Reader response
Myron Yaster MD
From Hugo Clifford M.D., Assistant Professor of Anesthesiology and Perioperative Medicine, Division of Pediatric Cardiac Anesthesiology, University of Rochester School of Medicine and Dentistry
I read with interest today’s PAAD on artificial intelligence in perioperative care here. I fully agree with your assertion that “the safest near-term role for AI is for the ‘A’ in ‘AI’ to stand for ‘augmented’ rather than ‘artificial’ or ‘automated’”. While machine algorithms may eventually provide unceasing vigilance, improved response times, closer hemodynamic control, and a host of other benefits, humans have been caring for the sick since before we were humans at all. While our scientific understanding has improved and the tools at our disposal have revolutionized techniques and outcomes, from prehistory the crux of quality care has been a recognition of our shared humanity with our patients. We must continue to advance our field, and capitalize on the advantages provided by AI, without losing our key role in understanding and steering the care being given. Because while a machine can be programmed to try and keep a child alive, it cannot understand the cost of failure.
From Sarah Rebstock, MS, MD, PhD, FAAP, Basic and advanced certified by the American Board of AI in Medicine
As an innovation director and having been in the innovation space since 2011. I have been keeping myself updated on where innovation in medicine is going. This is an incredibly timely PAAD. I just finished basic and advanced certification from the American Board of AI in Medicine.
I think Augmented is the important key word. Physicians are still responsible for decisions that are made by AI, and I am not sure this will ever change, making it incredibly important to understand and being able to discuss and explain medicine and the safe and ethical use of medical data for AI in medicine. The amount of medical data is doubling every 3-6 months, with providers generating around 137 terabytes of data on a daily basis, most of the data being unstructured and difficult to manage and use in a constructive way. Being able to understand what data engineers, data scientists do and be able to translate what we do into how they manage and structure AI programming for the data will direct the future of medicine.
We have had a couple of AI winters due to many historical environmental factors (internet of things, computer power, computer storage, and expectation management), so AI was slower to be adopted. However, this is no longer the case. AI is expanding exponentially into our lives. The current direction of AI and its ability to outperform humans in some areas is not going away, and in fact is becoming an imperative to be able for physicians to spend more quality time with the patients. We will not be replaced by AI machines, but we will be replaced by those who can work with AI and accept and manage its trajectory for medicine. AI is becoming medico-legally recognized, having cases against some radiologists who did not use AI in decision making. I think AI will eventually fundamentally change medicine, academia, and what we understand and accept as evidence-based medicine.
The use of AI should come in steps that we monitor and understand with ethics and patient safety in mind. We must be training physicians at all levels of their career about AI, data, ethics, and its use in medicine for the greater good of medicine. We should not be using AI and not understanding what we are doing, who owns the data we are putting into the AI models, and understanding the consequences. There are very wise voices in business stating that what we program AI to do should also be risk stratified, and if the consequences have catastrophic potential, then we cannot be allowing AI to run without governance, and intimate knowledge of what it is doing. AI is only as good as the programming of the model and the quality of data put into the model. Data scientists and engineers should be working side by side with physicians who understand their branch of medicine. Ergo the importance of augmented. Physicians must be intimately involved with AI data management, programming, and implementation going forward.
Courses for AI are available from the ABAIM. I have no financial interests or stock in any part of the ABAIM.
From Eric Jackson, MD, MBA, Disclosure: I am a paid consultant to Masimo Corporation and Chair of its Health Equity Advisory Board, and I formerly served as the company’s SVP and Chief Medical Officer. Masimo competes in the perioperative monitoring market. I am also a former Chief Innovation Officer at Nemours Children’s Health. I have no financial relationship with any manufacturer of the hypotension prediction technology discussed below.
Thank you for putting Han et al. article in front of the PAAD readership. This one lands close to home. I have worked these questions from both sides of the table: on AI governance at Nemours Children’s Health, and at Masimo, where I have argued the vendor’s case. The view is not the same from each chair, and that is most of my point.
First, the economics. The decisive question for a department is not whether a model predicts, but whether it pays. The prize is real: because complications are expensive and labor is our largest line item, a tool that truly prevents an event can improve outcomes, reduce workload, and lower cost at once. Economists call that dominance. The harder and more common case is a tool that costs more and works better. That turns on an incremental cost per quality-adjusted life-year judged against a stated threshold, paired with a budget-impact analysis, because cost-effective and affordable are not the same finding. Prediction alone establishes none of this. Intraoperative hypotension prediction is our most instructive case, being the one with real trial data. One randomized trial reduced the time-weighted average of MAP below 65 mmHg1; a larger pilot trial found no difference at all (0.14 vs. 0.14 mmHg)2; and a 2026 meta-analysis of 14 randomized trials (n = 2,030) found no reduction in acute kidney injury (RR 0.87, 95% CI 0.71–1.07), mortality, or length of stay3. Independent analyses suggest a simple MAP alert at 70 to 75 mmHg delivers nearly the same warning time4,5. A tool that moves a surrogate but not an outcome, at proprietary cost and with an added alert burden, has earned none of it even when it “works.” I have sat on the selling side of that distinction, and it is an uncomfortable one.
Second, governance. The figure you cite from Arina et al. deserves to be the headline: of 103 perioperative machine-learning studies, only 13% were externally validated and 90% carried high or unclear risk of bias6. Be precise, too, about what regulatory authorization means. Roughly 96% of FDA-authorized AI devices are 510(k) cleared, a finding of substantial equivalence to a predicate rather than an independent demonstration of benefit8; of 521 authorizations reviewed in 2024, 43% had no published clinical validation and 22 had been tested in a randomized trial9. Cleared is not approved, and neither is proven. Before any purchase, a department should require a named physician owner, a pre-specified performance-drift threshold, and an agreed sunset provision if the model fails to meet it. Algorithmovigilance without an accountable owner and a line-item budget is a slogan, not a safeguard.
Third, pediatrics. Antel et al. identified 40 studies, 60% of them risk-factor prediction, and noted the near-absence of external validation7. Children differ in physiologic reserve, weight-based dosing, and case mix, and the events that matter most are rare enough that samples stay thin. Adult models should be presumed non-transferable until demonstrated otherwise. That arithmetic favors children: a catastrophic event prevented in a four-year-old returns far more life-years than the same event prevented in an adult, so the binding constraint in pediatrics is not price but whether the effect is real. That is an argument for closing the evidence gap, not for tolerating it.
None of this argues against adoption. It argues for holding software to the standard we already hold any new device.
References
1. Wijnberge M, et al. JAMA. 2020;323(11):1052–1060.
2. Maheshwari K, et al. Anesthesiology. 2020;133(6):1214–1222.
3. Wang SS, et al. A&A Practice. 2026;20(4):e02180.
4. Enevoldsen J, Vistisen ST. Anesthesiology. 2022;137(3):283–289.
5. Rellum SR, et al. Eur J Anaesthesiol. 2025;42(6):527–535.
6. Arina P, et al. Anesthesiology. 2024;140(1):85–101.
7. Antel R, et al. BJA Open. 2023;5:100125.
8. Loganathan G, et al. J Med Artif Intell. 2025.
9. Chouffani El Fassi S, et al. Nat Med. 2024;30(10).
From Jenni Majumdar, PhD, CRNA, Assistant Professor, Hunter College, Nurse Scientist, Memorial Sloan Kettering Cancer Center, Writer, Art of Anesthesia, Editor-in-Chief, Journal of Nurse Anesthesia Education
The result that looks like a failure on paper is the one I'd point people to: voluntary reporting went up. That number only moves when staff believe a report will land on a process instead of on a person. Anesthesia has its own version of this in the mid-1980s decision to make pulse oximetry and capnography mandatory in every case — it worked because it removed the judgment call rather than asking people to be more careful. And the ARCC behavior that actually gets tested in my room is unglamorous: someone leaning on an arm that's already positioned, or reaching for the bed remote on a patient who can't object. How that gets received on the third case of the day tells you more about the culture than the stand-down does.