Why Healthcare GenAI is not your father's AI
Predictive AI has struggled to impact healthcare delivery, due to a fundamental mismatch between the technology and work to be done. Generative AI is different, and can deliver transformational change
Since the 1960s, visionary doctors and technologists have dreamed of computers’ revolutionizing medical care. Medicine is the original knowledge-based profession, and a computer’s ability to store, sort and analyse near infinite information should give it a fundamental advantage over the soggy grey matter of mortal physicians. However, attempts to replace even the most trivial aspects of clinical care have been, for the most part, dismal failures, from the “expert systems” of the 1970s to IBM Watson’s more recent attempts at precision oncology.
There are many reasons for this decades-long slew of disappointments, but the biggest issue has been trying to solve the wrong problems, often with the wrong data. The art of medicine has been developed over millennia, and the bar to add clinical impact to what the best doctors are already doing is quite high. Classical AI models excel at making reasonably accurate predictions at scale, but with limited precision and poor explainability. These are great characteristics for systems that optimize a power grid or serve online advertisements, where aggregate performance is what counts, but much less so for patient care, where every individual patient matters.
Sepsis detection is the best example for this to date. Sepsis kills many people, and early treatment saves lives. These twin facts have attracted large amounts of research efforts, but a high rate of “false positives” in these algorithms drives up unnecessary antibiotic use, and no high quality evidence has shown that an algorithm can actually affect patient outcomes. An ideal system identifies opportunities where a different clinical decision can actually save lives, an e-Cassandra that predicts the future without being able to change it has limited utility. AI research has often produced great ROC curves with impressive AUC statistics, but which don’t answer the core question - is this system actually capable of making people live better and longer?
Compounding this issue is the fact that AI models can only learn from data digitized at scale, and much of the information doctors use to diagnose patients (posture, skin color, facial expression) remains strictly analog. During my emergency medicine residency years ago, one of my teaching attending physicians set the expectation: “You have less than a second to look at a patient on an ambulance stretcher to decide if they are really sick or not.” Incorporating this data would require a massive dataset of patient images and video, linked with the rest of their health data. While not impossible, such an effort will take years, if not decades, to overcome the technical, cost, logistical and privacy hurdles entailed.
Why is Generative AI different? Fundamentally because GenAI does work. Rather than trying to create capacities that we are neither sure that we need nor how to use, it can unburden clinicians from the mountains of exhausting low-value scut work that drive burnout, cost, and poor quality care. Generative AI is a disruptive innovation - it excels at taking on low-value drudgery that no-one actually wants to do, with an order-of-magnitude cost and speed advantage. In a healthcare system where doctors spend more time looking at their screens than talking with their patients, and the the majority of Americans don’t receive basic preventative care, Generative AI is a perfect tool to pick up the slack and to allow us to deliver on the Triple Aim of Quality, Cost, and Experience.
It will take many years and billions of dollars to realize this transformation, and the political, regulatory, technical, data, and safety challenges are enormous (but not insurmountable). In this Substack series, I will explore the challenges and opportunities for Generative AI to transform healthcare, and the people and companies making them possible. I welcome your comments and questions, and please share with others you think might be interested!