Challenges for Artificial Intelligence in Medicine

I'm curious, since you have a background in fraud detection, are there parallels or insights you brought to healthcare from that area? I'm currently working in fraud detection, and I'd like to move to healthcare.

Also, what are your thoughts on operations research for healthcare? That is, not modeling individual health of patients, but instead improving scheduling or other operational aspects of a hospital or clinic.

2. Tooling. Spark, Tensorflow, etc.

3. Policy. Folks who make decisions in hospitals are finally coming around to this whole "computer" thing. Slowly, to be sure. Eventually those who don't figure it out will be bought by those who do.

In some cases, this is a simple problem of embedded traditions, or worse, resistance to something that could put you out of a job. In a lot of ways, though, the resistance to change in medicine is seen as an important safeguard; when you're dealing with people's lives, you hesitate before making any changes in procedure, because we know our current methods work okay. Even if there are others which seem to work much better, we should proceed cautiously. How worried are medical professionals about adopting the more opaque techniques of AI? Can you persuade people to accept a diagnosis from something that can't explain its reasoning? Are you worried about any "bugs" or undiscovered unusual behavior in edge cases?

In addition to the lack of (labeled) data, the deployment and adoption challenges, and the fear around regulation, I would add another challenge: patient data is very complicated and highly heterogeneous: think doctor and nurse notes, machine-generated imaging, all kinds of measurements, patient habits, patient medical histories, etc.

I suspect the first major hospital-facing implementations of machine learning will be in radiology, e.g.: http://suzukilab.uchicago.edu/, which has been diagnosing cancerous tumors in CT scans with neural networks since before it was cool (one reason you won't see the terms 'deep learning' in the literature since they were originally just 3-layer networks, before the term was even coined). IIRC it outperformed the average radiologist.

I wonder if the label problem could be less difficult for some low-hanging fruit. The CT scan neural network required something like 40k labeled scans from a radiologist, but it could come for free: many yes/no disease detections will eventually be resolved by human labeling anyway by your doctor. If you had access, say, to every CT scan taken, and electronic health records for the patients, your labeling is noisy and biased but at least at massive scale. The problem is (legitimately) restricted access to health records in the US. Maybe some European countries have better data access?

And the implementation problem will eventually disappear. I remember talking with a radiologist years ago, who remarked "some people in my field have no idea it's about to disappear". I'm not so sure there will be no more radiologists, but their role will definitely change. Hospitals would be okay with this, actually, since radiologists are expensive. Eventually radiology scans will probably be like ordering blood tests, where fewer and fewer MD's are required.

However, I feel like you're making it much simpler than it really is. You've got a cadre of PhDs at UCSF, how am I going to compete with your all-star team with all of your wealth of knowledge to release something truly innovative?

They've hit the point of gaining enough training data and were moving onto the next phase where they wanted to utilize deep learning to help augment the doctors decision making. There's a ton of red tape here with the FDA but that was the near term goal. Augment the doctors decision making but not replace them.

I believe they'll succeed in doing this one way or another. We had talked about also pairing brain imaging data with genetic data (eg. Alzheimers and mutations on APOE). The critical things we talked about were how we actually train the different models, what a sequential approach towards this type of software development would look like, etc. We believed that the most pertinent focus point would need to be a refined supervised deep learning model.

I can definitely sympathize with the complexities of deep learning in healthcare.

From the abstract: "MYCIN received an acceptability rating of 65% by the evaluators; the corresponding ratings for acceptability of the regimen prescribed by the five faculty specialists ranged from 42.5% to 62.5%." So, better than the human experts considered individually.

Expert systems are able to explain their reasoning, which is essential if they are to be used for diagnosis. Neural networks cannot.

Deep learning and similar approaches might be useful in two areas though: interpretation of images, e.g. what's on this X-ray or ultrasound scan, or what type of rash is this; and undiscovered associations, e.g. are patients who were given drug X combined with drug Y for disease P more likely to get disease Q later on in life?

Deep learning in medicine also has a downside even supposing it works: lots of patient records are required, and anonymous ones can be linked to they people they describe, so there's a confidentiality problem.

PS I notice that you appear to be agreeing with my in your article: "This was the MYCIN project, and in spite of the excellent research results, it never made its way into clinical practice." and even refer to the same paper, so we're really just using different criteria for success/failure. One of the problems of expert systems was getting them adopted by end users. I don't see how using neural networks will be any different in that regard.

Maybe the lack of adoption of machine learning (I don't like to call something AI that really isn't) is due to the fact that medicine is more demanding on high quality scientific evidence than the IT industry.

Related, Gerd Antes from Cochrane once wrote a very interesting piece on the promises of "Big Data" (which is kinda related) and that they still need to hold up to scientific evidence: http://www.labtimes.org/editorial/e_654.lasso

> If so... do the challenges listed in this post resonate? Do you believe the shifts identified are the right ones to focus on?

I'm applying deep learning in a healthcare application and certainly the points you raise resonate and are generally the right areas to address for utilizing AI (for points 2 and 3 really any innovative technology) but let me expand on them a bit.

#2: Deployment and the Outside-In Principle - Very often new technologies are coupled with business model innovation to bring about change. The complex and often perverse mechanics of the healthcare system in this country make this exceedingly difficult. I agree that models that more rationally couple risk with reward (e.g. Accountable Care Organizations, Bundled Payments, and payer/provider organizations like Kaiser) provide the right incentives to reduce cost while increasing the quality of care. This is an environment where technology can make a difference; not so in the fee for service model. I think the software "deployment" model is much less of an issue.

#3: Regulation and Fear - This is actually a significant challenge. The FDA has significant incentive to be very conservative with their approvals (lives may hang in the balance) and for them the risk of failure (i.e. a death caused by a device/test they approved and likely makes the news headlines) is MUCH more traumatic and negative than the rewards of success (some costs are reduced for a specific treatment or diagnostic that almost no one will ever hear about). Additionally the FDA, Doctors, and medical administrators suffer the same fear of formulaic decisions that most people do. We simply don't trust an algorithm to make decisions and so will resist them or hold them to significantly higher standards than we hold human driven decisions. This means even if you get through the FDA you'll potentially have resistance of doctors and patients.

If we export some of the other posts about radiology I can just imagine a patient's experience in a hospital. I'm sure a doctor or radiologist will share an anecdote like the one in this thread (radiologist saw an occlusion due to a poor image and detected a tumor) and scare a patient into thinking a radiologist is better than an algorithm at reading an MRI/x-ray despite statistics that will (eventually) show that the deep learning algorithm will be more consistent and accurate. To be clear I'm not arguing that a radiologist not continue to be involved and review the diagnosis but that it will be difficult for this technology to be established in high-stakes applications due to fear.

To be honest, even if we were to stumble across a revolutionary algorithm with 99% sensitivity and specificity, I am sceptical about the dissemination and use of such a product in medicine for at least the next decade.

However in the big picture I agree with you. Now is the time to be building these things.

Today, workflow integration issues still remain, there is still a lot of free text entered etc. However a more pervasive issue is the lack of outcome data against which to train. In other words, what are we optimising algorithms for? In many health care systems we capture raw data, e.g. observations and labs, but not patient outcomes that are meaningful (i.e. based on optimising patient utility vs some more easily captured data).

The final issue for deep nets and ML is that these are descriptive models, they learn from experience, where as we know there is huge variation in practice and outcomes. In medicine we may want normative models based on best evidence, or some combination. And then there's integration with individual patient utilities.

Biomarkers predict outcome, of drug effect or of toxins/safety. When used preclinically (non human models) biomarkers continue to be highly valued internally. If a computational model can reliably anticipate outcome, it can shorten trial time and cost. However, in clinical/human use, biomarkers seem increasingly fraught in recent years. However the FDA seems to be less receptive to surrogate outcome predictors, and more demanding of concrete quantifiable clinical adverse events (e.g. the LDL/HDL ratio vs stroke, BMD vs bone breakage, A1C vs retinopathy). As such, I'd be circumspect about developing biomarkers that lead directly to diagnosis. Like the PSA test, even a strong biomarker that also introduces false positives or uncertainty is likely to face opposition to adoption in standard medical practice.

As to the use of AI (esp. pattern recognition) to better measure drug response or toxin effect, this seems to be well received, at least for in-house use. Automation of signal acquisition or analysis, if it can reliably improve on the status quo, in my experience, is well received by my employer. As a drug development cost cutting measure or as a more reliable rater of symptom measurement, AI seems to be win-win. That doesn't mean I see a wholesale rush to adopt AI-related tech here, but the interest seems to be steady and positive. Presumably this should lead to greater use of AI by manufacturers of medical instruments, which frankly I have not seen (though Siemens certainly has hired its share of quants, presumably to serve such ends. However, these folks may well spend most of their time working their magic on external contracts.)

Often in pharma, I think AI, like math models, are perceived by biologists to be too synthetic and abstract, and lack the credibility of a well trod mouse model. Unless AI/quant models lead to a < .05 T Test and a visible separation of error bars between groups, it's unconvincing statistically. And unless the AI can be tied _convincingly_ and directly to an underlying chemical mechanism, it's unconvincing biologically. Scientists are a tough audience.

Take ECGs -- it's true that in a hospital, an automated ECG interpretation doesn't buy you much. But what about about the patient with a paroxysmal heart rhythm that doesn't show up when they're at the doctor's office?

I was at a patient conference recently, and people were describing the first time they felt atrial fibrillation (a common abnormal heart rhythm). Many times, by the time they got to the doctor, they were back in sinus rhythm and thus the ECG showed no abnormality. Some were told they were just feeling "anxious" or "going through menopause." It often took months of persistence just to get a diagnosis.

Now, if have cheap sensors + AI analyzing the patient's whole heart history before they walk in the door, you can do a lot of good for real people.

Google should write EMR software. They would probably be pretty good at it.

We have some pretty active users with great results, but doctors are super busy. Its hard to get them to use anything that isn't in their standard tool kit or tie to payments. And, that understandable when you see 14 + patients a day. Getting into the workflow if the real challenge for AI in my view.