> On January 17, 2024, the US Food and Drug Administration (FDA) announced it had __cleared DermaSensor[0]__ as the first AI-powered medical device able to...
From [0] (emphasis my own)
> On Friday, the FDA authorized __for marketing__ the DermaSensor Inc. DermaSensor device. It is a prescription device, indicated for the evaluation of skin lesions suggestive of melanoma, basal cell carcinoma, and/or squamous cell carcinoma in patients aged 40 and over to assist health care providers in determining whether to refer a patient to a dermatologist. __The device should be used in conjunction with the totality of clinically relevant information from the clinical assessment, including visual analysis of the lesion, by physicians who are not dermatologists.__ __The device should be used on lesions already assessed as suspicious for skin cancer and not as a screening tool.__ __The device ,,SHOULD NOT,, be used as the ,,sole,, diagnostic criterion nor to confirm a diagnosis of skin cancer.__ The FDA is requiring that the manufacturer conduct additional post-market clinical validation performance testing of the DermaSensor device in patients from demographic groups representative of the U.S. population, including populations who had limited representation of melanomas in the premarket studies, due to their having a relatively low incidence of the disease.
Even the product's website specifically says "clearer"[1]. And Reuter's article from 2 days ago[2]
@dang, I think this merits at minimum a title change but probably a link change. Current link is misleading and appears to be much more marketing focused.
I also thing the marketing release contradicts the terms of the cleared status by the FDA. The FDA specifically says this is not for screening or diagnosis. So essentially what does this device actually help with? A physician needs to be suspecting something abnormal with a lesion before using. If it is "negative" would a physician really want to trust a device in the off chance it misses a melanoma? If I were a family doctor and was suspicious of a lesion enough to use this device, wouldn't it make more sense to send to derm or just go ahead an biopsy?
It's not even that: it's a device for non-dermatologists, to decide if they need to make a referral. "Cleared" is definitely the right language here, this might only be a Class 1 medical device, it's not making any diagnostic claims.
There is an app for assistance in skin cancer detection (SkinIO) but of course it’s more limited by the cellphone’s camera vs. the device in the OP. I can see both approaches co-existing, especially as SkinIO’s database expands and mobile phone cameras and AI improve. But it would be great to have the professional device become cheap and small enough that it can be a phone accessory, or perhaps phones can pack more sensors if it somehow makes sense (ie if they can augment phone functionality in other ways).
Thanks for leaving the comment, because it was my first thought too.
What's peculiar is its accuracy doesn't appear to be much better than a visual only app (based on a quick google of iphone app accuracy levels).
Could be similar to the 'should self driving cars be visual only' debate - it might seem intuitive that more data sources should improve results, but in practice it's not necessarily the case.
Image resolution, lighting, and skin pigmentation will affect the accuracy of any image analysis app. I have a feeling a lot of those apps exist solely to collect and resell data. It doesn't actually matter that they are accurate or not, what matters is that you were concerned and that data point can be resold with the rest of the data you give them.
In general med tech = so expensive your eyes bleed and your brain melts.
I'm assuming they've dropped in price a pretty good amount now, but digital x-ray machines with the require software were near hundred thousand price range and typically tied around some windows operating system. So if the controller came with XP, never expect to go to a new operating system. The companies general answers were, oh, you need a controller with a new OS, spend another $100k for the next gen, which unless you're pumping out huge numbers of xrays was just not going to happen.
>So if the controller came with XP, never expect to go to a new operating system.
That's not necessarily a bad thing!
In our industry, the idea of something using XP is ridiculous. Applying our experience to the medical field is a bad idea!
It's an appliance. It doesn't really matter what OS it runs as long as its stable and certified to do. It spits out imagery which you can then process and view on any compatible device - often a reasonably modern computer running some sort of health informatics platform.
For what it's worth, the "XP" OS is often licensed and supported for well longer than a consumer SKU. IIRC microsoft provided support and security updates for embedded/appliance applications until 2019!
These devices are often not openly networked, or if they are, they are through some sort of gateway that's appropriately locked down and updated.
It's an appliance with huge gaping security flaws that its very creator says you should use under approximately no circumstances.
> For what it's worth, the "XP" OS is often licensed and supported for well longer than a consumer SKU. IIRC microsoft provided support and security updates for embedded/appliance applications until 2019!
You're 100% right but that still means the thing has been sitting around for half a decade with no security updates whatsoever.
There is a fine line between stability and simply being a miser and refusing to update things within a reasonable period of time. Healthcare, especially corporate healthcare run by people other than physicians, is far to much on the miser end of the spectrum.
> In our industry, the idea of something using XP is ridiculous.
Only because of the lack of (security, drivers) updates though right? If I couldn't use Linux or macOS I'd rather use XP (if it were current) than anything that's come since.
(I also don't need accessibility features which I'm sure have come a long way.)
Being in the medtech/AI industry (our company, Eyenuk, created the 2nd autonomous AI approved by the FDA), nothing seems too strange here.
I've seen subscriptions supported, and the pricing doesn't seem alarming. If you look at any medical bill that includes lab work, or anything medical related, this is really cheap comparatively.
The cost of everything is required to lure companies into the space. The cost of research, running a clinical trial, the FDA approval, certifications (SOC2, etc.), insurance, etc. still push many companies away. In fact, in the EU, where the cost is lower, the approval process is getting harder (MDR) and many companies are leaving the market.
Why can't companies just charge X price, once, since it's already FDA approved (hence I assume models are validated), and in the future any upgrades cost Y?
It comes across as predatory that users be charged subscriptions for something that seems like it should be fixed.
Aren't other medical devices and drugs sold for a fixed price, regardless of all the FDA approval and marketing they had to go through?
Like, no one pays a subscription fee for an X-ray machine, or a subscription for a course of antibiotics.
Antibiotics are consumables - in a B2B context (it's a medical service provider who is paying the subscription fee), antibiotics are basically a subscription.
Many (most?) diagnostic devices have consumables which is where companies make the bulk of their revenue. Failing that, service contracts are usually in place to lock in some sort of sustained reoccuring revenue.
There is no fundamental reason why a company couldn't do your proposed model, but upgrades are definitely challenging - if you make your model better, or you improve your hardware, or basically do anything that isn't a bug fix or an ancillary upgrade (like I dunno, you made the thing boot up 25% faster), will likely require additional testing and approval from the FDA.
One of the genius medical marketing tactics was to create disposable condoms for thermometers, ear probes, etc., so that you can use one up for each patient and then purchase a whole big box of them on a subscription basis, rather than going through a simple bottle of sanitizer that could clean anything and everything that needs it. (And then of course, nobody cleans the probe underneath, as it just gets slightly grodier with each use.)
That's not why disposables were adopted. A "simple bottle of sanitizer" is less likely to get rid of biofilm and spores, especially given the laziness and error of end-users.
Disposables are inherently safer. Lots of trash, and more expensive, but the best option for wealthy nations that can afford the best care.
I wouldn't pay for a 25% faster bootup unless it was an emergency life-saving device. You'd have to sell me something significantly better and worth my money, like any other seller in the world has to do with their product, from computers, to cars, to fast food.
I'm not sure that a software upgrade would require FDA approval, but product testing is a cost incurred by anyone in any industry. Even a new burger recipe would be tested at the restaurant's own expense. That doesn't seem to justify a subscription model (after paying the initial cost of the device, etc).
There are a lot of ongoing costs as well. Medical insurance is expensive, and there's a lot of process that is ongoing. A lot of effort goes into security, both monitoring and pushing updates. Medical device companies must also undergo regular audits of our quality systems, look up QMS/eQMS, where FDA reps will spend about 1 week a year reviewing everything we've done over the last year. Generating all the documents for those audits takes a lot of time (aka cost). That's not complete list, but this is in addition to regular SaSS costs.
An x ray machine does the same thing all the time, no Software maintenance needed. If you have a product that has a cloud component, any component to it where it needs to be maintenance from a security standpoint, that has cost, you need to cover that cost.
Modern X-ray machines use software to receive the image and view it.
If the models already work, why would it even require a cloud component? That's just risky, complex, and costly. Let the device do its thing, doc writes down the findings, data gets deleted, and device ready for next patient.
They aren't using AI or the cloud component to do what they can already do.
The whole promise that all these companies are pitching is to aggregate patient data over time, tie it to clinically measurable metrics so they can detect, prevent or treat things better. In your example the idea would be the model in the cloud is aggregating all the patient health record data, tying that to the xrays of those same patients, so that in the future the model can see things in x-rays that indicate disease or are predacessors to disease before we humans can even spot them.
Not quite the example I was explaining. In my example the validated/FDA-approved model runs on local hardware, at the physician's energy bill. Need an A6000 GPU? Fine, I'll buy it.
If the vendor wants to improve the model, they can ask for anonymized data from physicians, if they're willing to provide it. That's their cost of developing a better product, just like anyone in any industry pays a price in some way to improve their product.
A bit late here, but I thought I'd mention something in addition - most hospitals and doctors don't want to take on the burden or supporting something new and specialized. They will focus on providing care and buy support. Some have small IT departments, but even IT support is regularly contracted out.
This stuff can't run locally, it's cloud connected by design. It's cloud connected by design because the company's value pitch is on the data, not the hardware. They loose money or break even on the capital hardware. This is always the case, intuitive makes their money on the instruments, others make it on the subscription fee.
If you want it this way, you'll end up paying a lot more for thr capital equipment.
$400 a month, in a reasonably busy dermatologists office, would probably come to about $20 "per use". A lot less if you have multiple patients use it a day.
Primary use case would be PCPs who aren’t trained to diagnose skin cancer. There will be cost saving because the patient won’t have to visit a derm. Also it opens it up beyond dermatologists so it could be a PA or NP. That means in theory you could go to could go to a CVS minute clinic if you have suspicious mole.
> The FDA approval was based on results from a multicenter clinical trial led by the Mayo Clinic involving over 1,000 patients across 22 clinics.
Medical tech often requires clinical trials so there is some cost to that. And it can also limit how quickly competitors can create a competing device. But $399 monthly for unlimited use doesn't see so bad when a dermatology clinic might see hundreds of patients a month.
It's normal in the US. Outside the US things are more competitive as companies chase fewer buyers/dollars and have to sell things cheaper. Sometimes this makes certain countries worthless in the eyes of a medical device manufacturer
In theory this lowers the barrier to entry. A small clinic can buy in at the limited use subscription and they can pay to remove the limit as they grow. Having one hardware SKU for everyone and selling feature unlock licenses simplifies manufacturing.
Price discrimination, intellectual "property" enforcement, and DRM are nothing new in the medical industry.
A quick google shows iPhone apps detect skin cancers with up to 95% accuracy (as of 2020; may have improved a little since), I wonder what human level accuracy is (i.e. doctor carrying out a visual inspection), because surely it's not 100%? Looks like the accuracy of the FDA-cleared device is similar to the best apps (mid nineties).
It could be true that frequent inaccurate checks keeps more people alive than infrequent accurate checks. If so, an inaccurate device (or even iphone app) might be better than a doctor checking infrequently (e.g. every 2 years or so).
95% accuracy for a diagnostic test is not that impressive.
Wikipedia says one in 5 Americans will have a skin cancer in their lives. If each American has 10 moles, and we assume at most 1 is cancerous, that means a test which says "not cancerous" for all moles will have ~98% accuracy.
And 10 moles is an underestimation, not to mention that the same mole can be benign at first and develop cancer later, so we should really be looking over the number of moles * checkups, pushing the raw "accuracy" even higher.
It doesn't need to be 'xor' pairing between automated detection or a doctor. It can be both.
There's interesting work happening under places like ACEMID [0] on building models that boost detection accuracy. It's a research project so I have no idea what the final process will look like, but dropped in for a scan the other day, was imaged by a rather neat 360 camera rig [1], then had points of interest flagged for review by a dermatologist.
I don't think it's a stretch to envision a process that lets clinicians use more accessible/lower cost imaging devices to help focus their checks, particularly when you consider the ability to compare imaging across scans.
I don’t have the data for a cost-benefit-analysis but frequent inaccurate checks cause harm as well. The false positive rate is 3%; if all 330+ million Americans tested themselves each year, that’s millions of benign cases that will need to be followed up on which inflate costs and create backlogs.
> The false positive rate is 3%; if all 330+ million Americans tested themselves each year, that’s millions of benign cases that will need to be followed up on which inflate costs and create backlogs.
True. I also don't have the cost-benefit. I'm in sunny australia and get skin checks every 2 years, it's a slight PITA tbh, requires a couple of hours off work to commute to the doctor and get it done (the check itself is unobtrusive and usually takes only 10 min, it's getting there and waiting that's time-consuming).
Suppose a check via app takes 5 minutes (as opposed to 2 hours), and could alert you to true-positives sooner (possibly 1 year 11 months sooner), that could have a very significant benefit since skin cancers can go from harmless to harmful quickly.
The other thing we're missing is the false-positive rate for the human check; I suspect it's non-zero; possibly not trivial either.
I think that "accuracy" isn't as easy to define as you'd like. There are keratoses which just fall off if left untreated, and these are not considered cancer, but the cells are indistinguishable from certain skin cancers under a microscope. Since it may not be obvious whether something will just fall off if left to itself, and "wait and see" isn't an ideal approach for dealing with something that might be cancer, these often get treated, leaving whether they were 'really' cancer unknowable to modern medicine.
The article says this product uses wavelengths to detect abnormalities, DermAssist seems to be using pictures only. The latter seems like more of a toy and less a real diagnostic tool.
ML techniques based on cameras alone are reportedly beating dermatologists at identifying malignant melanomas. One study from 2019 (already ancient history in this field): https://pubmed.ncbi.nlm.nih.gov/31401469/
We already have something like this for at least a year in Germany.
Even my Family Practice is equipped with one.
So where is the big news?
This is hardly a new product.
But they charge only 5$ per Scan not 199$ for 5 Scans lol.
This is ridiculous
It’s news because the regulatory approval is very different between the US and the EU.
Being the first of a kind in the US means you go through a more in depth process with FDA (DeNovo). It’s a big achievement for the manufacturer. After that other manufacturers have it much easier to launch a similar product.
Oh and also the FDA puts more emphasis on devices working for all Americans, so this company had to prove that it works on all skin colors. This might work better than in the German device.
