Leave a comment

The Probability of Artificial Intelligence-Driven Miracles

Medical decision-making is a game of probabilities. Sapper Morton reminds us in Blade Runner 2049, “You newer models are happy scraping … because you’ve never seen a miracle.” In the real world, miracles do not exist. Not in a statistical sense, and reliably identifying and discounting “rare-events” can sharpen decision-making. The likelihood of an improbable event occurring somewhere at some point is, of course, high. People win the lottery, die on airplanes and are cured of incurable cancer: all low probability events. In medicine, misinterpretation of probabilities can have devastating consequences and improving medical decision-making should be a top priority. Artificial intelligence (AI) is finally emerging as a possible solution.

Medical judgment is complicated. Three tasks are paramount: 1) generate the correct diagnosis and assess potential risks (efficiently, if possible); 2) generate a treatment plan; and 3) implement the treatment plan. This may not necessarily be the treatment plan with the highest theoretical response rate but the treatment plan that, when implemented, has the highest probability of succeeding (e.g., if the patient cannot afford the medication deemed most efficacious, the probability of successfully implementing that treatment plan is lower than that of a treatment plan utilizing a less efficacious but more affordable medication).

Human-like cognition emerged roughly 2.6 million years ago and has allowed us to evolve into self-proclaimed masters of the planet. If we only focus on the last few millennia, however, we could jump back in time by thousands of years, pluck a child from the arms of her mother somewhere on the African savanna and watch her navigate our modern world without obvious deficiencies. Send her to Wall Street, and she may start selling toxic assets with the best of them. Send her to medical school, and she will emerge a doctor. Send her to Congress, and she too will forget her ethics.

This observation should be worrisome: It highlights the near absence of recent human evolution, even as our environment continues to change dramatically. The probabilities game of 10,000 B.C.E. did not present the same set of complexities one finds in today’s world. Navigating the challenges of medical school or managing the care of an intensive care unit patient requires the mastery of a completely different set of skills compared to hunting and killing a mammoth.

One hundred billion neurons and trillions of synaptic connections may be enough to split atoms, but they fail to completely shield us from egregious fallibilities. Well over 100 biases affect clinical decision-making, and diagnostic failure rates hover between 10 to 15 percent. Medical errors — some due to poor decision-making — cause the deaths of over 250,000 Americans annually and are the third most common cause of death in the United States after heart disease and cancer. Medical decisions are further obscured by “therapeutic illusion,” the overestimation of benefits of interventions, which prompts ordering of inappropriate tests and implementation of suboptimal treatment.

The stunning success of AlphaGo Zero further emphasizes the limitations of our analytical capacity and promises a cognitive breakthrough entirely independent of evolution’s sluggish pace. AlphaGo Zero is a program developed by DeepMind that used “reinforcement learning” to master and then dominate — human champions did not stand a chance — the Chinese game Go. Reinforcement learning (trial and error) is free of human guidance and is based on an artificial neural network that calculates the probability of maximizing the chance of victory with every conceivable move. In the process of mastering Go, AlphaGo Zero invented novel moves unidentified by human players. Could AlphaGo Zero minimize errors in health care and optimize outcomes?

“Good morning, Mr. Johnson. I am Doctor Meursault. You have high blood pressure, diabetes and obesity and you live in poverty. The probability of health improvement would be maximized if you found better employment, better health care coverage and a better neighborhood. Based on available data, your educational background and the current job market, region X of this country would provide you with the highest chance of success. You must complete ‘n’ job applications to maximize this probability. In city Z, your goal should be to move into neighborhood Y, which would give you access to affordable dietary options to prevent a health crisis. Crime is low in the suggested neighborhood, and you would be able to exercise outdoors. In the meantime, you might avail yourself of medications A and B. Re-evaluate shifts in probability in six months.”

The data available for AI-driven clinical decision-making as imagined in the previous paragraph are not broadly robust. Good data are the bottleneck. Without them, AlphaGo Zero cannot learn to maximize health care outcomes. For example, the information needed to broadly assess Mr. Johnson’s well-being is not available yet.

Nevertheless, several projects are ongoing to explore the potential of AI in health care. One such project is referred to as “Deep Patient,” in which AI analyzes electronic health records to generate patient-specific clinical predictions. Deep Patient trained using data from 700,000 patients and applies what it has learned to predict disease in newly encountered patients without any expert instruction. Deep Patient is particularly good at identifying cancers and schizophrenia. Diagnosing schizophrenia is notoriously difficult, and it remains a mystery how Deep Patient does it.

Broad commitment to AI would improve its chances of mastering clinical decision-making sooner rather than later. More projects like Deep Patient should be funded and pursued. As with the game Go, AI may discover hidden clues and patterns that we lack the imagination or bandwidth to appreciate and that may fundamentally shift how clinical decision-making is practiced. Despite its uncanny computational and analytical capacity, AlphaGo Zero cannot yet maximize its own utility and its implementation in healthcare remains a decision for the evolutionarily-limited to make. “AlphaGo Zero, what is the probability that you are a miracle?”

Tim Beck (4 Posts)

Medical Student Editor Emeritus

Drexel University College of Medicine

I am an MD/PhD Candidate at Drexel University College of Medicine/Fox Chase Cancer Center. My research focuses on cancer cell signaling, drug resistance, cancer cell invasion and discovery of prognostic biomarkers. Politics (national and international), foreign affairs and healthcare policy are additional topics I am particularly interested in.