Author: Ron Alterman
When AlphaGo, the machine learning algorithm that plays the ancient Chinese game 'Go', defeated Ke Jie, the world's top-ranked Go player, what was remarkable was not only that the algorithm won, but that it employed strategies that no human had imagined in the 2000 years the game has been played. Commenting on the match, Ke stated, "Last year, it (AlphaGo) was still quite humanlike when it played, but this year it became like a god of Go."1 Similarly, while one may easily imagine that an algorithm could grade diabetic retinopathy from digital retinal images more reliably than humans, one might not anticipate that the same algorithm could also use those images to predict an individual's risk of cardiovascular disease, their systolic blood pressure or their gender (with 97% accuracy).2 This is the power and potential of artificial intelligence (AI): the ability to analyze huge data sets and discover correlations that humans simply cannot imagine. Machine learning is rapidly transforming our economy and society, and make no mistake, it is ideally suited to transform the practice of medicine.
In the 20th century, physicians were selected for their ability to absorb and retain vast quantities of information and apply that information in the clinical setting while taking into account (as best we could) the infinite combinations of personal, social, cultural, and medical attributes that make each patient unique. We acquired that knowledge and its associated language in elite educational institutions, enjoyed almost exclusive access to that knowledge, and were entrusted by society to apply that knowledge equitably and in good faith. We embraced this vaunted role and protected it defiantly, claiming that medicine was an 'art' too complex to be subject to centralized planning or prescribed clinical decision algorithms. The practitioner's 'expert' opinion was what mattered.
In the early part of the 21st century, the foundations of that model have been severely eroded. Evidence-based clinical guidelines are known to reduce costly variations in care and improve outcomes.3 The field of behavioral economics has shown how all humans are subject to unconscious biases that lead us to make poor decisions, whether they are financial, personal, or other.4 Studies of resident performance have demonstrated that our already flawed human decision-making further degrades when we are tired or stressed.5 The institute of medicine estimates that hundreds of thousands of Americans die each year in hospitals as a result of human error.6
The internet has democratized access to medical knowledge so that patients often come to office visits knowing as much about their illness and treatment options as their physician does. Crowd-sourcing has been shown to neutralize individual biases and encompass multiple individual's incomplete knowledge sets into a more complete picture of complex systems.
Healthcare delivery researchers such as Atul Gawande have demonstrated both the sloth with which medicine adopts successful new treatment strategies as compared to other high-risk industries7 and how regional and individual practice variations that have no impact on outcome drive up the cost of healthcare.8 At the same time, rapid genomic and RNA sequencing techniques are allowing treatments to be tailored to a patient's specific genetics-an approach that is already yielding improved survival rates for some cancers. The reality we currently face is that medical knowledge is simply too vast and expanding too quickly for even the best of us to keep up.
Recognizing this, industry will in all likelihood turn to AI as a means both to contain the costs of healthcare and to improve healthcare outcomes. Large corporations such as CVS and Walmart will deliver basic healthcare services in a customer-friendly setting, employing allied healthcare providers equipped with advanced clinical decision algorithms based on more clinical data (i.e., experience) than any one physician could ever acquire. Machine learning algorithms will likely prove superior to humans when interpreting mammograms, chest x-rays, pathology slides, or any information that can be digitized and presented in conjunction with clinical outcome data in sufficient quantities for the algorithm to decipher key associations.
So, what does this mean for American Neurosurgeons? My guess is that a rapid adoption of AI into clinical decision making will result in a loss of autonomy as payors increasingly apply algorithms to dictate care (within reasonable boundaries), particularly for common and costly interventions (i.e., chronic pain and degenerative spine disease). Machine learning will likely allow for rapid and accurate interpretation of CT perfusion scans at all hours, determining which patients are eligible for clot retrieval procedures. Algorithms may prove superior to humans when predicting from CT scans which TBI or stroke patients will require decompressive hemicraniectomy, or interpreting MR spectroscopy to distinguish radiation necrosis from recurrent tumor without the need for biopsy. Likely, AI will impact neurosurgery in a manner that is totally unexpected.
Some will point to the recent high-profile layoffs from the IBM Watson team9 as evidence that the potential for AI is just hype and that human clinical decision-making will continue its long reign for the foreseeable future. To those individuals, I point to the Internet Bubble of 2000, when many early internet companies folded. Eighteen years later, the internet is the central technology of our lives, the means through which we accomplish almost everything. I would also point to the profound impact advanced document search programs have had on the legal profession, decimating armies of junior associates, who once were required to read documents and research precedents. In many respects, AI represents the fruition of the electronic medical record, finally putting all of that data we have so dutifully been uploading to good use. Certainly, AI will impact the so called 'cognitive' sub-specialties more than ours, at least for now. Luckily for us, it will likely be some time before our collective abilities to perform complex, life-sustaining procedures are surpassed by tireless robots guided by even more advanced algorithms. But it is foolish to think that AI will not soon play a significant role in our professional lives, and it is better for all that we embrace it.
THE REALITY WE CURRENTLY FACE IS THAT MEDICAL KNOWLEDGE IS SIMPLY TOO VAST AND EXPANDING TOO QUICKLY FOR EVEN THE BEST OF US TO KEEP UP.
1) “Google’s AlphaGo Defeats Chinese Go Master in Win for A.I.” The New York Times, May 23, 2017. https://www.nytimes.com/2017/05/23/business/googledeepmind-alphago-go-champion-defeat.html
2) Poplin R, Varadarajan AV, Blumer K, Liu Y, McConnell MV, Corrado GS, Peng L, Webster DR. Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning. Nature Biomed Eng 2018; 2:158-164.
3) Gawande A. “What Big Medicine Can Learn from the Cheesecake Factory”. The New Yorker, August 13, 2012; https://www.newyorker.com/magazine/2012/08/13/big-med
4) Thinking Fast and Slow, Daniel Kahneman, Farrar, Straus and Giroux, New York, 2011. “https://www.ncbi.nlm.nih.gov/pubmed/?term=Kahol%20K%5BAuthor%5D&cauthor=true&cauthor_uid=18194679” Kahol K, “https://www.ncbi.nlm.nih.gov/pubmed/?term=Leyba%20MJ%5BAuthor%5D&cauthor=true&cauthor_uid=18194679” Leyba MJ, “https://www.ncbi.nlm.nih.gov/pubmed/?term=Deka%20M%5BAuthor%5D&cauthor=true&cauthor_uid=18194679” Deka M, “https://www.ncbi.nlm.nih.gov/pubmed/?term=Deka%20V%5BAuthor%5D&cauthor=true&cauthor_uid=18194679” Deka V, “https://www.ncbi.nlm.nih.gov/pubmed/?term=Mayes%20S%5BAuthor%5D&cauthor=true&cauthor_uid=18194679” Mayes S, “https://www.ncbi.nlm.nih.gov/pubmed/?term=Smith%20M%5BAuthor%5D&cauthor=true&cauthor_uid=18194679” Smith M, “https://www.ncbi.nlm.nih.gov/pubmed/?term=Ferrara%20JJ%5BAuthor%5D&cauthor=true&cauthor_uid=18194679” Ferrara JJ, “https://www.ncbi.nlm.nih.gov/pubmed/?term=Panchanathan%20S%5BAuthor%5D&cauthor=true&cauthor_uid=18194679” Panchanathan S. Effect of fatigue on psychomotor and cognitive skills. Am J Surg 2008 Feb;195(2):195-204.
5) “To Err is Human” Institute of Medicine, November 1999.
6) The Checklist Manifesto, Atul Gawande, Metropolitan Book Company, Henry Holt and Co., New York, 2009.
7) Gawande A. “The Cost Conundrum”, The New Yorker, June 1, 2009.
8) IBM confirms layoffs in Watson Health. Boston Business Journal, June 4, 2018.