Will technology replace physicians?

Khosla writes well in his techcrunch article Do We Need Doctors Or Algorithms; talking about how technology and algorithms that will replace 80 percent of physicians.

It’s hard to not to get excited with the vision and examples that Khosla brings in the article but it’s easy to peel off the thin veneer of a VC hyping his portfolio companies once you stop panting from the excitement.

The vital signs could all be determined with the help of mobile devices, the operation of which do not require years of training and a certification. You will be able to do this by yourself—Philips already is using the iPhone camera to try to measure vital indicators, others will be even more innovative and as an insurance company it would be cost-effective to give them to every insured person for free. Skin Scan is measuring your risk of skin cancer from a photograph of a skin lesion. Telemedicine is accelerating and a Qualcomm company is measuring heart rates using an iPhone. Cell phones that display your vital signs and take ultrasound images of your heart or abdomen are in the offing as well as genetic scans of malignant cells that match your cancer to the most effective treatment. Ear infection and skin rash pictures and more will all be mobile phone based, often supplemented by the kind of (fractal) analysis that Skin Scan does, and more than what the doctors naked eye could usually see.The history of symptoms, illnesses, and test results could be accessed, processed, and assessed by a computer to see any correlation or trends with the patient’s past. You are the one providing the doctor with the symptoms anyway after all!Any follow-up hunts for clues could again be done with mobile devices. Etc…I have no doubt that better expert systems, high-power computing in smaller format devices, better graphics at lower costs, better understanding of what parameters count with better man-machine interfaces and fast access to expert answers could help replace 80% of physicians and could help more people be healthier and at a lower cost than the current healthcare system.

But they won’t. Not soon anyhow. There are a few bugs in Khoslas vision:

1. A VC doesn’t have to deliver outstanding value to patients. Venture capital business is about delivering outstanding value to investors by selling the startup to a big organization like GE Healthcare, often as an off-balance sheet R&D asset.An exit  event might take 10 years and that requires a lot of hype to keep funding flowing and deals coming.  An exit event requires “market validation” and “traction“, which translated into English, means that the technology works well for the portfolio company’s customers and that they’re able to collect real money for their products and services.Truth be told – there is an insanely cheap way to solve Americans’ health issues and that is for people to walk instead of driving and stop eating fried snickers for breakfast – but you can’t make money on people doing the right thing on their own.
2. Over-optimistic time estimations. It won’t be 10 years. It might be 100 years. My anecdotal evidence for this is based on an expert system we worked on almost 40 years ago for diagnosing and treating female infertility. The computing technology has approved by over 4 million times (consider Moore’s Law) but the problem has still not been solved to the point where we can do without 80% of the physicians in this space.It’s difficult to understand why a healthcare AI project you think should take 3 years actually takes 100 years. Beyond the usual software project estimation error, problems of software implementation and algorithm development there are many factors that we probably don’t know yet or won’t be able to solve with technology – like incompatibility between men and women for making children.
3. Confusing technology performance with the quality of algorithms and deep understanding of what parameters that we measure in order to analyze, diagnose and treat a particular clinical issue are really important.Yes –  a mechanical algorithm with a well-tested and proven record is better than doctors relying on hind-sight and intuition in many many cases.The challenge is arriving at the algorithm.It may take another 100 years of basic science research before the female infertility problem is cracked and then we will be in the realm of software implementation risk and project overruns where things still take 2-4x as long as we thought.Yes – brute force computing is a good substitute for inefficient algorithms but being able to run a lot of fast iterations with a crappy algorithm doesn’t help a woman trying to get pregnant as she tires of needles and measurements after 5 years or more of trying.
4. Big healthcare data is not retail data. Khosla makes a big point about the promise of big data for healthcare.What is the basis for this confidence in big data for healthcare?The US retail and financial services industries have been key influencers on the big data analytics space. Big thinking reasoning is that if it works for  commercial transactions, then big healthcare data  analysis should yield value as well.This is where I beg to differ.Big healthcare data sets are not the same as big commercial transaction data sets.

   Meaningful machine analysis of big healthcare data  could be possible but is probably currently  unfeasible except in well-defined and well-controlled use cases. The reason for this is that unlike transactions that are structured and have well-defined data values, big healthcare data is unstructured, poorly codified and of low evidence value.

   1. Current EHR systems store large volumes of healthcare data about diseases and symptoms in unstructured text, codified using systems like SNOMED-CT.Even if the big healthcare data sets were  perfectly codified, it is impossible to achieve meaningful machine diagnosis of medical interview data that was uncertain to begin with and not collected nor validated using evidence-based methods.
   2. Big healthcare data is retrospective data, which has low evidence value. Use of retrospective data, based on physician “experience”, may lead analysts to draw the wrong conclusions about cause and effect.   This is why, retrospective data analysis has given way to prospective, controlled data and  evidence based medicine. The US Preventive Services Task Force regards retrospective evidence as Level III, i.e. the worst possible. UK National Health Service regards this evidence as Level C or D on a scale of A through D.

One of the commenters (tcoyote) on the Khosla article said it well:

Big thinkers like Khosla have had an embarrassing track record investing in healthcare over the past twenty years. The Silicon Valley geninuses have squandered hundreds of millions of dollars of venture money in “molecule of the month” biotech companies, brought us “recreational genomics” from 23andme, etc., and $3000 a test single gene tests with debatable connections to future health risk. It’s been fun to watch. Glad not to be one of their limiteds.The healthcare landscape sure looks different from 40 thousand feet than it does from the ground. I think Khosla’s thesis is abject bullshit. It’s not just diagnosis physicians do, but manage evolving and complex situations in real time. Knowing the patient and knowing how patients think and act is really important in being a good physician. We’re a hundred years away from knowing enough about human disease to do what he suggests. Sounds great as a TED talk though.

It could happen and the technologist in me is sure it will, just not in the lifetime of Khoslas portfolio companies.

forward to a friend