Implicity

The platform's entire commercial argument is that models decide what a clinician sees, and without them the product is a viewer for data the manufacturers already provide.

The aggregation layer is real engineering. Normalising raw discrete data from Abbott, Biotronik, Boston Scientific, Medtronic and Microport into one interface, rather than working from the summary documents each portal produces, is difficult and valuable. It is also not the thing being sold. A unified viewer would reduce the number of logins and leave the alert volume untouched, and alert volume is the problem clinics are paying to solve.

Three cleared algorithms carry that weight. The ILR ECG Analyzer reassesses arrhythmia detections from insertable cardiac monitors. The atrial fibrillation alert management algorithm filters device notifications and classifies episodes against guideline thresholds. SignalHF analyses physiological trends to anticipate decompensation. Each is separately cleared, and clearing algorithms in their own right rather than as device features establishes them as the product.

The framing confirms it. The company describes its technology as artificial intelligence driven, positions alert filtering as the core capability, and quantifies its value entirely in model terms: 79 percent fewer false positives, 85 percent fewer unnecessary atrial fibrillation alerts. Those are model performance figures presented as the business case.

Graded A. Strip the algorithms out and a clinic still receives every alert its devices generate, which is the situation the product exists to change.

The autonomy is a filter rather than a diagnosis, and the company has tuned and reported it in a way that shows the safety logic has been thought through.

What the models do is decide which alerts reach a clinician. That is consequential in one direction specifically: a suppressed alert is an event nobody sees, and unlike a wrong diagnosis there is no reviewer positioned to catch it. So the entire safety question is sensitivity, and the company has tuned for it explicitly, reporting sensitivity of 98.3 and 94.3 percent across device generations while accepting specificity of 61.6 and 75.6 percent. Choosing to let non actionable alerts through rather than risk suppressing real ones is the correct tradeoff for this application, and publishing both numbers rather than only the flattering one is how a reader can verify the choice was made deliberately.

The classification logic is anchored externally rather than proprietary. Atrial fibrillation episodes are classified in line with European Society of Cardiology recommendations, so the thresholds separating actionable from non actionable follow published guidelines a clinician can inspect rather than a vendor's internal judgement.

The regulatory framing is also correct, with the software described as an adjunct to a remote monitoring platform for follow up rather than as a standalone diagnostic.

What holds this at B is that nothing published describes what happens to filtered alerts. Whether suppressed notifications remain retrievable, whether a clinic can audit what was filtered, and whether any sampling review of suppressed events occurs are all unaddressed, and those are the controls that would let a clinic verify the filter locally rather than trusting the published sensitivity.

Graded B.

The most completely documented performance claims of any record built in this session, achieved through footnoting rather than through a technical white paper.

The company publishes numbers on both sides of every claim. Sensitivity 98.64 percent with a false positive rate of 24.03 percent, given as a fraction so a reader can see the denominators: 509 of 516, and 68 of 283. False positive reduction of 79 percent alongside maintained sensitivity of 99 percent on Medtronic insertable monitor recordings. Alert reduction of 85 percent for atrial fibrillation. Patient reconnection of 28 percent within two days. Each carries a footnote to a study, a product code or an internal data observation with a stated date and population, and where a figure comes from internal data rather than publication the company says so.

The intended use statements are published verbatim, including the limitation that a cleared analyser supports only one manufacturer's monitor data, and the inputs to the heart failure model are named individually: arrhythmia burden, physical activity, nighttime heart rate and thoracic impedance.

Also disclosed is the architecture at platform level, cloud based, hosted on named public cloud infrastructure, no local installation, always current version, and the classification logic anchored to published cardiology society recommendations.

What is missing is model architecture and training corpus size. The development data source is named on the stewardship axis, and how large the training set was and what network design was used are not stated.

Graded A on the completeness and referencing discipline of the performance disclosure.

Development data provenance is named and externally governed, upstream data sources are enumerated, and model composition is undisclosed.

The development corpus is the strong element and it is unique in this session. The company states it develops its artificial intelligence using access to the Health Data Hub, identifies it as a French government established platform combining existing patient databases for research and development, and characterises it as one of the world's largest heart disease patient databases. A named, state operated source with its own authorisation regime is provenance a third party can verify, which is not true of any other training corpus disclosed in this session.

The operational data supply chain is equally clear. Five device manufacturers are named as the sources of the raw discrete data the platform ingests, so a buyer knows exactly which upstream parties feed the system and can reason about dependencies, including the risk that a manufacturer changes its data interface.

Infrastructure is named through the public cloud provider and marketplace listing.

What is absent is the models themselves: no architecture, no training set size, no acknowledgement of third party frameworks or pretrained components, and no bill of materials for a cleared Class II device. Nothing states whether the three cleared algorithms share a common technical foundation or were developed independently.

Graded B: the hardest provenance question in this index, where training data came from, is answered here better than anywhere else, and the model composition question is not answered at all.

The most honestly reported performance data of any vendor built in this session, and the honesty is what earns the grade rather than the numbers.

The EHRA 2026 work is the distinguishing element. Across implantable cardiac monitor generations the algorithm held sensitivity at 98.3 percent in devices carrying their own onboard artificial intelligence and 94.3 percent in those without, with specificity of 61.6 and 75.6 percent respectively and positive predictive value around 74 percent in both. Publishing specificity in the sixties is the opposite of what marketing incentive suggests, and it tells a clinician the actual shape of the tradeoff: this filter is tuned to miss almost nothing and will therefore pass through a substantial volume of non actionable alerts. That is the correct tuning for the clinical problem and most vendors would have reported only the sensitivity.

The supporting body is substantial and independent in venue. A retrospective study of more than 4,000 patients across four manufacturers published in the Cardiovascular Digital Health Journal, with a named academic cardiologist commenting on the clinical rationale. A study cited as reducing false positives by nearly 80 percent on more than 2,800 arrhythmia episodes. Separate published work on reconnecting patients who drop off remote monitoring, reporting 28 percent reconnected within two days.

Deployment is stated at more than 110,000 patients across over 250 facilities in the United States and Europe, which is real scale for a company of this size.

Graded A, and the record notes that the company also publishes its own instructions for use figures, sensitivity 98.64 percent with a false positive rate of 24.03 percent, giving both sides of the same coin.

The best training data provenance disclosure in this session, and the only one where the source is a publicly accountable institution rather than a private arrangement.

The company states it develops its artificial intelligence using access to the Health Data Hub, and explains what that is: a health data platform established by the French government to combine existing patient databases and facilitate their use for research and development, described as one of the world's largest heart disease patient databases. That single disclosure does several things no other record here manages. It names the source rather than gesturing at leading medical centres. It identifies a source with its own statutory governance, authorisation process and ethical review, so the consent and access basis exists independently of the vendor's assurances. And it is verifiable, because the platform publishes its own rules and its authorisations are matters of public record.

The practical effect is that a reader assessing whether these models were built on properly governed data has somewhere to look. On every other record in this session that question terminates in silence.

What is still missing is the operational side. Nothing states whether customer transmissions are retained beyond clinical use, whether they inform continued model development, what retention applies, or whether customers can decline secondary use. For a platform holding raw discrete cardiac data on more than 110,000 patients, that is a real gap.

Graded B rather than higher on that omission, and well above the session norm because the development corpus, which is where these questions usually go unanswered entirely, is disclosed and externally governed.

One substantive fact is disclosed that no other record in this index carries, and the contractual layer is unpublished.

The substantive fact concerns model development data. The company states it has access to the Health Data Hub, the health data platform established by the French government to combine existing patient databases and facilitate their use for research and development, described as one of the world's largest heart disease patient databases. That is a named, publicly accountable, state operated data governance framework with its own authorisation process, ethical oversight and access conditions. Naming the source of development data and having that source be a government platform with published rules is materially better provenance than the undescribed corpora on most records in this session, and it means a reader can go and examine the governance rather than take the company's word for it.

The operational data flows are extensive. The platform ingests transmissions from implanted devices across five manufacturers for more than 110,000 patients, holds raw discrete cardiac data rather than summaries, runs on public cloud infrastructure, and generates billing documentation. Every part of that involves identifiable patient data.

What is absent is the contract. No business associate agreement is offered or described for United States customers, no protected data handling summary exists, no retention position is stated, and nothing describes the European data protection basis for a French company processing United States patient data or the reverse.

The company states its solutions meet the highest security standards, which is an assertion rather than a posture.

Graded C.

No published security posture was located beyond an assertion. The company states that its solutions meet the highest security standards, and nothing sits behind that sentence: no trust centre, no service organisation control report, no information security management certification, no penetration testing statement, no vulnerability disclosure policy, no subprocessor list and no incident notification commitment were found.

The gap is more conspicuous here than the company's size would suggest, for three reasons. The platform holds raw discrete cardiac data on more than 110,000 patients across two continents, which is a substantial and sensitive archive. It ingests transmissions from five manufacturers' infrastructure, so it sits inside a data path involving several large medical device companies whose own security requirements for partners are demanding. And it is a European company subject to data protection law that expects demonstrable technical and organisational measures, which most vendors in that position evidence through a recognised certification.

A marketplace listing on a major public cloud implies the provider's own vendor requirements were met, which is assurance about the listing rather than about the software.

Regulatory clearance across FDA and CE routes means cybersecurity documentation exists in the technical files for the cleared algorithms, assessed by regulators rather than published for customers.

The contrast with this vendor's own behaviour elsewhere is the notable part. This is a company that footnotes its performance claims to specific studies, publishes both sensitivity and false positive rate, and names its development data source. On security it offers a single unsupported adjective.

Graded D on published evidence.

Three separately cleared algorithms across two regulatory systems, with the class and directive cited precisely rather than summarised.

The ILR ECG Analyzer holds FDA 510(k) clearance as a Class II device and CE marking as a Class I device under the medical device directive, with the intended use stated exactly: assessment of arrhythmias in insertable cardiac monitor electrocardiographic data, by qualified healthcare professionals, interfaced with remote monitoring platforms that supply the data. The company also states plainly which devices it supports, initially Medtronic insertable monitors, rather than implying universal coverage.

The atrial fibrillation alert management module is CE marked as a Class I device under the same directive, with a product code and the reimbursement provision both cited. SignalHF is stated as an FDA cleared Class II device and described as the first algorithm of its kind cleared and compatible across multiple manufacturers.

What distinguishes this from a simple clearance count is the referencing discipline. The company footnotes its claims to specific product codes, device classes, directives and study citations throughout its material, so a reader can trace each assertion to what authorised it. Publishing the intended use text verbatim, including the limitation on which manufacturer's data a cleared algorithm supports, is the behaviour of a company that expects to be checked.

One consideration a buyer should note: Class I CE marking under the older directive is self certified, and the in vitro diagnostic and medical device regulations have since replaced that framework, so European status may need refreshing.

Graded A.

Performance is broken down by a dimension that genuinely matters, and not by patient characteristics.

The breakdown that exists is technical and it is unusually thoughtful. The EHRA 2026 work reports performance separately for implantable cardiac monitors that carry their own onboard artificial intelligence and those that do not, finding sensitivity of 98.3 versus 94.3 percent and specificity of 61.6 versus 75.6 percent. That is a vendor examining how its filter behaves when the upstream device has already filtered, which is a real and non obvious source of variation, and reporting that its own algorithm performs differently depending on what generation of hardware produced the signal. Most vendors report a single aggregate figure and let a buyer assume it holds everywhere.

The multi manufacturer evidence base extends the same logic, with the atrial fibrillation study drawing on more than 4,000 patients across four manufacturers, so performance is demonstrated across the heterogeneity the platform claims to handle.

What is absent is the patient dimension entirely. No performance figures by sex, age or comorbidity were located, and no model card or training population composition is published despite the development data source being named. Cardiac device populations skew older and male, and atrial fibrillation burden and heart failure decompensation patterns differ by sex, so a filter tuned on that distribution may behave differently for the patients least represented in it.

Graded C: genuine and unusual technical subgroup reporting, no demographic reporting.

No published liability position, and the failure mode here is more clearly defined than on most records, with the published data allowing a buyer to size it.

The risk is suppression. A filter that decides which alerts a clinician sees fails dangerously in one direction only: an event filtered out is an event nobody reviews, and there is no downstream reader positioned to catch the omission. That is a genuinely different liability shape from a diagnostic model, where a clinician sees the output and can disagree.

What partly answers it is the published performance. Sensitivity of 98.3 and 94.3 percent across device generations, and 98.64 percent stated in the instructions for use with the fraction given, mean a clinic can quantify the residual risk rather than accept an assurance. A vendor that publishes both sensitivity and false positive rate has given a customer the material to make an informed decision about accepting the filter, which is a substantive form of accountability even without a contractual term.

Regulatory clearance adds manufacturer accountability with post market surveillance and adverse event reporting attached, and the intended use text positions the software as an adjunct to a monitoring platform rather than as the monitoring itself, which places responsibility with the clinic operating the service.

What is missing is everything contractual and one operational control. No indemnity, limitation, performance warranty or service level was located, and nothing describes whether filtered alerts remain retrievable or auditable, which is the mechanism by which a clinic could investigate a suspected miss.

Graded C.

Interoperability is the entire product rather than a feature of it, and it is demonstrated at the hardest layer in this market.

The achievement is manufacturer neutrality at the data level. The platform aggregates, normalises and standardises transmissions from Abbott, Biotronik, Boston Scientific, Medtronic and Microport, and does so from raw discrete data supplied by the manufacturers rather than from the summary documents their portals generate. Working from raw data is the difference between a genuine unified record and a document repository, and it is what allows a single algorithm to run across every device a clinic follows. Cardiac device manufacturers have no commercial incentive to make this easy, so building it across five of them is a substantial and defensible position.

The cleared algorithms inherit that neutrality. SignalHF is described as the first algorithm of its kind cleared and compatible across multiple manufacturers, which turns the aggregation layer into a clinical capability rather than only a convenience.

Downstream integration is addressed through automated billing workflows and report generation, with customers reporting more than 90 percent billing compliance, so output reaches the administrative systems that determine whether the service is paid for. The platform is delivered as a cloud service with no local installation, available through a public cloud marketplace, which lowers the deployment barrier considerably.

What is not documented is electronic health record integration specifically. No record system is named and no interoperability standard is cited, so how a reviewed transmission reaches the patient chart is unstated.

Graded A on the strength of the manufacturer layer, which is the harder problem and the one this market actually has.

The hosting platform is named and the regional specifics are not, on a company operating across two regulatory regions from a European base.

What is disclosed is more than most. The solution is cloud based, hosted on a named public cloud provider and available through that provider's marketplace, with no local installation and no customer managed updates, so every clinic runs the current version. Naming the infrastructure provider tells a security reviewer which underlying control environment applies before any conversation begins, and marketplace availability means some buyers can procure through an existing cloud agreement.

What is absent is region and residency. The company is headquartered in Paris with a Cambridge, Massachusetts presence, and monitors more than 110,000 patients across the United States and Europe. So European patient data is processed by a company with United States operations, and United States patient data by a French company, and nothing published states where either is stored, whether regional segregation exists, or what transfer mechanism applies. For a French company subject to European data protection law and serving American clinics, that is the question a buyer on either side would ask first.

The Health Data Hub relationship adds a further consideration, since that platform's own governance includes conditions about where data resides and who may process it, and nothing describes how development work under that authorisation relates to the commercial infrastructure.

No subprocessor list, retention position, export term or availability commitment was located, and availability matters here because a filtered alert stream sitting in a diagnostic pathway needs to be running.

Graded C.

No price is published and two structural disclosures make the commercial model unusually legible for this index.

The first is reimbursement. The company states that its atrial fibrillation alert management module is covered under a named article of the French social security code, which is a specific citation to a reimbursement provision rather than a vague claim of coverage. For a French clinic that is the number that matters, and citing the legal basis is more useful than a price.

The second is the billing capability itself. The platform automates billing workflows and reporting and the company reports customers achieving more than 90 percent billing compliance. Remote monitoring reimbursement depends on documented periodic review, and clinics routinely lose revenue by failing to evidence it, so a platform that generates the documentation is selling into a measurable revenue gap. That is a commercial argument a practice can evaluate against its own billing data.

Availability through a public cloud marketplace is a third route, which for some buyers means procurement through an existing cloud agreement rather than a new vendor contract.

What is missing is every figure. No per patient or per clinic price, no module structure across the platform and the three algorithms, no implementation cost and no contract term. Whether the cleared algorithms are included in a base subscription or licensed separately is unstated, and with three distinct cleared modules that is the question a buyer would ask first.

Graded C.

Manufacturer coverage is the coverage claim on this record, and it is complete in a way nothing else in the index achieves.

Five manufacturers are named and supported: Abbott, Biotronik, Boston Scientific, Medtronic and Microport. That is effectively the entire cardiac implantable device market, and the significance is structural rather than commercial. A clinic following implanted patients has whatever devices were implanted over the past decade, by whichever surgeons, so a platform that covers four of five leaves a residue requiring a separate portal and a separate workflow. Vendor neutrality here is not a philosophical position, it is the difference between replacing a process and adding to one.

Device type coverage runs across pacemakers, implantable defibrillators, resynchronisation devices and insertable cardiac monitors, spanning both arrhythmia and heart failure populations.

Clinical coverage extends beyond rhythm. The heart failure module tracks arrhythmia burden, physical activity, nighttime heart rate and thoracic impedance as an indicator of pulmonary fluid accumulation, so the platform addresses decompensation as well as arrhythmia, which are the two reasons these patients are followed.

Geographic coverage spans the United States and Europe with regulatory standing in both, and user coverage is explicitly designed for the range of staff involved, with material aimed separately at physicians, and at the nurses and technicians who do most remote monitoring review in practice.

Graded A.