certified health data servers in the European Union and is generally not transferred outside it; the exception named is analysis of product use, under standard contractual clauses. Two gaps remain.

The policy lists AI model providers among recipients without naming them or stating whether they may train on what they receive. And nothing is published on safety engineering: how a wrong or missed alert from the prescription analysis modules is detected and handled beyond the vigilance obligations a device maker carries.

Synapse Medicine is French and states its role under the regime that governs it. Its product privacy policy (July 2026) says it processes personal data as a processor acting only on the instructions of the health professional, who is the controller, and as a separate controller for defined purposes such as device vigilance and consented research, with retention periods for each, a named data protection officer, EU hosting on certified health data servers, and the rights data subjects can exercise.

That is the equivalent of stating business associate status and supporting it with a substantive document. The company publishes no data processing agreement text, subprocessor list with names beyond the host, or subcontractor terms. It also runs an English site aimed at the United States, and no business associate agreement appears for any US covered entities it may serve.

No certification of Synapse Medicine's own is published: no ISO 27001, no SOC 2, no trust center, no penetration testing cadence and no vulnerability disclosure policy. The certification it does publish belongs to its host. The product privacy policy states that data sits on servers certified under the French health data hosting scheme (Hebergeur de Donnees de Sante), with Microsoft Azure named as that certified host, which answers the legal hosting requirement for French health data and is the host's attestation rather than the vendor's. The Class IIb device marking and the type test certification of its prescribing software concern clinical and quality conformance, not information security.

The regulatory position is stated and in the home market it is a device position. In February 2025 Synapse Medicine announced CE marking of its prescription analysis modules as Class IIb medical devices under the EU Medical Device Regulation (2017/745), five widgets and one API set marked separately (adverse effects, clinical analysis, drug interactions, dosage analysis and geriatric recommendations), with Synapse Medicine as manufacturer; every page of its site repeats the statement, adding that other components are Class I devices under the earlier directive.

Its prescribing support software, version 3.0 with the Theriaque drug database, is also certified by type test by SGS against the French health authority's March 2021 functional reference. The notified body and certificate number are not published, and nothing published says which modules and versions the certificate covers. The EU device database may hold that registration detail. No United States position is stated, which matters if the company sells there. No FDA warning letter or French safety notice names the company.

Synapse Medicine states a governance position without an evaluation behind it. It describes the CE marking of its prescription analysis modules as Class IIb devices under the EU Medical Device Regulation, announced in February 2025, as the framework under which its AI modules are developed, tested and validated.

The Regulation requires a quality management system, risk management, clinical evaluation and post market surveillance assessed by a notified body; it does not require an evaluation of performance across patient groups or a bias assessment, and neither is published. The product privacy policy adds that device vigilance data is kept 10 years. The exposure specific to this architecture is coverage rather than demographics, and it sharpens with the company's move into other markets.

A knowledge base assembled from French and European regulatory documentation inherits European approvals, labeling conventions and recommendations, so the same pipeline meeting American or Japanese labeling meets different documents and different ways of expressing the same warning.

Nothing describes how coverage and extraction quality are assessed when the corpus crosses a jurisdiction, or how results differ for patient groups such as the elderly, pregnant patients or those with organ insufficiency whose data the modules take in. Nor is there a published change plan for the models approved by the notified body.

The clinical products are advisory: prescription analysis reaches a prescriber or pharmacist who decides, and Synapse's privacy policy states it makes no automated individual decisions with legal effect. That is review before commit. Nothing stands behind the output from the vendor's side. The patient's data protection rights, including rectification, run to the health professional who is the controller, not to Synapse as processor, so they are not a correction route against this vendor.

The architecture has a specific and well defined gap: natural language processing reads regulatory and reference documents and turns them into structured medication knowledge, so the product's correctness is the extraction's correctness. An extraction pipeline has an accuracy, a recall and a characteristic failure mode, and none is published, nor is how extracted content is validated before it reaches a clinician or how often the corpus is refreshed against updated labeling.

A medication reference that lags a labeling change is confidently wrong rather than merely incomplete. The Medical Device Regulation requires a clinical evaluation for Class IIb devices, but its performance figures are not published.

The distribution model is the notable feature. Alongside direct hospital deployment at more than a hundred institutions including the largest French university hospital groups, the platform is sold modularly to software editors as widgets and an application programming interface embedded in the management software health professionals already use, and dozens of telemedicine companies embed it.

That makes this vendor a medication intelligence layer inside other people's products as well as a product in its own right. No hospital information system certifications, marketplace listings or interoperability standard conformance statements are published. French hospital software estates differ enough from American ones that this integration record says little about readiness for the United States market the company also addresses.

Hosting is on Microsoft Azure's certified health data hosting in the European Union, named in the company's product privacy policy (July 2026), which also states that data is generally not transferred outside the Union, with the stated exception of analyzing product use and improving the product under a recognized legal mechanism such as standard contractual clauses.

The product is delivered as cloud widgets and an API that partner software embeds, so patient data can reach the platform through a partner's product rather than through a hospital's own systems, and the contractual chain lengthens accordingly. Tenant isolation is not described, and the processing path for the named exception, what leaves the Union and to whom, is left open.

How the product is sold and what drives the cost can be established from Synapse Medicine's own material, though no price is published. The products are sold modularly as widgets and an API that software editors embed in the management software health professionals already use, and the product privacy policy records how access is metered: each professional holds a license type, a subscription start and end date, and primary or secondary user status, and Synapse invoices the partner editor.

So the cost scales with licensed professionals through a partner's contract, with direct hospital and pharmacovigilance deployments alongside. No price, unit rate, contract term or implementation fee is published. The company states that it is independent of the pharmaceutical industry, and its investor base of a healthcare professionals' insurer, a mutual insurance fund and public investment banks is consistent with that.

Coverage follows the medication rather than a specialty, and the settings are varied: acute hospital care at major university hospital groups, ambulatory care through the practice software that embeds its widgets, virtual care through telemedicine partners, pharmacovigilance through the adverse event coding product, and public health institutions including the medicines regulator and the digital health agency.

The prescription analysis modules take in the patient characteristics that change prescribing, including age, weight, pregnancy, breastfeeding and organ insufficiency, and include geriatric recommendations. The international footprint described in company material, with sites in English and Japanese and stated operations in the United States and Japan, rests largely on a 2022 funding announcement describing plans for the following two years, and nothing published since says whether local presence or validation outside France followed.