DermaSensor

Elastic scattering spectroscopy returns an optical signal and nothing a clinician could read. Every clinically meaningful output this device produces is inference performed on that signal.

The distinction matters because this is physical hardware and could be mistaken for an instrument with software attached. It is the reverse. A spectrum describing how light scatters through subsurface tissue carries no interpretation on its face; there is no reading mode, no threshold a physician could apply by eye, and no fallback operation without the model. The algorithm is what converts photons into a risk result, and the risk result is the entire product.

The regulatory record confirms it. What was authorised is an artificial intelligence enabled device, and the authorised algorithm is locked such that it cannot change without new studies and new authorisation, which is a statement about the model being the regulated article rather than the optics.

The hardware is genuine optical engineering and it exists to feed the classifier. Graded A on the same reasoning applied to the radar vendors in this index, where a sensor producing uninterpretable raw output makes the interpreting model constitutive rather than merely central. Every question worth asking about this product is a question about the algorithm.

The scope boundary is drawn by the regulator in negative terms, the operating point is published in full, and the effect on human performance was measured rather than assumed.

Labelling states explicitly that the device is not a screening tool, not a sole diagnostic criterion, and not intended to confirm a clinical diagnosis. Three negative statements defining what a product must not be used for is the clearest scope discipline in this index, and the company repeats the adjunctive framing in its own material rather than softening it in marketing.

The operating point is published rather than withheld: sensitivity of 96.3 percent across common skin cancers, standalone melanoma sensitivity of 90.2 percent, negative predictive values of 98.1 and 96.6 percent, positive predictive value of 61 percent at the highest risk band, and specificity of 20.7 percent. A physician can therefore know both what the device catches and what it costs them, which is what this index has asked of vendor after vendor.

Most importantly, the human machine system was tested. Three multi reader multi case studies with more than 280 physicians measured referral accuracy with and without the device, which evaluates the thing that actually happens in a clinic rather than the algorithm in isolation.

The residual risk is the one this design creates. With specificity at 20.7 percent, most positive results are benign, and a physician who learns that positives are usually nothing may begin discounting them, which is the failure mode an adjunct with a low specificity invites.

Ask what proportion of scanned lesions return positive in commercial use, and what guidance addresses alert fatigue.

The complete operating characteristic is in the public record, with one important asymmetry in how the company presents it.

What is published is comprehensive by the standards of this index. Sensitivity of 96.3 percent across common skin cancers, standalone melanoma sensitivity of 90.2 percent, negative predictive value of 98.1 percent for melanoma and 96.6 percent overall, positive predictive value of 61 percent at the highest risk band, and specificity of 20.7 percent. Study designs, site counts, enrolment figures and lesion counts accompany them, and the sensing method is named specifically as elastic scattering spectroscopy rather than described as proprietary optics. Peer reviewed publication means an independent reader can examine the derivation.

The asymmetry is worth stating plainly because it shapes how a physician will read the product. Sensitivity and negative predictive value appear prominently across company material; specificity of 20.7 percent is the number a reader is most likely to encounter first through independent analysis of the authorisation rather than through the company's own presentation. Both figures come from the same study and both are necessary to understand the device, and only one is foregrounded. That is selective emphasis rather than concealment, and it matters because specificity is what determines the workload a practice takes on.

Graded A because the full operating point is genuinely public and peer reviewed, which almost nothing else in this session achieves. Ask for the confusion matrix from the pivotal study and the risk score distribution observed in commercial use.

The model and its training basis are unusually traceable through the regulatory and published record, and the operational chain is undisclosed.

The favourable point is genuine. The algorithm is proprietary, developed in house, and locked under authorisation, so no third party model provider sits in the inference path and no external service interprets patient data. More unusually, the development and validation corpus is substantially documented rather than opaque: ten studies, more than 1,300 subjects, over 2,000 pathology verified lesions and more than 30 sites, gathered prospectively under trial protocols with ethical approval and informed consent. Most vendors in this index cannot say where their training data came from; here it came from clinical trials with a paper trail.

The hardware is the company's own instrument, and the sensing method is named rather than obscured.

What is not disclosed is everything around the device. No cloud or hosting provider for the portal, no sub processor register, no component or optics supplier, and no statement on whether commercially captured scans support future submissions.

That last question is live rather than theoretical. More than 20,000 lesions had been scanned commercially by late 2025, and expanding the evidence base is a stated use of the recent financing, so whether real world scans feed future authorisations, and under what consent, is worth establishing.

Ask for the portal hosting provider and sub processor register, the optics supplier, and whether commercial scans support future regulatory submissions.

A real clinical trial programme with the strongest study design recorded in this session.

The pivotal study is the reason. DERM-SUCCESS was international, multicentre, prospective and blinded, running at 22 sites with 30 primary care physicians across 1,005 patients and 1,579 lesions, with pathology verification as the reference standard. Prospective and blinded against biopsy confirmed truth is a materially higher bar than the retrospective dataset evaluations that constitute most evidence in this index. The wider programme comprises ten studies, more than 1,300 subjects, over 2,000 pathology verified lesions and more than 30 sites.

The utility evidence is equally well constructed and answers the right question. Three multi reader multi case studies involving more than 280 primary care physicians measured whether the device changes what physicians do, rather than only whether the algorithm is accurate. Aided physicians correctly referred 91.4 percent of cancers against 82.0 percent unaided, roughly halving missed cancer referrals. Measuring the human plus machine system is what an adjunctive device should be evaluated on, and most vendors here never attempt it.

Publication is real, spanning peer reviewed dermatology and primary care journals, and the authorisation attracted independent analysis in a digital medicine journal, which is external scrutiny the company did not commission.

Two honest limits. The utility studies are reader studies rather than real world deployments, so behaviour under time pressure with actual patients is untested. And no published outcome shows earlier stage at diagnosis or downstream benefit.

Ask for real world referral and biopsy yield data from commercial sites.

A dedicated pass located no encryption statement, no retention schedule, no access control model, no deletion process and no position on whether captured data contributes to model development.

Two included services establish that data movement occurs. A web based software portal accompanies every subscription, and remote device monitoring is a standard inclusion, which means the vendor maintains a connection to devices in clinical use and receives something from them. Nothing describes what.

The locked algorithm partly closes the question that usually matters most here. Because the authorised algorithm cannot change without new studies and new authorisation, captured data cannot be quietly feeding continuous model updates, which is a genuine structural protection and worth crediting. It does not mean data is not collected for future submissions, and no statement addresses that.

The device is a spectroscopic instrument rather than a camera, which is a real privacy advantage worth noting: an optical scattering measurement is not a photograph of a patient's skin, so the most identifiable artefact in dermatology technology is never created.

What remains unaddressed is the record. A stored result attaching an elevated skin cancer risk score to a named patient, held in vendor infrastructure with no published retention limit, is the exposure here.

One pre emptive note: further clinical evidence cannot move this grade. Only a published data handling position will.

Ask what the portal captures and retains, for how long, and whether data supports future submissions.

A dedicated pass located no health privacy position of any kind: no compliance statement, no business associate agreement template, no execution requirement and no processing terms.

The silence is easy to overlook because the device looks self contained, and the product description makes clear it is not. A web based software portal is included in every subscription, and remote device monitoring is listed as a standard service, so data leaves the practice and reaches vendor infrastructure as a matter of routine operation. What travels is unstated: whether lesion spectra alone, or accompanying patient identifiers, lesion location, images or clinical context.

The buyer type sharpens the omission. This is sold to primary care and urgent care practices, which are covered entities but are also small organisations without procurement functions or in house counsel. They are the least equipped of any buyer in this index to discover a missing agreement and the most likely to sign whatever is placed in front of them, which is precisely when a vendor's published posture matters most.

A result recorded against a patient indicating elevated skin cancer risk is sensitive, and where it is stored and who can retrieve it are unaddressed.

Graded D because nothing was located after a dedicated search. Ask what the portal transmits and stores, for the agreement template, and whether the practice or the vendor controls the resulting record.

A dedicated pass located no security page, no external attestation, no trust centre, no penetration testing statement, no vulnerability disclosure policy and no documentation offered under agreement.

One distinction needs stating because this vendor's regulatory record invites conflation. De Novo authorisation involved review of device safety and effectiveness, and modern device submissions include cybersecurity documentation, but authorisation is not an information security attestation and covers the device rather than the vendor's hosted infrastructure or corporate controls. This record does not credit one as the other.

The surface extends beyond the instrument. A web based portal holds clinical results, remote monitoring maintains connections to devices in clinical use, and firmware presumably updates over that channel. A connected medical device with a remote management path is a recognised category of risk, and nothing describes device authentication, update signing or how the fleet is protected.

The buyer type is the aggravating factor rather than a mitigating one. Small primary care practices lack the security review capability that a hospital would bring, so they will not surface these questions themselves, which places more weight on the vendor publishing a position unprompted.

One pre emptive note: restating regulatory authorisation cannot move this grade. Only an external information security attestation, or documentation available under agreement, will.

Ask whether an information security attestation is held, how devices authenticate and receive firmware, and what documentation can be shared.

De Novo authorisation, which is the hardest regulatory route represented in this index.

Authorisation was granted on 17 January 2024 following Breakthrough Device designation. The De Novo pathway applies where no predicate device exists, so the applicant must establish that a novel device type is safe and effective and, in doing so, creates the classification that later entrants will use. That is a materially heavier undertaking than clearance against an existing predicate, and it makes this company the reference point for artificial intelligence enabled skin cancer detection rather than a follower of one. The resulting classification is Class II adjunctive diagnostic. European marking is held alongside.

The locked algorithm provision deserves particular attention. The authorised algorithm cannot change in any way without new studies and new authorisation, which means the device a practice buys today behaves identically to the device evaluated in the pivotal trial. Across this index the more common pattern is a model that evolves continuously with no published change control, so a buyer cannot know whether current behaviour matches published evidence.

The labelling limits are part of the regulatory achievement rather than a caveat on it, since the negative statements about screening and sole diagnostic use were part of what made authorisation possible.

What is not published by the company is the authorisation number or the full order text, though both are matters of public record.

Ask for the De Novo number, the complete indication for use, and what a future algorithm update would require.

The one bias question that matters most in this field is addressed directly and with regulatory backing, which no other record in this index can claim.

Skin tone is the defining failure mode of dermatology artificial intelligence. Published image datasets underrepresent darker skin, several widely reported tools have performed worse on it, and the clinical consequence is severe because melanoma outcomes are already substantially worse for patients with darker skin who present later. Against that background this device is authorised across all six Fitzpatrick skin types, clinical studies included patients across the full range, and performance is stated as consistent regardless of pigmentation.

The modality may be doing genuine work here rather than the claim alone. Elastic scattering spectroscopy measures subsurface tissue optical properties rather than surface appearance, so the confound that defeats image based models is less directly in play, and that is a plausible mechanism rather than only a marketing assertion.

The locked algorithm adds a second governance property, since performance cannot drift silently away from what was evaluated.

What holds this short of the top is that the claim is stated without the numbers. Consistent performance across skin types is asserted rather than published as sensitivity and specificity per type, and consistency at a sensitivity of 96 percent and consistency at a specificity of 21 percent are different findings with different implications. No breakdown by sex or by age band within the eligible population was located, and no external algorithmic audit exists.

Ask for sensitivity and specificity by Fitzpatrick type, and enrolment counts per type in the pivotal study.

No contractual terms are published, and this vendor supplies the two things that make responsibility assessable.

The first is a published operating point. Sensitivity of 96.3 percent across common skin cancers and 90.2 percent for melanoma standalone means the miss rate is known: roughly one in twenty five common skin cancers and one in ten melanomas will not be flagged. A practice adopting this knows what it is accepting, which is the precondition this index has written into several records as the only thing that would move a liability grade. The second is the labelling, which states in negative terms that the device is not a screening tool, not a sole diagnostic criterion and not a diagnostic confirmation, so responsibility is explicitly retained by the physician and the device's role is bounded by regulatory language rather than by contract alone.

The locked algorithm reinforces both, since the device in use is the device that was evaluated.

What is absent is conventional and real. No service level agreement, warranty, indemnity or remediation position was located, and no published terms describe recourse if a device malfunctions or a result is wrong.

The exposure a practice actually faces runs in the direction of over referral rather than under detection. At a specificity of 20.7 percent most positives are benign, and a physician who refers on every positive generates volume the dermatology system may not absorb, while one who begins discounting positives has quietly lost the benefit.

Ask what the agreement warrants on device function, and what support exists for referral volume management.

A dedicated pass located no electronic health record integration, no interface standard, no marketplace listing and no published application programming interface.

The result is delivered on the device screen, and a web based software portal accompanies the subscription. Neither reaches the chart. Nothing indicates that a risk score, the lesion assessed or the referral decision posts to the record as a discrete entry.

The consequence is specific to how this device is used. A primary care physician scans a lesion, receives a result, and must then type that result into the note by hand for it to exist in the patient's record at all. That makes the finding dependent on manual transcription for continuity, for any subsequent dermatology referral to carry the context, and for the practice to demonstrate that the assessment occurred.

Billing makes the same gap consequential in a second way. Reimbursement support is an included service, and a service that cannot be evidenced in the record is difficult to substantiate to a payer, so the absence of chart integration works against the reimbursement pathway the vendor is helping practices pursue.

The buyer type compounds it, since small primary care practices have the least capacity to build integrations themselves.

Graded D consistent with the treatment of an absent record system integration elsewhere in this index. Ask whether results post to any record system, through what mechanism, and how practices document the assessment for billing.

A dedicated pass located no hosting provider, no region statement, no residency commitment, no tenancy model and no description of what leaves the device.

The physical side is straightforward and favourable. This is a handheld instrument used at the point of care, and inference appears to occur on the device itself, since a result is returned on screen within seconds during a consultation. Local inference means the clinical function does not depend on connectivity, which is the right design for a device used in a primary care room.

The hosted side is undescribed. A web based software portal is included in every subscription and remote device monitoring is a standard service, so vendor infrastructure exists and devices communicate with it. Where that infrastructure sits, what it receives, and whether patient identifiable data reaches it is not stated anywhere.

European marking implies an intent to deploy under a regime with data location requirements, and nothing addresses how that is handled or whether processing differs by region.

One question follows from the service model and is unanswered. Remote device monitoring implies devices report status and possibly usage, and whether that telemetry includes clinical content, scan counts or results is precisely the distinction between benign fleet management and a clinical data flow.

Ask where the portal is hosted, what remote monitoring transmits, whether inference is fully on device, and how European deployments are handled.

The commercial model is described in structure and inclusions, with no figures attached.

What is published is more than a contact form. Pricing runs as subscription plans scaled to anticipated device usage, and the included services are enumerated: customer success and support, certified reimbursement support, a web based software portal, remote device monitoring, and on demand clinical resources. A one time device setup and activation fee also exists, disclosed indirectly through a promotional waiver, which tells a prospective practice there is a capital component alongside the subscription. A practice therefore knows the shape of the commitment and what it covers before enquiring.

Usage based subscription is also the right model to disclose here, because a primary care practice can estimate lesion volume from its own patient mix and reason about whether the tier fits.

What is missing is every number. No subscription rate, no usage tier boundaries, no setup fee amount and no contract term was located.

The reimbursement position is the more consequential unknown and is only partly addressed. Reimbursement support is included as a service, which implies billing is possible and non trivial, and no established coverage policy or payment rate was identified. Independent commentary notes limited public signal on durable coverage, which leaves practices absorbing the cost directly rather than recovering it. For a primary care practice operating on thin margins that determines adoption more than the subscription rate does.

Ask for the subscription tiers and rates, the setup fee, the contract term, and what codes practices currently bill successfully.

A single indication in a single setting, bounded tightly by the label, addressing a genuinely underserved gap.

The scope is defined by authorisation rather than by marketing. The device is indicated for use by non dermatologist physicians, on lesions raising concern for melanoma, basal cell carcinoma or squamous cell carcinoma, in patients aged 40 and above. Each of those is a real constraint: a dermatologist is outside the intended use, a patient of 35 is outside the population, and a lesion the physician is not already concerned about is outside the workflow since the device is not a screening tool.

The setting choice is well judged rather than narrow by accident. Dermatology training is limited in most family and internal medicine residencies, only around 16 percent of primary care clinics report access to a dermatoscope or training in its use, and referral access is constrained in many regions. Building for the front line rather than for specialists targets where the diagnostic gap actually sits.

Coverage of skin types is complete across the Fitzpatrick range, which broadens the population meaningfully within the age bound.

What limits the grade is that everything else is out of scope. No paediatric or younger adult use, no dermatologist workflow, no lesion tracking over time, no body mapping, and no second condition. Geographic reach covers the United States commercially with European marking held.

Ask what evidence exists for use below age 40, and whether the age bound reflects a data limit or a performance limit.