What do healthcare professionals think of Vinod Khosla’s prediction that automation and technology will replace 80% of physicians in the …

Answer by Jae Won Joh:

Let's not beat around the bush: it's complete bullshit.

Why are we even listening?

Khosla's words exemplify the type of incendiary arrogance that makes people in medicine hate to work with people in tech, particularly those of the hype-loving Silicon Valley variety.

Would you care about what a plumber has to say about quantum physics? No? Then why are you listening to Khosla babble on about a domain in which he has no training or experience? What, because he's rich, suddenly his words are golden and infallible?

The average 3rd year medical student has already forgotten more medicine than Khosla will ever learn in his entire lifetime, but Khosla seems to have no respect for this. Where did this (subconscious?) contempt for medicine come from?

I challenge you to name a single other field that a VC is perfectly comfortable dismissing 80% of people who spent 7-13+ years intensively training in a specific discipline before they're even allowed to sit for certification and become independent practitioners, particularly when a significant chunk of those years are spent working the only job in America that, due to regulations instated in 2003 after a young girl died, now has hours unwillingly capped at the equivalent of 2 full-time jobs.

What Khosla doesn't see

Read the way he describes a doctor's visit. It's blindingly obvious that he has no idea what goes on in his doctor's head.

I'll use an analogy from the coding world: it's because we in medicine "black box" our thinking. A doctor's thought process is much like an well-written function in code: you can call upon it without knowing (or needing to know) what's going on under the hood. All you have to do is give it input, and it gives you output.

But much like a well-written function, it can be orders of magnitude more complicated than you think. All you see is the serene smiling face, but underneath is a brain actively assessing your "cough" for pneumonia, bronchitis, asthma (exacerbation), COPD (exacerbation), upper respiratory infection, tumor, heart failure, bleed, trauma, to name but a few.

And all that analysis begins the moment I lay eyes on you.

I look at everything about you. The way you're dressed. Your posture. Your facial expressions. Your hair. The shape and proportions of your body. Your skin.

That "friendly banter"? To you, that's a social greeting. To me, that's more data. If you can speak, your cardiovascular system is intact enough to perfuse your brain and provide adequate oxygenation for thought, and your neurological system is intact enough to coordinate thought from your brain and translate to muscular movement/coordination, and your language centers are jiving to process my speech and formulate an appropriate response; it also tells me your airway and lungs are likely ok.

Your accent? The specific pattern of words you use? To you, that's just how you communicate. To me, that's more data. I get some idea of whether or not English is your first language, and if I need to raise my suspicion for tropical diseases or pathologies more common in foreign countries.

The way your eyes move? To you, that might mean nothing. To me, that's more data. If you're a woman and your eyes keep flitting back to your "boyfriend" as if for approval before you answer any of my questions, I'm going to ask him to leave and take steps to ensure you have the privacy/safety to tell me if you're being abused.

Literally everything about you is input to me. All you might see is me doing a "tongue and throat check" or "listening to the breath and vibrations in the abdomen", but there is way more going on in my head than you will ever know. So before you think my job is simple, it would behoove you to learn a little more about it.

Some unwarranted optimism

It's clear that Khosla has strong faith in AI/machine learning/etc. That's nice. I went to a talk last year at MIT about "Modeling and Prediction with ICU EHR Data". IIRC, Professor Marlin and his group had apparently spent the last 2 years on this research, and with a lot of fancy math on a dataset of vital signs from a pediatric ICU, had determined in an intensive care setting, the following were associated with mortality:

  • Low blood pressure
  • Prolonged cap refill
  • High heart rate
  • High respiratory rate
  • Low SaO2
  • Low pH
  • Low TGCS
  • Shock and depressed cognitive function

Everyone in the audience except me appeared to be thinking, "WHOAAAAA, WE CAN NOW TELL WHAT FACTORS ARE STRONGLY ASSOCIATED WITH MORTALITY??? BALLERRRRRRRRRR!!!"

My reaction as the only person with clinical training/experience in the audience: "…I could've told you all this as a first year med student."

It was kind of awkward, but in all seriousness: we've known about those factors for over a century. They're not new. What, you thought we called them "vital" signs for no reason? We call them that because when they're abnormal, things tend to be going south.

Suffice it to say I left the room shaking my head in bewilderment at the notion that "computers will completely revolutionize medicine".

More unwarranted optimism

Khosla also appears to have strong faith in the notion of EMR-based data analysis. Where this faith comes from, I have no idea–even a cursory search of peer-reviewed literature reveals that there are significant limitations.

For example, consider the common disease process known as sepsis. It's responsible for ~20% of ICU admissions and is also the leading cause of death in non-cardiac ICU's. Mortality for severe sepsis is usually cited as being between 30-50%, with an annual estimated 751,000 deaths nationally costing ~$16.7 billion/year in the U.S. [1-4]

Given such a common critical condition and wealth of data, you would think it would be a total cakewalk to just amass and analyze the data. Not so. Why?

When a patient sees a physician in our current system, that encounter's data is collectively billed under a set of ICD codes. The billing code used and the patient's actual condition are often not the same. For example, "food poisoning" may be billed as "Nausea & vomiting". A cold may be billed as "Congestion" + "Sore throat" + "Headache".

This causes problems when you try to go back and do retrospective analysis, because in order to find the patient population or disease process you want to study in your EMR, you have to basically play a sophisticated guessing game as to what billing codes to search for.

In other words, there's ample room for significant disparity in the results depending on what codes you use. Someone who searches for "flu-like symptoms" is going to get a rather different dataset than someone who searches for "fever"+"congestion"+"cough".

There's actually a recent paper from Sweden[5] that illustrates this perfectly: they took three independently published and peer-reviewed ICD criterion for identifying severe sepsis and applied it to a single starting dataset. Guess what? Each criterion produced wildly different patient numbers: 37,990; 27,655; and 12,512.

The differences between those numbers aren't exactly…negligible. So…where is this sophisticated EMR-based analysis going to come from? There's a limit to how much you can compensate for crappy data, so I really don't care how much data you have. What I want to know is, who's going to be the mastermind that determines how sub-groups are made for clinical recommendations? When you're trying to analyze a condition with a given dataset, is analysis on a 96% sensitive but only 53% specific subgroup better than a 71% sensitive but 89% specific subgroup, or should they be used for different analyses? Who will make that call to determine how conclusions are drawn for human lives?

"Everybody lies"

Fans of the show "House" will recognize the star character's favorite line. While a jaded view, it's not without reason. Khosla's view that "You are the one telling your doctor your symptoms" relies on the highly faulty (and naive) assumption that the patient is always telling the truth. That often isn't the case.

I would love to see Khosla's theoretical algorithm take on a manipulative benzo addict.

"I'm experiencing alcohol withdrawal, Dr. Algorithm."

"When was your last drink?"

"A week ago."

"Are you having tremors?"

"Yes, see, my hands are shaking right now."

"Have you withdrawn from alcohol before?"

"Yes, several times. Please, I need help, can you give me something?"

"Thank you for this information. You will be receiving valium shortly."

It would be absurdly easy to trick a computer into dispensing unnecessary drugs and interventions. Just read on the internet what symptoms you need to endorse and what signs you need to fake, and the computer would give you what you want.

I get flat-out lies all the time from people seeking secondary gain. Part of my job is to figure out who really needs my help vs. who is trying to score more vicodin they can sell on the street. Can your computer deal with blatantly false input? Can it be made to have the clinical acumen to know when someone is lying? Somehow, I don't think most patients will appreciate being hooked up to a fancy lie detector every time they see Dr. Algorithm.

The constraints of practicality

Your robot will never match me in speed of clinical assessment and intervention, and this is most obvious when patients truly need a doctor to live.

I had a patient not long ago who came in with "shortness of breath". The first thing I see is pink frothy foam coming out of their mouth, and I immediately know this patient will need to be intubated ASAP; the scary part is that once the meds are given to sedate and paralyze the patient I'll have about 20 seconds to do it or else I might lose this particular airway, causing the patient to die.

Quick, draw up the meds. Give 'em, let's go go go.

It's done. Patient's paralyzed. I put the laryngoscope in their mouth, try to visualize their vocal cords. Oh god there's a ton of secretions in the throat. Quick, I need to suction it out. Perfect, I can see cords, but there are more secretions actively coming out of the trachea. Quick, give me the endotracheal tube before my view gets obscured again. Bam. Success.

I'd like to see your robot of the future match the lifesaving procedure I did in ~15 seconds. Is computer vision going to be perfect by 2025? Is its analysis of a difficult airway situation going to be perfect? Because…it has to be. Will it know when to try to tube the patient vs. when to try airway adjuncts? Will it even be able to quickly recognize a patient holding hands with Death just by laying eyes on them?

Conclusion

I respect that Khosla is a successful man. I respect his confidence. I respect his optimism. I have no respect for his belief that 80% of doctors can be replaced by algorithms in the near future. I find a complete and utter lack of credible evidence to support the claim. I read his TechCrunch article with initial interest which quickly turned into despair, that such a powerful man views my field with such disdain despite having such vast ignorance about how it works.

References

[1] Rivers E, Nguyen B, Havstad S, Ressler J, Muzzin A, Knoblich B, Peterson E, Tomlanovich M. Early goal-directed therapy in the treatment of severe sepsis and septic shock. N Engl J Med 2001 Nov 8;345(19):1368-77.
Early goal-directed therapy in the treatment of… [N Engl J Med. 2001]

[2] Shapiro NI, Wolfe RE, Moore RB, Smith E, Burdick E, Bates DW. Mortality in Emergency Department Sepsis (MEDS) score: a prospectively derived and validated clinical prediction rule. Crit Care Med 2003 Mar;31(3):670-5.
Mortality in Emergency Department Sepsis (MEDS… [Crit Care Med. 2003]

[3] Bone RC, Fisher CJ, Jr., Clemmer TP, Slotman GJ, Metz CA, Balk RA. A controlled clinical trial of high-dose methylprednisolone in the treatment of severe sepsis and septic shock. N Engl J Med 1987 Sep 10;317(11):653-8.
A controlled clinical trial of high-dose methyl… [N Engl J Med. 1987]

[4] Angus DC, Linde-Zwirble WT, Lidicker J, Clermont G, Carcillo J, Pinsky MR. Epidemiology of severe sepsis in the United States: analysis of incidence, outcome, and associated costs of care. Crit Care Med 2001.
Epidemiology of severe sepsis in the United St… [Crit Care Med. 2001]

[5] Wilhelms S, Huss F, Granath G, Sjoberg F. Assessment of incidence of severe sepsis in Sweden using different ways of abstracting International Classification of Diseases codes: difficulties with methods and interpretation of results. Crit Care Med 2010 Jun;38(6):1442-9.
Assessment of incidence of severe sepsis in Sw… [Crit Care Med. 2010]

View Answer on Quora

Advertisements
This entry was posted in Uncategorized. Bookmark the permalink.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s