AI in the CICU: From Theaters to Monitors

The medical artificial intelligence revolution is coming and it will not be televised. It might show up on your new virtual reality headset, but even if it doesn’t, it will nevertheless arrive in your cardiac intensive care unit unabated. Medical AI is poised to reshape the majority of the care provided in cardiac intensive care units, but the pace of that change and, importantly, the smoothness of that transition remain unknowns.

With the explosion of AI in recent press coverage and in the public zeitgeist you would be forgiven for thinking that AI is a shiny, new marvel of the 21st century. But it is important to recognize that medical AI is old, decades old. The automatic interpretation on the printout of an ECG and even plain vanilla multivariable logistic regression are both examples of machine learning, a subgroup of AI. I point this out because not all AI is the same. If a medical AI company salesman with an expensive watch on his wrist pitches your critical care division on a slew of fancy, cutting-edge AI products, it will be worth your time to peer under the hood of the expensive car you’re considering buying. If you find that the engine is running on multivariable logistic regression, that model may still be valuable, but don’t pay V8 engine prices for it. You might even consider building it yourself inhouse.

If medical AI is older than most of our careers, the reason we are hearing so much about it now is because medical AI has recently improved and expanded exponentially. Advances in computing power and the development of neural networks, including the transformers used in large language models like ChatGPT, have combined to create algorithms with capabilities that far exceed what was possible even a few years ago. As those advances leave the lab and enter the implementation and commercialization phases, their impact will become increasingly apparent.

While these advances are both promising and inspiring, the media has overhyped our expectations for currently available AI applications. The trough of disillusionment will be rough. Elon Musk’s serial overpromises about the capability of his self-driving cars is an apt cautionary tale. While the community continues to make incremental but important progress towards a safe, true self-driving future, promises that we would have completely autonomous vehicles by 2017 are laughable if not misleading in retrospect. Medical AI applications in the ICU will likely face similar challenges, because when lives are on the line, every mistake counts.

I for one, though, am playing the long game, and I look forward to the breakthroughs we’ll see on the slope of enlightenment. Those breakthroughs will hopefully decrease the drudgery and allow for more time spent on the parts of care that humans are good at and which they like doing. AI excels at data collection, curation, and analysis, streamlining decision-making processes. It can process almost any kind of data, from medications documented in the MAR to extremely granular data like waveform signals (e.g., arterial blood pressure, telemetry). AI algorithms will be at the vanguard of precision medicine, by for example, tailoring medication regimens based on pharmacogenomic drug interactions. At the same time, it can leverage population or cohort level data, highlighting when a repaired Tetralogy of Fallot patient starts to deviate from the recovery trend of his digital twins. And it does all this without sleeping, without tiring, and without a drop in performance.

However, realizing this bright future isn’t guaranteed. Artificial intelligence, like device and drug development, is an expensive business. The GPUs in high performance computing clusters cost millions of dollars alone, and they consume a small country’s worth of electricity when running at full throttle. While there are many business cases to train very large models for medical applications in adults, the cost of training such models on pediatric disease is harder to justify for a single organization to take on. For the health of all acutely ill children, it will be imperative for children’s hospitals to pool their resources, sharing both data and expertise, to get pediatric specific algorithms into our ICUs. Only then will we be able to bring the newest advances to the children that need them most.