Deep Medicine: How Artificial Intelligence Can Make Healthcare Human Again, by Eric Topol
Eric Topol hit the news big time in 2004 as the chief whistleblower questioning the validity of research that led the FDA to approve the Merck drug, Vioxx. Amid scandal, Merck withdrew Vioxx and paid a big fine. The New York Timesinvestigated Topol’s own connections with the pharmaceutical industry. Merck protested his role in its Vioxx debacle to the Cleveland Clinic, which eliminated Topol’s position as its chief academic officer.
But Topol rebounded higher. The Scripps Research Institute in San Diego beckoned him. There he soon founded the Scripps Genomic Medicine program, launching him deep into the possibilities for artificial intelligence in medicine. However, Topol is very human: a scandal survivor, a patient with a difficult condition, a practicing physician, a top cross-disciplinary researcher, and no stranger to controversy.
Consequently, Topol reviews health care from many perspectives. Broader than many books on AI and healthcare, this book readably synthesizes in detail many areas about which most readers know nothing. Of his 30 books, plus stacks of research papers, Topol says that Deep Medicine is the most difficult writing he has ever attempted. It merits a super long review.
What’s Artificial Intelligence (AI)?
AI is training search-and-relate algorithms to analyze patterns in massive data sets. Artificial intelligence has evolved into a new capability for humans to see what otherwise we could never see, opening new worlds of discovery, not unlike the development of microscopes 400 years ago. The new scientific possibilities are hypothesis-free discovery to see patterns not predicted, and eliminating the need for statistical sampling and probability estimating because one is examining a total known database. Deep Medicine covers its profound effects in health care.
Medicine creates massive data sets. Genetic databases alone gobble terabytes of storage like M&Ms. More data keeps piling up: genomics, epigenomics, proteomics, and microbiomic symbionts (the little critters that are 90% of the cells we carry around). In addition, using AI for drug discovery explores a number of possible combinations beyond human comprehension, 1060 being a rough estimate. That’s a bigger number than all the atoms in our solar system.
Impressed? Topol is impressed with the possibilities; not so much with accomplishments so far. Yes, there are achievements like using an app to triage an emergency patient in 30 seconds that would busy an appless physician for 4 hours. But most applications have unsung challenges, like incomplete data sets, biased data sets (because some types of cases were excluded or mislabeled), plain error-ridden data sets, and biased training vectors for algorithms that take off from biased human logic doing diagnoses.
Well developed, AI is vastly superior to humans in narrow tasks. One of those is reading X-rays and other medical images. Enthusiasts can jump to the conclusion that if an AI scanner has 92% accuracy and a human radiologist only 85%, replace the human with the AI – to do the same process better and cheaper. That’s an administrator’s business logic.
Not so fast, says Topol. Why not combine the two and get more like 99% accuracy, reducing the expense and anxiety of false positive readings, eliminating many patients re-taking tests? A human radiologist may be excellent detecting what she expects to see, but miss what she is not looking for. AI can be trained to seek multiple possibilities.
But most of all, AI is not human. Topol questions whether AI can ever be trained to have an artificial sense of empathy for humans. That’s the essence of “good bedside manners” by a human physician undistracted by clinical analysis, administrative busy work, and organizational money incentives. The best potential for medical AI is to free physicians from technical tasks so that they can resume being Dr. Welby’s (the 1950s TV doctor).
Combining AI with human empathetic capabilities to forge a better civilization is a theme threading through Deep Medicine,building to its grand finale. Along the way, Topol reveals much that this reviewer did not know. AI is not an unmixed blessing. We must stir our ethics and empathy to learn to control the purposes for which it is used. Doing that might be the biggest boon to come from AI. If we can’t do that, we may have created a digital Frankenstein.
Reductionism, RIP
Twenty years ago, geneticists hit the peak of reductionist naiveté by assuming that human genomes could predict each individual’s risk of various diseases. They assumed that a gene, or a simple combination of them, was a direct causal precursor of each disease. That’s reductionist, or “linear” thinking. Additional research quickly disabused geneticists of that notion.
Researchers quickly began to see that the natural world at the micro level is horrendously complex. Causality loops entangle the possibilities. At a minimum, significantly rolling back this uncertainty requires all the data from each person’s Electronic Health Record, assuming that it is accurate, which is very unlikely today. Then one must consider microbiomes, environmental influences changing gene expression (epigenetics) … and the list goes on. The intrusion of gathering such detailed personal data would be unacceptable to many of us. Even if we get past these objections, and AI does its thing, the environment we are exposed to is always changing – sometimes suddenly. Once true is not always true.
Gathering massive volumes of data requires automating data collection with smart phones, watches, wrist bands, and patches. How wired do you really want to be? And if you are, is the data valid? AI enthusiasts are warming up people to accept constant monitoring using slogans like, “Manage your life, digitally,” but only one organization, iCarbonX in China, is today trying to validate data collection with the gold standard – randomized, controlled trials.
For the record, Topol is adamant that AI is a black box spitting out solutions. None should be accepted to guide clinical practice unless humans validate its sources of data and deduce the chain of logic by which the black box discovered patterns hidden in its input data, or the conclusions that it reached. Humans must retain the prerogative of questioning both the input to AI and the output from it.
As the iCarbonX project shows, China is capturing the lead in AI research, and not just in health care. Millions of Chinese are amenable to intrusive surveillance for data collection – Big Data. They can’t opt out of it. A growing army of Chinese AI researchers is poised to surpass the United States. And China hopes to use medical AI to supplement a shortage of physicians by augmenting the knowledge of lesser trained health workers with AI. They have incentive to move far, fast. In 2018, China proclaimed world leadership in AI to be their moonshot, like the United States’ Apollo Program earlier. Our media obsession with China’s Belt and Road initiative is so 20thcentury. China foresees artificial intelligence becoming the base of 21stcentury intellectual and commercial superiority, and probably military superiority too.
But for now, a snag for everyone in AI is that massive data computation consumes massive energy. Redesigning integrated circuits around more specialized functions might cut the energy needed by an order of magnitude or more (Amazon is working on this). However, energy requirements could still be a show stopper.
Electronic Health Records (EHR)
American Electronic Health Records as well as manual records are a mess: narrow, incomplete, and error ridden. EHRs are structured to aid billing, not deep clinical learning. Worse, they are collected, bought, and sold for commercial intelligence and marketing.
Topol insists that if AI is to progress in medicine, each person must own their own EHR, to share willingly. They should not be shared surreptitiously, as traded commodities. Owning your own EHR should be a civil right. Otherwise, the average patient will not trust AI. Topol devotes two pages of Deep Medicine to 24 cautions safeguarding your personal health records.
Patients’ disclosures and EHRs are ethically tricky, especially in areas like mental health, where even being examined may deny an applicant a job. Health insurance companies access EHRs to screen applicants for pre-existing conditions. But interviews to collect health data have “interesting” anomalies. For instance, some patients divulge deeply embarrassing conditions more readily to an AI robot than to a live physician.
The issues with EHRs come down to cleansing them of incorrect data (mistakes are made correcting mistakes), restructuring them for clinical learning, keeping them up to date, and preventing EHRs from being used for purposes other than health improvement. Given the detailed fuss work (and inevitable errors) doing this, no wonder AI companies feel compelled to capture medical measurements as they happen, compiling their own databases for deep AI learning. (This is reminiscent of the tribulations of many companies computerizing inventory systems. To rely on numbers from computers, they had to discipline their data accuracy.)
We Are What We Eat: Deep Diet
Topol devotes a whole chapter to diets and dubious dietary research. Nearly all diet studies rely on human self-reporting, which is notoriously inaccurate. The most meticulous humans can’t document exactly what they eat, and may not know what they are eating. (Do you carefully read the ingredients on every food package? And nutrient info based on test averages doesn’t tell you what is in the tidbit you are wolfing down. And plain ignorance is overwhelming. Medical students have very little training in nutrition, and much of that is misleading.)
AI has started to carve into this chaos. Massive data reviews show pretty convincingly, for instance, that high carbohydrate intake, not fat intake, is more responsible for heart disease. Another review of 700,000 people concluded that about 45% of all deaths from heart disease, stroke, or diabetes can be attributed to the decedents’ diets – high in processed meats, salt, etc. But these studies still analyze noisy data for dietary factors common to everyone.
The same diet is not good for everyone. Individuals have unique allergies and unique gut microbiomes. Topol is amazed by the huge discrepancies in dietary recommendations. Diet fads are justified by incomplete, often cherry-picked research. Food companies influence dietary studies, sometimes blatantly, and promote profitable offerings more than nutritious ones.
Topol sees a beacon of hope in research from the Israeli Weizmann Institute showing that foods trigger glucose response peaks that appear to be unique to each individual. Summarizing findings as averages makes this variation disappear. Identifying diet response profiles for individuals requires intense personal monitoring and processing huge gobs of data, AI’s forte. This study is summarized in a book, The Personalized Diet. A web site explains the data gathering and solicits people to participate. The more data that is gathered, the more the AI system refines its knowledge. Of course, each participant should benefit too.
Intrigued, Topol contacted the Weizmann Institute, visited, and went through the two-week regimen necessary to collect data for their database – including a stool sample to characterize his microbiome. This disturbed his work routines. He also noted that knowing he was being monitored altered his eating habits.
Nonetheless, Topol discovered things about himself, and despite reservations, he emerged convinced that this line of research holds promise. Perhaps a personal diet that “optimizes” personal health will become possible, or at least a diet that identifies personal eating choices that are bad for you.
Runaway Health Care Costs
The best health care is maintaining a state that does not need curative care. Everyone knows this, but health care costs keep rising anyway. Most of this expense is for curative care. One in eight working Americans works in health care. Health care is edging toward being 20% of all GNP.
Some see health care as a job growth sector. This is incompatible with keeping curative care affordable, much less making it unnecessary. And curative care is bloated. There are 175,000 professional medical coders at the core of a labyrinthine billing system. Topol’s personal experience is that billing costs a medical practice about $20 for each office visit.
Hospital errors and infections are a huge problem. In ICUs, 10% of all patients die of sepsis, and about 20-30% of all hospital deaths while admitted are from sepsis, uncontrollable infection. There is a high risk of patients contracting clostridium difficile infections, highly resistant to antibiotics. One in 25 of all hospital admissions contract a nosocomial infection, one that has its source in the hospital. Treating these cases is extra work, extra expense, and dangerous for patients
Topol realizes the importance of improving the flow of patients through a series of procedures, but never uses the term “lean” to describe it. He trusts AI to reduce the bottlenecks slowing flow. One application is using Medstar AI to dredge up relevant information from 60 pages or so of medical records in a flash. Another is AI assistance weaning patients off ventilation. This now takes hours of close observation. Patterns detected by AI can give earlier warnings about unintended side effects setting patients back. Using AI to interpret tests improves prediction of patient progress with less observation. That is, Topol sees AI as cutting physician technical work time, but to do what? Spend more time getting to know each patient. Physician behavioral support and intuitive knowledge will remain very important to improving patient recoveries, and there’s too little time for it now.
The Personable Physician
Today, about half of all MDs burn out. Physician suicide rate is high. A major reason is that doctors are frustrated by not being able to help patients as well as they know how. They can’t get to know a patient if their noses are always stuck in a screen or in paperwork. It’s worse if they are pressured to do all this while treating as many patients as possible – maximize billing. Doctors become psychological victims of this system as much as the patients they treat.
Topol recommends recruiting youth that are more empathetic into medical education; then train them a little less on bioscience and biotechniques, and coach them more on bonding with patients. Prepare them to become trusted advisors to patients and their families in times of stress, and in practice allow them the time to do it. Trust in a doctor is a wonderful placebo. Nurses and technicians need development for knowledgeable empathy too, but patients need it most from the person controlling their treatment process.
Wisely developed, AI might open this kind of world in healthcare. Turn health care workers’ automatable tasks over to computerization while encouraging them to develop the tacit knowledge to question outcomes from a black box. On top of this, charge physicians with keeping their healthy patients healthy. And compress the financial bloat that now encumbers effective treatment. Topol does not suggest how to find jobs for all the people that might be released from health care work if this is done. His dream is super health care for everyone, even those that are out of a job.