Agile Diagnosis (YC S11) Launches: Helping Doctors With Diagnosis | Hacker News Reader

Agile Diagnosis (YC S11) Launches: Helping Doctors With Diagnosis

techcrunch.com

73 pointsjonlee22 comments

Every time I go to the doctor with anything that isn't obvious I'm always amazed that some basic ML tools aren't part of the standard practice. Especially since the standard reaction now is "here take a bunch of expensive tests, you won't be paying for most of them out of pocket so who cares!" We've had the tools for decades now to at least say "given these symptoms, it's very likely you X, Y or possibly Z". Additionally I'd be surprised if there weren't some unexpected correlations between various features that most doctors would miss.

Of course I guess I shouldn't be too surprised since medical records keeping is still 30 years out of date. The fact that it's still common practice to share medical records by filling out a consent form and having one office fax paper forms to other is ridiculous.

I really hope we start to see disruption in this industry, so much potential for not only cost savings but improved care.

re: medical record sharing

I don't think it is that ridiculous really. Assuming that the policy to require a signature from the patient stands, fax seems like a pretty natural choice for sending a paper form. Alternatively, they could take the time to scan the form then email it.

Ideally, both you and I would probably prefer something completely digital (both forms and health records). Towards that goal, I think the biggest problem would be interoperability. Someone like Kaiser could (and maybe they do) use all digital things between themselves, because they can enforce a uniform software platform.

One nice thing that I've seen about some digital medical record systems is that they print out nice reports. That is a good fallback system, so that you can still fax or email simple documents around.

Note that in the US, protected health information (PHI) may not legally be transmitted across the internet unencrypted, according to HIPAA, which renders emailing records much more cumbersome than simply faxing them, which is cheap and ubiquitous.

Kaiser uses Epic, so they have a system called Intraconnect; now Care Everywhere to share information. This also works between other Epic customers with Care Everywhere installed. For the non-Epic customers there is something called Care Elsewhere.

But there are a LOT of smaller practices without an EMR or that don't want to go through the work of initial integration.

IMO, the problem is not so much about the tools, it is more about Doctors "perception" of such tools and to a certain extent the "culture" and their past medical training. There is tons of evidence/literature in form of clinical guidelines[1], RCTs (randomized controlled trial)[2], and information resources (such as UpToDate). There are several studies [3] on why doctors continue to ignore software alerts (that incorporate such knowledge) in EHRs at the point of care. Although, I do agree that across the board the UI/UX of the EHRs currently in the market is horrible.

I believe there is a fine line between a good "decision aid" and "annoyance" and getting this right for an expert working in a chaotic, collaborative, time-sensitive, law-suit-attracting environment is extremely challenging.

1. http://guidelines.gov/

2. http://en.wikipedia.org/wiki/Randomized_controlled_trial

3. http://www.informatics-review.com/wiki/index.php/Alert_fatig...

I found this answer on Quora, by a medical student, to be quite illuminating on why machine diagnosis isn't going to be any good, any time soon. http://www.quora.com/Why-is-machine-learning-not-more-widely...

This is an older post describing the subproblem of ontology/vocabularies:

http://lingpipe-blog.com/2008/11/17/accident-involving-space...

It took 3 days of work-up and 4 teams' brains to make the final diagnosis of early Crohn's disease.

3 days and 4 teams working on one guy's stomach ache? Was this guy a Saudi prince? Seriously, I think we've already established that the richest of the rich will always be able to get great health care; for myself, I just don't want to have to convince a doctor that a 103 degree fever and occlusion on a chest x-ray aren't caused by my asthma acting up. I don't want to go into the doctor after vomiting constantly for three days, just to be told "that happens". Personal care from a team (or a few teams) of doctors probably can't be beat, but machine diagnosis can probably beat any health care I've ever received.

Jae is awesome, but picked a particularly difficult case.

There are certain kinds of diagnosis which are hard for humans and easy for computers, relatively, and some which are hard for computers and easy for people. (well, more likely hard for one and impossible for the other).

Truly novel things, where you aggregate data across multiple sites, are IMO the most amazing. Or, really rare but well defined conditions; doctors, especially busy ones, have a much smaller in-memory working set than computers.

I am very excited by machine diagnosis to augment humans. I don't think it will replace humans for a long time, but making humans even 1% better saves many lives and improves quality of life (and saves money).

In order for ML to be effective you first have to provide the correct inputs as discrete, coded data that the software can accurately parse. We have a number of different coding systems for observations of patient symptoms such as SNOMED-CT. But you would still need an experienced healthcare provider to figure out what observations to make, and then turn those into specific codes. That's very much a subjective process in many cases, and there are often multiple valid ways to code the same observation.

All of that data entry would take time that physicians don't have. It would only be potentially worth doing for complex cases. Unless you can find a way to pay providers for doing the extra coding and data entry. Insurance generally won't pay extra for data entry beyond what they need for billing, and billing data isn't sufficiently detailed for diagnosis.

ML also doesn't account for human intuition and tacit knowledge. There are many subtle cues which experienced physicians can pick up on to make a diagnosis. It's impossible to codify them all; some are probably only noticed subconsciously.

This industry is not very open to change. But, Borna and Eric are very hard working (and equally smart) guys -- just the kind of guys who can bring disruption to the industry. I have very high hopes!

I think it's great that there's an increase in attention to the medical space from smart young folks. From my experience though, being smart and hard working have very little to do with success in the medical-tech space. It mostly has to do with 1) if you can show ROI with the least (can't emphasize this enough) change to anyone's daily workflow, and 2) having a very very hard working sales team

Another critical aspect is to target the tech-savvy younger generation of students and residents, which Agile is doing with their recent offer of free lifetime access for participating medical schools. Almost all of us are using tools like Uptodate on a daily basis on the floors. Agile's algorithmic presentation of the same information makes a compelling case to use their product. Early adoption by trainees can lead to future paradigm shifts in the whole industry.

I have chronic nerve pain in my entire upper body. I used to be a patient of a world class university pain clinic. I left there for an individual pain doctor because... I saw 4 doctors in 4 visits, none of whom knew my diagnosis, read my file, etc.

If this flow chart was on the door of the patient room, my quality of care would have been MUCH better.

I find it fascinating that AgileDiagnosis went through YC, and then six months later through RockHealth. Would any of the founders mind commenting on how/why they chose this path?

Both incubators have been crucial in their own ways. YC has an immense experience and expertise in start-ups of every kind to draw from that helps us tackle general start-up challenges decisively. Rock Health has a deep network of health and medical partners and mentors for opportunities and challenges unique to health-tech start-ups.

In retrospect, would you have done them in that order, or done Rock Health first and YC second, or does it not matter?

I don't think anyone can touch YC for the core value proposition, but domain-specific programs (for medical, enterprise, government, ...) might make some sense, and I could see the value in something like StartX as a pre-YC while-still-in-college thing.

I hunted down a long HN comment thread where I proposed something similar about 2 years ago: http://news.ycombinator.com/item?id=1399953

Pretty excited to see someone is executing on this

A space that IBM is hoping to own with it's "Watson for Healthcare" initiative: http://www-03.ibm.com/press/us/en/pressrelease/36989.wss

Amazing work JonLee and crew! Congratulations on the launch!

<conspiracytheory> http://paulgraham.com/ambitious.html was written before YC S11. </conspiracytheory>

I'm sure a lot of people here have read Atul Gawande (of the New Yorker) article about The Checklist, in which a simple list of steps before surgery, such as "wash your hands", vastly reduced infections at hospitals. Yet the checklist faced resistance, including from doctors who thought that having such an obvious list was beneath them...even though even the smartest doctors mess these steps up in the rush of an operation. http://www.newyorker.com/reporting/2007/12/10/071210fa_fact_...

This is the mindset that any disruptive enterprise must face. It's not like in the coding community, where spending a substantial amount of time writing seemingly trivial and boring tests -- i.e. an automated checklist -- is considered to be a necessary step in writing software.

Reply on news.ycombinator.com