According to Forbes, European researchers built an AI “patient preference predictor” that analyzed 61 characteristics of patients aged 50+ and accurately predicted their end-of-life preferences 71% of the time. This outperformed both the average surrogate score in medical literature (68%) and the 59% accuracy among couples predicting their partner’s wishes. In a separate St. Louis initiative across 8 BJC HealthCare hospitals, sharing AI-generated death risk predictions with clinicians led to a surge in end-of-life planning and palliative care use, while the death rate within 30 days of discharge dropped by a third from expected levels. Both studies used sophisticated machine learning rather than chatbots, with researchers emphasizing human involvement throughout the process.
Better than humans?
That 71% accuracy number is both impressive and deeply unsettling. Think about it – an algorithm could potentially understand your end-of-life wishes better than your own spouse. The European study specifically looked at preferences for interventions like CPR, and the AI beat human partners by a significant margin. But here’s the thing: accuracy in predicting static survey responses is one thing, capturing the nuance of real human relationships and values is entirely another.
Preferences change over time in ways that can’t be predicted, as the commentary on the European research pointed out. What you want today might not be what you want when you’re actually facing that medical crisis. And let’s be honest – filling out a survey about hypothetical end-of-life scenarios is very different from making those decisions when they’re real.
The human safety net
What I find reassuring about both studies is how carefully they built in human oversight. The St. Louis approach was particularly thoughtful – AI alerts weren’t sent directly to doctors. Instead, another clinician reviewed them first, taking about 2-3 minutes per patient. They also trained over 300 clinicians on “goals of care discussions” and made sure there were enough palliative care specialists to handle increased referrals.
This is light years ahead of the 1990s APACHE III system, where stressed ICU patients and families were suddenly confronted with survival odds by clinicians with no special training. That earlier attempt actually didn’t improve end-of-life decisions overall, probably because timing and delivery matter just as much as the prediction itself.
The moral AI question
Now we’re getting into really interesting territory. A University of Washington researcher is proposing an AI surrogate that considers the “moral adequacy of representation” – basically, whether the AI honors the patient’s values, relationships, and cultural worldview. Muhammad Aurangzeb Ahmad is in the “conceptual phase” of testing this using Harborview patient data, though no patients have actually interacted with the model yet.
But can you really algorithmize morality? And what happens when patients start copying their medical records into chatbots and getting conflicting survival prognoses? We’re already seeing this in other areas of medicine – patients showing up with ChatGPT printouts challenging their doctor’s recommendations. In end-of-life care, that could get really messy really fast.
The real solution
Here’s the bottom line: the best technology in the world can’t replace actually having conversations with your loved ones about your wishes. As physician Tal Tova Patalon emphasized, we need to avoid dependence on AI by taking initiative in advance care planning.
The St. Louis results are promising because they used AI as a tool to start conversations, not make decisions. Death rates dropped because people had the opportunity to express preferences before crises hit, not because algorithms were pulling plugs. That’s the sweet spot – using technology to facilitate human connection and understanding, not replace it. Because when it comes to life’s final chapter, no algorithm should have the final say.
