1,743 karma · joined December 29, 2010
You can certainly use high-level ML tools (like Keras), but it takes a great deal of work to wrangle your data into a usable format, and even more knowledge to debug an ineffective network.
I'm biased, but I don't agree with the premise of the article; Doctors should be among the highest earners in any country, it's the price you pay to ensure the brightest students are attracted to medicine despite the long training. I do agree with the following:
A second route would be to end the requirement that foreign doctors complete a U.S. residency program in order to practice medicine in the United States. This means setting up arrangements through which qualified foreign doctors could be licensed to practice in the United States after completing an equivalent residency program in another country. The admission of many more doctors would put downward pressure on the pay of doctors in the United States, as insurers would have a new pool of physicians to add to their networks who will accept somewhat lower compensation.
I've met many foreign-graduates who have already completed residency. In makes ZERO sense to have these individuals repeat a residency in the U.S. They have already undergone 12+ years of training, actively treat patients in their home countries, and there's a need for physicians in the U.S. This is very much cartel-esque of the AMA. I know it would result in lower salaries, but this would (hopefully) translate into lower-cost medical school in the U.S.
This is a situation that may benefit from tech solution. There are already certain states allowing foreign medical graduates to practice in the U.S. If you're looking for a "startup pain point", let me present you with the two sided market:
1. There were ~7,500 international medical graduates who applied to residency in the U.S. (http://www.nrmp.org/wp-content/uploads/2016/09/Charting-Outc...) Presumably, these are individuals who want to practice in the U.S., but must undergo a U.S. residency. I don't know how many of them have already completed a residency in their home country, but let's conservatively say 25% or ~1,800.
2. The AAMC is estimating a physician shortage of 30,000 by 2030 (https://news.aamc.org/medical-education/article/new-aamc-res...).
Even if all of the foreign train medical graduates were allowed to practice in the U.S. w/o residency, it still wouldn't be enough. Maybe there's a startup opportunity that matches foreign MDs to positions in the U.S. As I mentioned, states like Missouri have already allowed some non-residency-trained physicians to practice (https://www.statnews.com/2017/05/15/missouri-doctor-dearth/). My preference would be to have a foreign doctor who has undergone residency in his or her home country rather than a U.S. grad with no residency.
Also, note that SF isn't the fastest growing housing market anymore, it's Seattle (https://www.seattletimes.com/business/real-estate/seattle-no...). Seattle didn't feel as much housing-pressure with Microsoft due to the remote location of its campus in Redmond/Bellevue. That's changing with Amazon; the sales giant is located in prime real estate, just like the tech companies on the peninsula.
Cronote made next to no money for 5 years, but slowly migrated its way to the top of Google for search of "text message reminders" and "sms reminders." I did a ton of work, made an iPhone app, changed the billing system four times and rebuilt the interface three times.
I now have ~60 paying customers, but it only makes $300 per month after all expenses. I don't think I've recouped the man-hours spent, but I've certainly learned a lot along the way. I learned Python/Django (website back-end), Objective-C and Swift (iOS app), and lots of fun stuff about timezones and calendaring APIs.
FWIW, it took 10x as long to get the DLL call to work as it did to drag-and-drop the coefficient of variation box into the GUI. There are certainly rapid prototyping benefits to be had with higher-level languages. Nowadays I prototype in Python and rewrite line-by-line in C++ if I really need the speed. I think GUI programming languages might serve a similar purpose.
The solution would be some combination of video or audio that records the clinical encounter and automatically generates a note based on what was discussed and performed. It falls under Paul Graham's "Schlep Task" (http://www.paulgraham.com/schlep.html); you'd have to work with individual clinicians, get their (and the patient's!) approval to record the encounter, record it from multiple angles, build tech into dumb devices (e.g. stethoscope, Dopplers for pulses, O2 saturations, and somehow use machine learning to integrate all that sensor data into a plain-text note.
It's probably the #1 problem from a provider's viewpoint right now, especially on the primary care side. Any individual provider will see 10-20 patients per day and write just as many notes.
Edit: I'd add that you'd also have to have access to existing notes because progress notes typically summarize the patient's prior visits.
The difference is that Autodesk relies on a mature, deterministic technology (3D graphics rendering). Deep learning is a stochastic process that depends on the data and the model. The training code, and especially the framework hooks, is the least important part. The example they give is three lines of code to train a cat vs. dog classifier. I've tried this classifier on a different binary image classification task: livers with and without tumors. It didn't work very well. There's lots of reasons: little variability between images, grey-scale images, different resolutions, etc. You can tweak the network, throw in more middle layers, try different kinds of layers, whatever, to get better results. All of that is guesswork if you don't understand what the CNN is doing at each stage. At this point in time you do need a formal education in linear algebra, calculus and statistics to investigate why a model does/does not work. It's not enough to know how to use the libraries.
On the flipside, you also need to know how to manipulate data and parse it into the correct format. This generally requires a year or two of programming practice in a good scripting language like Python. I will echo their thoughts that Ian Goodfellow's Deep Learning Book is remarkably lacking in this area. As a simple example, you cannot even use AlexNet without pre-processing your images to be 227x227 or 224x224 for GoogleNet. That's 10,000 images resized, labeled and loaded into the model before training can take place.
tl;dr IMHO in terms of being a competent user of deep learning: mathematics >= programming >>> knowing how to use a framework
Edit: This is not a new idea, there are lots of companies already doing this using natural language processing.
https://www.zillow.com/homes/for_sale/East-Palo-Alto-CA/cond...
There are no 1 bedrooms from Santa Clara to Burlingame for < $350K, and only one two-bedroom for less than $500K.
Medicine-wise: I'm in the bottom quartile of my class. I'm terrible at sitting there and studying. It's a completely different mind set. I did really poorly on UMSLE Step 2 in particular; most med students do +20 points higher relative to Step 1...I did 40 points lower. That test is entirely memorization, we're talking memorizing immunization schedules for children, the screening guidelines for low-dose CT, the warning signs of an acute abdomen, and so on...
In terms of benefits, I'm much faster at research data analysis. I'm doing imaging work and I can take a DICOM, convert it to NIfTI, run an algorithm in Python or MATLAB, and dump the results in under and hour. That kind of stuff is second-nature to me, but takes my fellow lab mates an indeterminate amount of time. I've written a number of 20-line scripts for my classmates that do basic stuff like vertical lookups in Excel and searching for files on the filesystem. Really simple stuff that would take a person a long time by hand.
One more thing, the MCAT was a blast for me. It was a wonderful test. It emphasizes integration and application of what you've learned on novel problems, very much like computer science. The USMLEs were hell. I got 98th percentile on the MCAT; 40th percentile on the USMLE.
Surprisingly, teamwork is one area where CS wins out. I've found that despite the professed importance of the healthcare team, healthcare often doesn't function like a team at all. Doctors, nurses, PAs, patient care assistants, administrative staff, can make or break your life by providing support or creating obstacles. I think it's a result of a inherently tiered system with a single person (e.g. the attending) on top who decides the treatment plan. In CS you almost never have a single person who is overly-trained relative to peers. Maybe you'll have a better developer, UI designer or manager, but none of these attributes makes that person the sole decider. The pay structure in medicine reflects its hierarchical system: software managers don't earn 10x what software developer make, whereas attendings easily make 10x a nurse's salary.
"The sentiment neuron within our model can classify reviews as negative or positive, even though the model is trained only to predict the next character in the text."
If you look closely at the colorized paragraph in their paper/website, you can see that the major sentiment jumps (e.g. from green to light-green and from light-orangish to red) occur with period characters. Perhaps the insight is that periods delineate the boundary of sentiment. For example:
I like this movie. I liked this movie, but not that much. I initially hated the movie, but ended up loving it.
The period tells the model that the thought has ended.
My question for the team: How well does the model perform if you remove periods?
I read a lot of articles about AI in medicine, pretty much anything I can get my hands on. I also read generic tech articles related to everything from Nintendo Switch, Tesla, Brain-Computer interfaces, and other popular media articles.
-How many articles do you read each day? Likely 10+. These aren't high-brow articles, just random blog posts and pop culture tech. I read about 2-3 research abstracts per day in medicine and maybe skim the text of 1-2 articles.
-They're usually related to your job or to some side projects? Usually they are related to my interest in medicine or technology. Sometimes they are related to my job (I work as a part-time developer / data scientist). I also run a small website (https://www.cronote.com). I encountered a number of issues with time-zone switching and the daylight savings change on March 12th. Read about 20 articles having to do with correctly implementing timezones in Python.
-Do you usually read about a variety of topics or it's focused in 2 or 3 topics only? Topics cover a vast span of medicine and computer science. I enjoy computer science more than medicine so it's a 20:80 split.
-Do you usually read during some time of the day or it's usually random? I read whenever I'm behind my computer, usually alternating between work and browsing the Internet. This amounts to ~5 hours per day.
http://adilmoujahid.com/posts/2016/06/introduction-deep-lear...
It's the only way I could get a working model of Caffe while understanding the data preparation steps. I've already retrofitted it to classify tumors.