Following on the heels of The Creative Destruction of Medicine and The Patient Will See You Now, Dr. Eric Topol has published his newest book, Deep Medicine, earlier this year. While the first two books highlight how technological disruption will first digitize and then democratize medicine, his newest is about deep learning, and how artificial intelligence (AI) has the potential to ultimately deepen the patient-physician bond. It is a lofty claim, but after reading the book, his hopeful and optimistic vision of medicine seems plausible with deliberate and well-intentioned decisions regarding AI. Before delving into Deep Medicine, however, a quick aside about predictive medicine.
Machine learning, a subset in the broader AI category, performs tasks by relying on patterns and inference – a divergence from older rules-based algorithms. Using this type of pattern-based recognition, machine learning can help to calculate a person’s risk for disease based on the risk factors that previous patients had, allowing for earlier intervention. It can help us to treat risk, rather than disease.
While these types of analyses are useful in some cases, in others, they fail to take into account iatrogenics, i.e., the harm done to the patient because of the medical intervention itself. To do this, it is critical to assess the number needed to treat (NNT) and the number needed to harm (NNH). In other words, how many people must be given the intervention for one to benefit, and how many people will be harmed, by way of adverse side effects, per number of people given the treatment? When taking these factors into account, interventions that once seemed like a good idea, may no longer appear to provide net benefit.
To quote Nicholas Nassim Taleb, we want to make the system antifragile, which we do by minimizing Black Swan events. In medicine, the Black Swan, or devastating event, is the diseased state. Thus, we intervene when there is great benefit and little downside, as in disease. However, when there is no disease, and we intervene anyway using risk prediction, we may produce little upside, and more profound downside. The downside is the fragilities we have introduced into the system by way of medical side effects (that may compound with further treatments of those side effects) when treating a healthy patient. This is a fancy way to say that we intervene when the NNT is low, and the NNH is high, i.e., high upside, little downside.
Returning to Deep Medicine, Dr. Topol begins the book with an explanation on System 1 and System 2 thinking, adapted from economist, Daniel Kahneman. System 1 thinking is reflexive and largely unconscious, while System 2 thinking is reflective and cognitive. When making diagnoses, doctors are trained in System 1 thinking, which is surprisingly accurate but also subject to biases.
To give an example of this type of thinking in medicine, we can examine current medical training. When going through cases, an attending physician will often start with a keyword, such as “pain radiating to the back.” With this clue, students reflexively answer aortic dissection or pancreatitis. Depending on the next clue, for instance, uncontrolled hypertension, the student can hone in on the diagnosis of aortic dissection. The point is students are trained to form intuitive, largely unconscious associations between patient symptoms, risk factors, diseases, and treatments – System 1 thinking. But what if the disease never comes into the physician’s mind to begin with? There are more than thousands of diseases, and as research continues, we are able to subtype those diseases to deeper levels. On a broad scale, epilepsy is due to synchronous neuronal activity in the brain. At the genomic level, however, it can be due to mutations in hundreds of genes, which may make a difference for long-term management.
Systems 1 thinking is crucial, but with clear limitations given our big data age. It is in this space, Dr. Topol argues, that machines can assist by aiding in diagnostics and limiting iatrogrenics. To borrow again from Nicholas Nassim Taleb, technology is best leveraged when used for Black Swan events (severe disease states) or in a via negativa way, i.e., addition through subtraction. Deep Medicine is ripe with examples of AI used in this way.
Early in his book, Dr. Topol demonstrates the case of an eight-day-old newborn who presents to the emergency department with intractable seizures. While doctors are initially perplexed by the diagnosis, genome-sequencing revealed a variant in the ALDH7A1 gene as the most likely culprit, producing an exceedingly rare metabolic defect. Dietary changes, including adding some amino acids and restricting others, allow the boy to live a healthy life without development delay. This case demonstrates AI at its best, aiding in the case of a profoundly sick patient and intervening in a way that is life-changing. It is working on the Black Swan events, providing immense upside for little downside. It has a low NNT.
Later in the book, Dr. Topol shows how factoring in genetic data using inexpensive gene arrays can help predict who will benefit from a statin. Rather than the current population-based approach that prescribes to many people to benefit only a few, this new approach utilizes genomic information to attain more precise diagnostics. This is an example of the via negativa approach. By using AI to be more precise in medication administration, we can save on the adverse effects for a patient who may not benefit, and help a patient who may. In this way, it would increase the NNH for common medical interventions.
Deep Medicine is an audacious book, and I can only imagine the work Dr. Topol must have done to keep up with the ever-changing role of technology in health care. Although I was initially skeptical of how the digital revolution would transform medicine, Dr. Topol has demonstrated that AI used on Black Swan events and in a via negativa way has the potential to improve the current system. If done with caution and care, perhaps AI can restore trust in medicine, and create empathic patient-physician relationships.
John Paul Mikhaiel is a medical student.
Image credit: Shutterstock.com
