Will AI and Machine Learning bring refocus on the individual, but defined by the patient journey?
“You should have your internist prescribe anti-depression medications.” That’s the startling way Dr. Eric Topol leads off his new book, Deep Medicine, recounting his knee replacement surgery. After being in constant pain for several months, his surgeon suggested having his internal medicine physician put him on anti-depressants. Of course, being in pain for months on end will make you feel depressed. It's natural. So, in this case, depression was not the root cause. Pain from the wrong physical therapy regimen due to a rare condition he had was the source of the problem. He goes through how a particularly enlightened physical therapist, who had seen cases like his, found the root cause and got him on the road to recovery.
Topol’s story is a prime example of how understanding the individual at a deep level, can make all the difference. In fact, he says deep knowledge of the individual is a cornerstone of the coming wave of Artificial Intelligence, AI, in medicine, the topic of his book. Artificial Intelligence and Machine Learning in medicine, will work best when they lead to deep understanding the individual.
For now, we have very superficial knowledge of patients, and very little knowledge of what’s happening in their daily lives. Over the last several decades, we’ve mostly understood medicine by looking at populations, through randomized clinical trials. But in speaking with Machine Learning experts recently, it’s clear that Machine Learning may be best suited to understand us as individuals at a very deep level. Machine learning can lose its predictive power when attempts are made to apply it to populations, even as many are trying. Those efforts may be misplaced.
A recent New Yorker article on the difficulties of getting people off of psychiatric drugs highlights the kinds of issues population-based research can cause:
"A decade after the invention of antidepressants (in the 1950s), randomized clinical studies emerged as the most trusted form of medical knowledge, supplanting the authority of individual case studies. By necessity, clinical studies cannot capture fluctuations in mood that may be meaningful to the patient but do not fit into the study’s categories. This methodology has led to a far more reliable body of evidence, but it also subtly changed our conception of mental health, which has become synonymous with the absence of symptoms, rather than with a return to a patient’s baseline of functioning, her mood or personality before and between episodes of illness.”
Machine learning is well-suited to understand changes and deviations from baseline over time, but when we don’t understand patients in the real world before and after treatment, we may lose the target. We lose what might be normal for that patient. Anti-depressants have been a godsend to many patients, but without looking deeply at individual cases, we can lose the best benefits for each patient.
The New Yorker piece quotes an inventor of one of the earliest anti-depressant drugs, who understood this even at the outset:
"Roland Kuhn, a Swiss psychiatrist credited with discovering one of the first antidepressants, imipramine, in 1956, later warned that many doctors would be incapable of using antidepressants properly, “because they largely or entirely neglect the patient’s own experiences.” The drugs could only work, he wrote, if a doctor is 'fully aware of the fact that he is not dealing with a self-contained, rigid object, but with an individual who is involved in constant movement and change.’"
There’s a lot invested into making sure people are treated for depression. For good reason, it can be deadly. Suicide is the second leading cause of death among young people, aged 10-34. Meanwhile, about 1 in 8 adults in the U.S. is taking anti-depressants. Anyone who’s been to a primary care doctor recently has likely been asked about depression.
Of course, these issues are not limited to mental health. A major RWE evidence study of 165K patients in the UK of statins, the cholesterol-lowering drugs, shows that they 51.2% of patients get sub-optimal results. What happens in the real world? Is it behavior, is it genetics?. Can we better identify who belongs in which 50%? What we’re seeing is sub-optimal value for those patients. As we move to value-based care, we need answers to know our money is well spent.
As we enter a new era of artificial intelligence and machine learning, understanding the patient journey deeply will help to augment population studies to give the best results for each individual, leading to better value-based care. New ways of doing medical research will need to follow, driven by populations of 1. There’s a long way to go, but we’re starting to get the real world data we need to get there.