25 Takeaways from Eric Topol’s new book, Deep Medicine
Every once in awhile there comes a well-researched and clearly-written book that deserves to be read end-to-end. While Dr. Eric Topol’s Deep Medicine’s compass points to the future, the book’s heart is rooted in compassion and deep concern for patient care.
My printed copy has so many dog-ears that I’m simply going to list my takeaways for you – with a few comments (>>).
25 Takeaways from Deep Medicine
1) Delays in diagnosis or failure to spot a diagnosis account for a significant number of law suits. The underlying reason is poor chart documentation. AI tools (e.g. Ada, Your.MD, Babylon, Buoy Health) can help.
2) Crowdsourced diagnosis is an upcoming area. Figure One, Health Tap, DocCHIRP, Medscape Consult, CrowdMed (competitions) are some interesting examples. Human Dx has been used by 6,000 doctors and trainees. In the future, AI tools + crowdsourced diagnosis can be real aids for doctors.
>> Our medical knowledge is expanding and growing in complexity. ICD-10 diagnosis codes are 150,000 in number. It’ll soon be impossible for doctors to see patients without the aid of technology.
3) Machine learning works best with raw data and lots of data. If there’s enough data, then the noise gets filtered out on its own.
>> If we take stock of ALL the electronic medical records circulating out there, we’ll notice that the majority is templated and not pertinent. Using that data to create algorithms is dangerous and possibly futile. But with more data, the machine can possible figure out what’s useful and what’s not.
4) AliveCor’s Kardia band is possibly the first FDA approved AI algorithm to help patients with self-diagnosis. It can help users detect atrial fibrillation.
5) Gary Kasparov said IBM’s Deep Blue “didn’t enjoy beating me”.
>> We vastly underestimate the role of human emotion (in this case the joy of competing) in our transactions with machines.
6) 4 areas where deep learning has made big impact already: games, images, voice and speech, and driverless cars.
7) If a doctor makes a medical mistake, it can result in one death or coma. If an AI makes a mistake, it could be devastating – possibly resulting in hundreds or thousands of deaths!
>> Imagine if an algorithm that determines drug dosages for a certain condition goes awry. It’ll be far, far worse than a clinician overdosing one patient.
8) Problem of using current genetic data as input for AI algorithms is recipe for trouble. Because most of the data is from people of European ancestry.
>> The problem is we don’t really have that much data from other ethnicities. So there’s a lot of room in the future to make these algorithms more relevant and accurate.
9) NHS used an app (Streams) that helped triage patients in 30 seconds vs up to 4 hours in the past.
10) The paradox of driverless cars. In the accident that occurred in 2018, the AI detected a pedestrian in the dark but didn’t stop and the human backup driver trusted the AI.
>> I hear this all the time with staff at medical sites: “The computer told me to do this” and they simply do it.
11) In healthcare, there’s scope for building algorithms that are unethical, channeling patient care in a certain direction based on insurance or income levels.
>> Think of someone at a hospital saying this: “the computer suggests that it’s better for you to get a stent placement. Do you want to go ahead?” The problem is a patient will almost never question who told the computer to suggest that and on what basis.
12) In an experiment, 83% of radiologists missed a man in a gorilla suit shaking his fist into images, while reviewing for cancer!
13) It’s not about AI replacing dermatologists. But it’s about AI assisting family doctors who are called to do dermatology cases.
>> I wrote this article about a DNA stool test and colonoscopy and it created much angst among gastroenterologists. The reality is it’s general practitioners who are prescribing alternatives to traditional procedures because patients are approaching them first, not specialists.
14) AI speech processing exceeds performance of human transcriptionists. Can it be used in the medical office to speed up EHR documentation?
15) We are trapped in a binary world of medical diagnosis – normal or abnormal and ignoring rich, finer data on various possibilities.
>> We are barely scratching the surface of what’s possible. The future of AI-based medicine is vast and yet to be uncovered. Google’s using 46 billion data points to predict medical outcomes.
16) Surgery 4.0. Imagine cloud-connected surgeons sharing data to democratize surgical practice.
17) Based on several acoustic biomarkers (voice), AI can sense various mental conditions – from depression, schizophrenia, bipolar disorder and so on.
18) In a study, AI learnt from electronic medical records of 160,000 patients to predict death with great accuracy.
>> Imagine what would happen if this information is released to patients or worse their insurers or employers!
19) Insurers like United Healthcare are experimenting with voice AI to replace keyboarding.
20) It’s likely that AI for medicine will take hold outside of US in countries like China or India. Economist characterized China as “the Saudi Arabia of data” (hospitals are training AI using massive amounts of data).
>> Healthcare is likely to be globalized and more and more uniform in the future. Discoveries in China will impact medicine in US and elsewhere. It’s in our interest to develop a more global mindset in healthcare and be open to new possibilities.
21) Nutrigenomics is an evolving field. Can we personalize food to control disease?
>> It’s too, too early. I doubt we are at the point where we can trust the food we buy yet. How do we then customize it?
22) With $600 million in funding, Chinese company iCarbonX is attempting to collect data on a massive scale across “lifestyle, DNA sequencing, proteomics, metabolomics, immune system via autoantibodies, transciptomics, the gut microbiome, continuous glucose monitoring, and the use of smart toilets and mirrors beyond smart phones.” The plan is to develop a more accurate AI health coach.
>> There many companies that are pursuing similar goals. For the machine to guide our health. We’ll see many forms of AI coaches – possibly not for all therapies but definitely for certain conditions.
23) In 1975, the term “health system” wasn’t coined yet, US spent $800 per patient per year and less than 8% of the GDP. Today, we spend $11,000 per patient, over $3.5 trillion per year and close to 19% of GDP.
>> This is so absurd. And what’s disturbing is that there’s no stopping it. It’ll increase more and more, with no guarantee of better outcomes.
24) AI will profoundly change medicine. Some of these changes will be for the better. Some could make it far worse. We have a choice right now to nudge AI development in the right direction.
25) AI can offer the gift of time to doctors by unshackling them from the burden of data entry into electronic health records. But this requires collaboration across the board from administrators to physician leaders and EHR vendors (some with “gag clauses”). The gift of time is not necessarily to keep seeing more and more patients but to go deeper instead.
My final takeaway: Machine-based medicine need not be the future of healthcare. Instead, AI can move us in the direction of greater human empathy and make space for rebuilding a genuine physician-patient relationship. If we allow it to, it can recreate the space for caring, trust and empathy. To ultimately result in healing.
>> The last chapter had a profound impact on me. It seemed to say, let’s rapidly embrace the future the right way so that we can go back to a simpler past.
Re-reading my dog-eared takeaways, I see that they hardly do justice to what the book represents in its totality. Deep Medicine adds up to much more than this.
Read it if you are in healthcare. But read it even if you are not. Because sooner or later, healthcare affects all of us. So would AI.
Image Credit: Pexels and Amazon