Before neural networks got deep, there was a lot of very impressive work applying neural networks to medicine. Example:

http://suzukilab.uchicago.edu/research.htm

IIRC they were outperforming the average radiologist on some tasks 10 years ago.

Why haven't they replaced the average radiologist yet then?

Having worked in the field, the model that works well is decision support = augmenting an expert with a computer-aided diagnosis. This can work in a couple of ways, but the two classic ones are:

* Following behind the expert to give a second opinion

* Going ahead of experts to screen cases that should be read by an expert

Humans are imperfect and have their sources of unwanted variability. Algorithms may be less variable and reach near-human performance in controlled settings, but are often not flexible enough, not good at incorporating multi-modal patient history, and sometimes fail in spectacularly bad ways.

Social skills. You still need them, patients still demand them. I am a physician myself and can assure you I think my profession will go the way of the dodo in the next 50 years, starting with surgeons and going on to conservative diagnostics/therapy. The only reason it is not happening immediately is patients still (and for the foreseeable future) demand and expect the physician to be a human as well: showing empathy and bonding the so called "therapeutic alliance" with them.

As soon as AI manages to show that capabilities, it's the end. Algorithms are much more accurate in combining all the available biomedical information needed to decide for diagnosis and therapy.

but doesn't radiology specifically have the reputation for being the "unsung heroes" or at least the unsung among physicians ? with the exception of interventional radiology I thought they hardly ever interacted with pts ?

Indeed. My comment was more general in nature. Specialisations of medicine that have less direct interaction will be the first to be automated. Laboratory ones are pretty much already.

I don't think radiologists are going anywhere, but these tools should at least be used by radiologists. I think one reason they haven't been adopted is for reasons of safety and liability. It would be negligent to run these kind of tools without confirming the results, and if they have to inspect the results anyway, why bother with the computational approach?

I suspect one reason is the schleppiness of implementation. Healthcare is hard enough to navigate as a doctor or patient.

Another is the existence of edge cases. While not quite as bad as self-driving cars, they still require human review.

One radiologist I spoke with was under the impression that his field was going to disappear in the next n years. I don't think that will be the case entirely, but it will change.

One reason why image detection is ripe for this kind of "disruption" is that some of the highest paid medical fields (dermatology, radiology) basically employ expert image recognizers.

On the other hand, I'm not sure what percentage of health care costs goes toward employing doctors directly, but I don't think it's huge compared with the rest of the facility and equipment and administration cost.

The real benefit will come when people don't have to go in to the doctor at all. Which makes the smartphone "take a picture" aspect pretty sweet.

Healthcare plans do not reimburse machines yet.

I'm not sure if you're joking or not, but as someone who works for a medical technology company this is very true. Whether or not the use of a particular technology is billable to first order determines whether the technology is adopted.

There may be incentive soon enough, via ACOs and bundled payment programs. When dollars saved go to the bottom line, folks start trying to save dollars.

If we could have a preliminary diagnosis quickly from a machine and only then and if necessary talk to a specialist doctor, the savings in costs could be outstanding. Talk about Affordable Care. Maybe an act of Congress could move this along.

It all goes well until the machines terminate Ms. Butler's pregnancy.

I'm an average radiologist, and would be happy to answer any questions you might have.

Indeed, I second that the time is now. There are several imaging modalities/organs that provide diagnostic information about a variety of human functions or diseases. For example, retina is a unique organ that is amenable to imaging of central nervous system, cardio-vascular system, and microvasculature without any incision. This allows us to detect, screen, monitor, and predict risk for diseases such as diabetic retinopathy, macular degeneration, glaucoma, and even cardio-vascular risk, Alzheimer's, and stroke.

If this interests you and you have developed the expertise, there is another startup opportunity to explore -- please email me at [email protected]. We are bunch of machine learning PhDs, developing/publishing and commercializing deep learning algorithms for disease/risk identification from retinal photography. We play with millions of retinal images, and it's a lot of fun!

> If you've developed expertise in deep learning and want to apply your skills to healthcare in a startup

What if we haven't but we do?

If you're a strong software engineer but without any particular expertise in deep learning, feel free to email me as well. [email protected] :)

We (Cardiogram) in particular are hiring for: mobile engineers (building an Apple Watch app used by 100,000 active users), storage and data infrastructure engineering (how are you going to store 10 billion sensor measurements those users are producing?), and machine learning engineers (now that we have the data, what can we do with it?).

In practice, many people fit multiple categories. There's a tight relationship between data infrastructure and machine learning, for instance, since new infrastructure often enables new algorithms. Likewise, building a new feature in Cardiogram for Apple Watch may give us 10x more labeled data, and therefore make the deep learning algorithm perform better. Interdisciplinary is good.

Is remote working an option?

Not for now unfortunately--but maybe once we get a little larger.

I heard before that a problem with heart monitoring for ML is that most of the samples out there are with abnormal hearts rather than normal and false alarms. As in, there's not enough published data to establish a baseline for high accuracy across general population. Most claims like this cite rules protecting medical records/data. Never got to ask a specialist in the field for confirmation or rejection of that claim.

So, how true or false is it?

True -- the available data sets (like MIT-BIH arrhythmia database) are quite small. In our case, we launched a study with UCSF Cardiology which has recruited about 10,000 people so far, and that's how we get a data set large enough for deep learning.

Very interesting! What is Cardiogram's business model?

The app is free to you, but if we can detect (say) diabetes early and route you to the right medical care, we save the healthcare system money. We've built a platform called Cardiogram Connect which optionally lets you share aggregated data in exchange (usually) for financial rewards. You'd choose to connect with whoever pays for your healthcare--depending on your situation, that could be a health system, accountable care organization, insurer, or employer--and that entity is the customer that pays Cardiogram.

Is it working? It seems a bit convoluted. Why not just let users agree to sell their aggregated, deidentified data to whoever wants it and then you take a cut kinda like the app stores.

I think the parent was asking 'who pays you money?' not, 'for whom do you provide value?'