Deep Medicine

chapter one

INTRODUCTION TO DEEP MEDICINE

“YOU SHOULD HAVE YOUR INTERNIST PRESCRIBE ANTI-DEPRESSION medications,” my orthopedist told me.

My wife and I looked at each other, bug-eyed, in total disbelief. After all, I hadn’t gone to my one-month post-op clinic visit following a total knee replacement seeking psychiatric advice.

My knees went bad when I was a teenager because of a rare condition known as osteochondritis dissecans. The cause of this disease remains unknown, but its effects are clear. By the time I was twenty years old and heading to medical school, I had already had dead bone sawed off and extensive reparative surgery in both knees. Over the next forty years, I had to progressively curtail my physical activities, eliminating running, tennis, hiking, and elliptical exercise. Even walking became painful, despite injections of steroids and synovial fluid directly into the knee. And so at age sixty-two I had my left knee replaced, one of the more than 800,000 Americans who have this surgery, the most common orthopedic operation. My orthopedist had deemed me a perfect candidate: I was fairly young, thin, and fit. He said the only significant downside was a 1 to 2 percent risk of infection. I was about to discover another.

After surgery I underwent the standard—and, as far as I was told, only—physical therapy protocol, which began the second day after surgery. The protocol is intense, calling for aggressive bending and extension to avoid scar formation in the joint. Unable to get meaningful flexion, I put a stationary bicycle seat up high and had to scream in agony to get through the first few pedal revolutions. The pain was well beyond the reach of oxycodone. A month later, the knee was purple, very swollen, profoundly stiff, and unbending. It hurt so bad that I couldn’t sleep more than an hour at a time, and I had frequent crying spells. Those were why my orthopedist recommended antidepressants. That seemed crazy enough. But the surgeon then recommended a more intensive protocol of physical therapy, despite the fact that each session was making me worse. I could barely walk out of the facility or get in my car to drive home. The horrible pain, swelling, and stiffness were unremitting. I became desperate for relief, trying everything from acupuncture, electro-acupuncture, cold laser, an electrical stimulation (TENS) device, topical ointments, and dietary supplements including curcumin, tart cherry, and many others—fully cognizant that none of these putative treatments have any published data to support their use.

Joining me in my search, at two months post-op, my wife discovered a book titled Arthrofibrosis. I had never heard the term, but it turned out to be what I was suffering from. Arthrofibrosis is a complication that occurs in 2 to 3 percent of patients after a knee replacement—that makes the condition uncommon, but still more common than the risk of infection that my orthopedist had warned me about. The first page of the book seemed to describe my situation perfectly: “Arthrofibrosis is a disaster,” it said. More specifically, arthrofibrosis is a vicious inflammation response to knee replacement, like a rejection of the artificial joint, that results in profound scarring. At my two-month post-op visit, I asked my orthopedist whether I had arthrofibrosis. He said absolutely, but there was little he could do for the first year following surgery—it was necessary to allow the inflammation to “burn out” before he could go back in and remove the scar tissue. The thought of going a year as I was or having another operation was making me feel even sicker.

Following a recommendation from a friend, I went to see a different physical therapist. Over the course of forty years, she had seen many patients with osteochondritis dissecans, and she knew that, for patients such as me, the routine therapeutic protocol was the worst thing possible. Where the standard protocol called for extensive, forced manipulation to maximize the knee flexion and extension (which was paradoxically stimulating more scar formation), her approach was to go gently: she had me stop all the weights and exercises and use anti-inflammatory medications. She handwrote a page of instructions and texted me every other day to ask how “our knee” was doing. Rescued, I was quickly on the road to recovery. Now, years later, I still have to wrap my knee every day to deal with its poor healing. So much of this torment could have been prevented.

As we’ll see in this book, artificial intelligence (AI) could have predicted that my experience after the surgery would be complicated. A full literature review, provided that experienced physical therapists such as the woman I eventually found shared their data, might well have indicated that I needed a special, bespoke PT protocol. It wouldn’t only be physicians who would get a better awareness of the risks confronting their patients. A virtual medical assistant, residing in my smartphone or my bedroom, could warn me, the patient, directly of the high risk of arthrofibrosis that a standard course of physical therapy posed. And it could even tell me where I could go to get gentle rehab and avoid this dreadful problem. As it was, I was blindsided, and my orthopedist hadn’t even taken my history of osteochondritis dissecans into account when discussing the risk of surgery, even though he later acknowledged that it had, in fact, played a pivotal role in the serious problems that I encountered.

Much of what’s wrong with healthcare won’t be fixed by advanced technology, algorithms, or machines. The robotic response of my doctor to my distress exemplifies the deficient component of care. Sure, the operation was done expertly, but that’s only the technical component. The idea that I should take medication for depression exemplifies a profound lack of human connection and empathy in medicine today. Of course, I was emotionally depressed, but depression wasn’t the problem at all: the problem was that I was in severe pain and had Tin Man immobility. The orthopedist’s lack of compassion was palpable: in all the months after the surgery, he never contacted me once to see how I was getting along. The physical therapist not only had the medical knowledge and experience to match my condition, but she really cared about me. It’s no wonder that we have an opioid epidemic when it’s a lot quicker and easier for doctors to prescribe narcotics than to listen to and understand patients.

Almost anyone with chronic medical conditions has been “roughed up” like I was—it happens all too frequently. I’m fortunate to be inside the medical system, but, as you have seen, the problem is so pervasive that even insider knowledge isn’t necessarily enough to guarantee good care. Artificial intelligence alone is not going to solve this problem on its own. We need humans to kick in. As machines get smarter and take on suitable tasks, humans might actually find it easier to be more humane.

AI in medicine isn’t just a futuristic premise. The power of AI is already being harnessed to help save lives. My close friend, Dr. Stephen Kingsmore, is a medical geneticist who heads up a pioneering program at the Rady Children’s Hospital in San Diego. Recently, he and his team were awarded a Guinness World Record for taking a sample of blood to a fully sequenced and interpreted genome in only 19.5 hours.¹

A little while back, a healthy newborn boy, breastfeeding well, went home on his third day of life. But, on his eighth day, his mother brought him to Rady’s emergency room. He was having constant seizures, known as status epilepticus. There was no sign of infection. A CT scan of his brain was normal; an electroencephalogram just showed the electrical signature of unending seizures. Numerous potent drugs failed to reduce the seizures; in fact, they were getting even more pronounced. The infant’s prognosis, including both brain damage and death, was bleak.

A blood sample was sent to Rady’s Genomic Institute for a rapid whole-genome sequencing. The sequence encompassed 125 gigabytes of data, including nearly 5 million locations where the child’s genome differed from the most common one. It took twenty seconds for a form of AI called natural-language processing to ingest the boy’s electronic medical record and determine eighty-eight phenotype features (almost twenty times more than the doctors had summarized in their problem list). Machine-learning algorithms quickly sifted the approximately 5 million genetic variants to find the roughly 700,000 rare ones. Of those, 962 are known to cause diseases. Combining that information with the boy’s phenotypic data, the system identified one, in a gene called ALDH7A1, as the most likely culprit. The variant is very rare, occurring in less than 0.01 percent of the population, and causes a metabolic defect that leads to seizures. Fortunately, its effects can be overridden by dietary supplementation with vitamin B6 and arginine, an amino acid, along with restricting lysine, a second amino acid. With those changes to his diet made, the boy’s seizures abruptly ended, and he was discharged home thirty-six hours later! In follow-up, he is perfectly healthy with no sign of brain damage or developmental delay.

The key to saving this boy’s life was determining the root cause of his condition. Few hospitals in the world today are sequencing the genomes of sick newborns and employing artificial intelligence to make everything known about the patient and genomics work together. Although very experienced physicians might eventually have hit upon the right course of treatment, machines can do this kind of work far quicker and better than people.

So, even now, the combined efforts and talents of humans and AI, working synergistically, can yield a medical triumph. Before we get too sanguine about AI’s potential, however, let’s turn to a recent experience with one of my patients.

“I want to have the procedure,” my patient told me on a call after a recent visit.

A white-haired, blue-eyed septuagenarian who had run multiple companies, he was suffering from a rare and severe lung condition known as idiopathic—a fancy medical word for “of unknown cause”—pulmonary fibrosis. It was bad enough that he and his pulmonologist had been considering a possible lung transplant if it got any worse. Against this backdrop he began to suffer a new symptom: early-onset fatigue that left him unable to walk more than a block or swim a lap. He had seen his lung doctor and had undergone pulmonary function tests, which were unchanged. That strongly suggested his lungs weren’t the culprit.

He, along with his wife, then came to see me, very worried and depressed. He took labored, short steps into the exam room. I was struck by his paleness and look of hopelessness. His wife corroborated his description of his symptoms: there had been a marked diminution of his ability to get around, to even do his daily activities, let alone to exert himself.

After reviewing his history and exam, I raised the possibility that he might have heart disease. A few years previously, after he began to suffer calf pain while walking, he had stenting of a blockage in his iliac artery to the left leg. This earlier condition raised my concern about a cholesterol buildup in a coronary artery, even though he had no risk factors for heart disease besides his age and sex, so I ordered a CT scan with dye to map out his arteries. The right coronary artery showed an 80 percent narrowing, but the other two arteries were free of significant disease. It didn’t fit together. The right coronary artery doesn’t supply very much of the heart muscle, and, in my thirty years as a cardiologist (twenty of which involved opening coronary arteries), I couldn’t think of any patients with such severe fatigue who had narrowing in only the right coronary artery.

I explained to him and to his wife that I really couldn’t connect the dots, and that it might be the case of a “true-true, unrelated”—that the artery’s condition might have nothing to do with the fatigue. His underlying serious lung condition, however, made it conceivable that the narrowing was playing a role. Unfortunately, his lung condition also increased the risk of treatment.

I left the decision to him. He thought about it for a few days and decided to go for stenting his right coronary artery. I was a bit surprised, since over the years he had been so averse to any procedures and even medications. Remarkably, he felt energized right after the procedure was done. Because the stent was put in via the artery of his wrist, he went home just a few hours later. By that evening, he had walked several blocks and before the week’s end he was swimming multiple laps. He told me he felt stronger and better than he had for several years. And, months later, the striking improvement in exercise capacity endured.

What’s remarkable about this story is that a computer algorithm would have missed it. For all the hype about the use of AI to improve healthcare, had it been applied to this patient’s data and the complete corpus of medical literature, it would have concluded not to do the procedure because there’s no evidence that indicates the opening of a right coronary artery will alleviate symptoms of fatigue—and AI is capable of learning what to do only by examining existing evidence. And insurance companies using algorithms certainly would have denied reimbursement for the procedure.

But the patient manifested dramatic, sustained benefit. Was this a placebo response? That seems quite unlikely—I’ve known this man for many years, and he tends to minimize any change, positive or negative, in his health status. He seems a bit like a Larry David personality with curbed enthusiasm, something of a curmudgeon. Ostensibly, he would be the last person to exhibit a highly exaggerated placebo benefit.

In retrospect, the explanation likely does have something to do with his severe lung disease. Pulmonary fibrosis results in high pressures in the pulmonary arteries, which feed blood to the lungs, where the blood becomes oxygenated. The right ventricle is responsible for pumping that blood to the heart; the high blood pressure in the arteries meant that it would have taken a lot of work to force more blood in. That would have stressed the right ventricle; the stent in the right coronary artery, which supplies the right ventricle, would have alleviated the stress on this heart chamber. Such a complex interaction of one person’s heart blood supply with a rare lung disease had no precedent in the medical literature.

This case reminds us that we’re each a one-of-a-kind intricacy that will never be fully deconvoluted by machines. The case also highlights the human side of medicine: We physicians have long known that patients know their body and that we need to listen to them. Algorithms are cold, inhumane predictive tools that will never know a human being. Ultimately, this gentleman had a sense that his artery narrowing was the culprit for his symptoms, and he was right. I was skeptical and would certainly not have envisioned the magnitude of impact, but I was thrilled he improved.

AI HAS BEEN sneaking into our lives. It is already pervasive in our daily experiences, ranging from autocomplete when we type, to unsolicited recommendations based on Google searches, to music suggestions based on our listening history, to Alexa answering questions or turning out the lights. Conceptually, its roots date back more than eighty years, and its name was coined in the 1950s, but only recently has its potential impact in healthcare garnered notice. The promise of artificial intelligence in medicine is to provide composite, panoramic views of individuals’ medical data; to improve decision making; to avoid errors such as misdiagnosis and unnecessary procedures; to help in the ordering and interpretation of appropriate tests; and to recommend treatment. Underlying all of this is data. We’re well into the era of Big Data now: the world produces zettabytes (sextillion bytes, or enough data to fill roughly a trillion smartphones) of data each year. For medicine, big datasets take the form of whole-genome sequences, high-resolution images, and continuous output from wearable sensors. While the data keeps pouring out, we’ve really processed only a tiny fraction of it. Most estimates are less than 5 percent, if that much. In a sense, it was all dressed up with nowhere to go—until now. Advances in artificial intelligence are taming the unbridled amalgamation of Big Data by putting it to work.

There are many subtypes of AI. Traditionally machine learning included logistic regression, Bayesian networks, Random Forests, support vector machines, expert systems, and many other tools for data analysis. For example, a Bayesian network is a model that provides probabilities. If I had a person’s symptoms, for example, such a model could yield a list of possible diagnoses, with the probability of each one. Funny that in the 1990s, when we did classification and regression trees to let the data that we collected speak for itself, go into “auto-analyze” mode, without our bias of interpretation, we didn’t use the term “machine learning.” But now that form of statistics has undergone a major upgrade and achieved venerability. In recent years, AI tools have expanded to deep network models such as deep learning and reinforcement learning (we’ll get into more depth in Chapter 4).

The AI subtype of deep learning has gained extraordinary momentum since 2012, when a now-classic paper was published on image recognition.²

The number of new deep learning AI algorithms and publications has exploded (Figure 1.1), with exponential growth of machine recognition of patterns from enormous datasets. The 300,000-fold increase in petaflops (computing speed equal to one thousand million million [10¹⁵] floating-point operations per second) per day of computing used in AI training further reflects the change since 2012 (Figure 1.2).