Deep learning on electronic medical records is doomed to fail

This title is somewhat click bait though, because the fault is really with EMR systems and (esp) the American Healthcare system, not deep learning.

The entire system is designed around billing and decisions are made my hospital and insurance executives that are generally not technical. There is no incentive to clean up the system or work on a well structured open protocol for interop the same way there is in say banking. Plus, the author gives some good examples like pulse ox%, doctors and nurses are not at all concerned with or trained to record data in a way that makes sense to use programmatically. They're typically thinking only as if they're recording it for another human to read.

Deep learning could probably be quite useful in the medical field, but we won't know until someone comes along and disrupts the system top to bottom similar to how Tesla has done with not only manufacturing, but the sales process and shirking the dealership model. This would probably look something like Forward[1], but with a crazy amount of funding, so that insurance companies and billing codes could be ignored entirely.

[1]: https://goforward.com/p/home

This would be like asking programmers to standardize on the recommended programming language.

we would love to just use the recommended workflow, if it worked for our hospital. There are differences in the patients, doctors, local regulations, existing systems, etc between hospitals.

Patients: Top cancer hospital does a lot of clinical trials, so some of the forms require you to fill out clinical trial information for every patient. In a maternity ward, it would not be appropriate to ask about clinical trials for every patient.

Doctors: Hospitals are staffed differently. If the hospital has residents, some of the work can be delegated to residents. If not, someone else has to do it. The workflow needs to account for who is actually available to do the work.

Local regulations: Medicine is highly regulated, and each state and hospital has its own rules.

Existing systems: Hospital computer systems have been around for decades, and usually it's not possible to migrate everything to a new system, so the new system needs to integrate with the old systems that couldn't be upgraded.

- at some point someone, without evidence, speculated that cervix dilation should proceed along some curve

- cervix dilation is actually measured by hand - literally inserting fingers and having the doctor practice "so many fingers = so many centimeters". There's plastic sheets with holes in them so they can practice measuring the size of holes with fingers.

- the OBGYN knows that the cervix dilation curve should look like, and kinda sorta maps their hand readings to what it should look like

- the OBGYN has a general sense as to how labor is going, and will game the cervix dilation stats to match their expectation, e.g. if labor is going well but the cervix hasn't dilated then they'll kinda sorta report progress anyway

Anyway, given the above it seems like the data around cervix dilation is suspect - the measurements are fitted to what the curve should look like, and then the data matches the curve, and that makes people more confident in the curve, and so on.

The point is, can you really apply ML to the EMR of cervix dilation? Does it make sense, could you really draw conclusions from this?

You just described every admin software for every industry I've worked on.

The problem is individual orgs dictating software architecture. When each purchase is in the millions, you accommodate every whim no matter how absurd... and then you end up with these bloated, messy systems.

For software systems to REALLY, shockingly improve efficiency in an organization, all the processes in the org need to change to accommodate a new overarching system design. Tailoring software to mirror legacy processes defeats the purpose almost entirely.

I think there is a truly absurd competitive advantage in doing this right but you seldom see enough leverage to completely overhaul every department in order to implement software admin systems.

I know this because I worked as an ML engineer at an extremely successful company that automated medical coding using deep learning.

The confusion stems from conflating a "perfect solution" with a "human augmented" one.

90% of coding cases are trivial, have low value and can be done by a model. 10% are really subtle and need human expertise.

That's fine. You can make a billion dollar company on low hanging fruit. I think it's best not to conflate the perfect solution with a very good solution.

However- this has been truly shocking to all of the non-technical stakeholders I’ve worked with. They take the stance that any large amount of data can be used to do ML on, presumably because they don’t know too much about what doing ML is like.

So I’m convinced the author is right, and I’m also convinced that there will be many attempts to use ML on EMRs.

Many years ago, prior to anything like ML, Canada figured out that cystic fibrosis patients whose weight is higher than 50th percentile, had significantly better lung function. Nobody really understood why, but the correlation was so strong (.85 or something like that) it could not be ignored. Treatment protocols for CF changed to encourage weight gain, and lifespan outcomes have steadily improved over the years.

What other oddball correlations are hiding in the depths of bloodwork, weight/height, etc. for patients? We've teased out all the easy ones, the ones that are left are combinations nobody thought to even measure.

Regarding the EHR debacle, I'm optimistic that something could be worked out as a standard and implemented across the board. Expensive? Sure, but it's an investment that pays off pretty quickly, I think.

"I don’t think deep learning models on EMRs are going to be useful any time soon .. Clinical expertise is absolutely necessary to ask the right questions, to set up the inputs to the model, and to sift through the findings. Following that, clinical research will be necessary to validate the discovery"

A health outcome is correlated with age? Ya don't say.... This is only a problem that literally every single economist and sociologist checks for _first_. Agents in causal graphs respond to their inputs? Shocker, that's only.... The definition of an agent.

There are a ton of applications out there where a fairly naive model built by someone who took a Pytorch Coursera course can provide decent improvements. Healthcare is not one of them, and nobody serious ever thought it was. Bringing modern learning tools into healthcare is going to require a lot of smart people who know what they're doing, working cautiously to introduce these improvements into an operational quagmire with high stakes.

But this article reads like: "I tried making healthcare 'smart' in a Jupyter notebook over a weekend and it didn't work: the effort is doomed".

On billing: I'm convinced that the primary reason why healthcare (at least in the US) is so complex - is because of the dichotomy of saving people at all costs, while doing so fiscally responsibly. It is fairly common for large healthcare organizations to have ACTING doctors in their c-suite, who's primary goal is not to make money - it's to save lives. The people who care about saving money, reducing cost and increasing efficiency have no control over the organization. I'm not saying this is a bad thing, but IMO it's the largest contributing factor as to why healthcare billing is so complex, and healthcare costs get as high as they do (at least in the US).

France's system is universal and private (private docs, insurance, and reference pricing -- a service has a single fee across coherently regulated payers). They have had a standardized medical record for 20+ years. One system and one medical record for everyone. WHO ranked #1 healthcare in the world (to US ~40th) at 1/2 the cost per capita of US healthcare. This is catastrophic legislative failure for a problem largely solved by lots of other people around the world.

The legislative failure has created vast administrative overhead (10, or more, staff per doc at a hospital) and corrupt insurance companies. When an insurance company has to pay a claim, they call it a "medical loss" (they had the money and they lost it). They make their money on poor service and deceit -- charging wildly different prices for the same product where they can get away with it). In France, an insurance company, by law, has to pay a claim to a practice in a few days. Imagine the decreased capital needs for running a medical practice or a hospital.

The hospitals are not blameless in all this, but the heart of it is the payer system/s.

> Did you know that a blood oxygen saturation of 0% is highly correlated with healthy outcomes? No, I didn’t get the percentage backwards. A 0% reading is what you get when the nurse looks at you and decides you’re too obviously healthy to bother with putting the pulse oximeter on your finger. The empty field value gets saved as a 0, of course.

Well, statistical methods (no matter if classical, Bayesian, Deep Learning, or anything, as long as it goes beyond linear methods) will perfectly capture the special case of 0% and predict accordingly. These methods are free of our biases and will take consistent approaches (e.g. empty columns, columns that de facto mean something else, common typos, etc).

Sure, interpretability is problematic, and we often need to consider the knowledge of physicians and medical institutions rather than be based on raw data.

There has been modest progress through wider efforts:

- Standard vocabularies, eg LOINC codes for different kinds of lab tests

- Mappings between vocabularies, eg OMOP

- Semantically rich vocabularies, eg OBO's OBI

The important thing to remember about AI is you need reliable data to train it as well as reliable data to test it. If you don’t the FDA will not allow you to employ it medically.

The author's attempt at using structured EMR data is the root cause. We have found that structured data, which the author attempted to use, is at best 35% accurate. Sure it's better than claims, but it does not reach the level of quality necessary to inform clinical decision-making. The reason for this is that almost everything clinically relevant is captured in freeform text fields--clinical notes. To build proper models from information in EMRs, you have to start with processing the narrative data, which is a hard problem.

Training models to interpret clinical notes requires clinical expertise. Clinicians record facts differently in different locations, and there are many different ways to say the same things, and sometimes they skip underlying facts because some other fact implies the rest. Different specialties record things differently too. You really cannot just throw some data into a notebook and hope it works. Even with clinician input, we still find that high quality results require ensemble models with multiple techniques; plain NLP doesn't work either.

Take for example, non-alcoholic steatohepatitis (NASH), the leading cause of liver failure requiring liver transplant. NASH is a complication of non-alcoholic fatty liver disease, in which your liver has unusually large deposits of fat. NAFLD is not coded in structured data. To identify it from unstructured data, you have to extract concepts related to liver cancer, pre-diabetes, alcohol use, liver fibrosis, cirrhosis, jaundice, fatigue, and loss of appetite. To make a long story short, you cannot do these things using structured data or naive NLP approaches. F1 is zero.

So maybe his point, "Data encodes clinical expertise" is worthwhile, but the rest of the article...not so much.

Source: My company, Verantos https://verantos.com , specializes in the generation of high-validity evidence from data we abstract from EHRs using machine techniques.

Pap smears are also a really good sex bucketing signal, and there are lots of diseases that are more prevalent in one sex, so I would expect Pap smears to be correlated and anti-correlated with a lot of other diseases as well.

I think we have a term for this in both areas: garbage in, garbage out.

All the problems the author brings up, while legitimate, are being worked on.

For example, on the interoperability front, TEFCA is making big strides on government-mandated nationwide interoperability: https://www.healthit.gov/topic/interoperability/trusted-exch...

> In January 2022, ONC and the RCE announced the publication of the Trusted Exchange Framework and the Common Agreement (TEFCA). Entities will soon be able to apply and be designated as Qualified Health Information Networks.

Google also has made significant strides on deep learning on EHRs: https://ai.googleblog.com/2018/05/deep-learning-for-electron...

Which is to say, it is and will be difficult to create a business based on applying a novel DL method to a particular problem space. We're seeing a consistent trend that focusing on the other aspects of the problem such as data, tooling, or end to end services tends to be much more successful.

The continual war between hospitals having to opportunistically charge for their services vs. the insurance industry having to take a default stance of deflection creates the massive, meaty layer of coding and billing waste. Thousands upon thousands of jobs exist just for this purpose, and I think any inefficiency in a single-payer system is more than offset by getting rid of that layer and everyone benefits.

The cadence, uniformity and alignment of event stores to underlying pathways does pose an opportunity (but there's so little research and I suspect Brian didn't have access to this space). An EMR is a projection/point in time snapsot of these events (basically API calls). There's also inherent natural labels because inferences are evaluated against an actual pathway.

Specifically speaking about intra-episodic inferences -- longitudinal predictions (over several episodes -- or a patient life time) becomes wildly inaccurate. There's a lot of research to demonstrate this no matter which models are used.

So much of what government does (aside from the bombings etc.) is really about providing and enforcing a framework for people to work together. And in this era that needs to be a high tech framework.

Actually, it needs to not only be very high tech, but also very cutting edge, decentralized, sufficiently holistic but also flexible enough to evolve.

Which is incredibly hard, and we probably will not get due to greed, stupidity and politics, and that may be the actual reason that human civilization is superceded by AI civilization.

This goes hand in hand with Konerding's 3rd empirical observation, which is that there is a never ending supply of "machine learning geniuses" who are naive about how health care works in the US, and spend some 20 years learning how hard it is to actually do anything actionable with health care data.

Medical record systems, as other comments point out, have been constructed for the benefit of administrators, recording clinical data is mainly to support administrative needs.

In reality manifestations of a given illness vary continuously across a wide spectrum. It means patients with the same diagnosis have differing sets of symptoms and course of illness. As I like to say it, "no two patients have exactly the same disease."

Official disease classification schemes embody the "splitter" model which attempts to fit continuous data into discrete categories. Of course sharp-edged distinctions suit purposes like billing and other management operations.

However diagnosis is often ambiguous, mixed or multiple, but no matter what categorical diagnoses must be assigned. Unsurprisingly doctors are biased to select the choices with the greatest reimbursement, and "stretching" criteria to cover the patient's condition (or vice versa) is not at all uncommon. (Also, there's not assigning diagnoses where it might negatively impact payment.)

EHR clinical data follows the discrete assumptions of the system design. This can create an impedance mismatch between clinician observations and data input. This may be troublesome, for example, in specialties (behavioral health) where data is complex with many overlapping subtle but meaningful variations.

I've often received records after a patient has been hospitalized. More than not it's difficult to understand the course of treatment as there's no coherent summary or narrative description provided. The "pile of data" literally transcribed from the EHR isn't very useful to human readers. To be sure information like lab reports, etc., are good to have, but the marginalized human-to-human element is troublesome.

EHR clinical data failing to serve ML purposes could only point to problematic EHR system design. Though I've been aware of EHR limitations, I wouldn't have guessed about ML issues. The article taught me something about the problem that exists. Now it remains to be seen what can or will be done about it.

Which makes it hard to use models in clinics.

Doctors are smart about this. Every providier we spoke to about survival models based on biomarker data did not want supervised learning models. The explainability, trust, authorization just aren't there and the risk of misuse is too high.

For now, research is a better use case, but less commercial funding.

It is not hopeless, but your basic LSTM is probably not going to revolutionize medicine.

I work mostly on claims data now, and building models to identify problematic claims I assure folks is feasible.

The point I am raising is that there is an embedded assumption that this info should even be available for deep learning to work over.

    1.billing
   2. government mandates
   3. billing
    4.helping doctors keep track of their patients’ records

If not, I wouldn't call it "deep".

A lot of countries don't have the same drivers.

Perhaps things are better in some of the many other health care systems around the world.

ECG is the classic case of data fragmentation that he's talking about and if you want to export raw data from the popular ECG machine like Mortara, good luck with that. Heck even Apple currently does not support exporting raw data ECG out of the box that's crucial for deep learning, apart from the PDF image file for the ECG waveform [1]. Literally there are more than a few dozens open formats for the ECG [2],[3] and this obligatory XKCD comic comes to mind except that there are 39 standards instead of 15! [4]. Even if the ECG machine manufacturer is using one of the formats there are still serious interoperability issues down the road.

Rambling aside, there's yearly cardiology global challenge competition organized by Computing and Cardiology Conference (CINC) and recently there are machine learning and deep learning techniques proposed for multi-lead ECG diagnostics but the results are not that great [5]. Hopefully it will start an impetus to deep learning in EMR similar to ImageNet Challenge that gave rise to the actual deep learning algorithm that drived the community past the winter AI.

[1] Accessing the ECG Data of the Apple Watch and Accomplishing Interoperability Through FHIR:

https://pubmed.ncbi.nlm.nih.gov/34042901/

[2] A review of ECG storage formats:

https://pubmed.ncbi.nlm.nih.gov/21775198/

[3] A Review on Digital ECG Formats and the Relationships Between Them:

http://diec.unizar.es/~imr/personal/docs/paper12IEEETITB1.pd...

[4] How standards proliferate:

https://xkcd.com/927/

[5]The PhysioNet/CinC Challenge:

https://cinc.org/physionet-cinc-challenge-awards/