Geriatrician warns AI mortality models require context, not just accuracy

Artificial intelligence models are rapidly becoming fixtures in clinical workflows, predicting everything from sepsis risk to patient mortality directly within electronic health records. But a geriatrician who has built several such models is urging colleagues to look beyond performance metrics and consider how predictions actually get used in practice.

James Deardorff, an assistant professor in the division of geriatrics at the University of California San Francisco, has developed AI models that predict outcomes for older adults, including mortality and nursing home placement needs. His message: technical accuracy alone doesn't guarantee good patient outcomes.

Why it matters

As health systems embed AI predictions into routine care, the gap between a model's statistical performance and its real-world impact is becoming a critical concern. For older adults especially, how clinicians interpret and act on algorithmic outputs can mean the difference between a helpful conversation and a life-altering decision made on incomplete grounds. Understanding this distinction is essential as AI tools proliferate across medicine.

The context problem

Deardorff's concerns center on how predictions get translated into action. In a commentary published this month responding to a large analysis of Epic's proprietary end-of-life prediction model in JAMA Network Open, he and a co-author illustrated the stakes with concrete examples.

If a one-year mortality prediction prompts an open-ended discussion about a patient's care goals, the potential downsides are minimal. But when the same prediction informs higher-stakes decisions—such as transplant eligibility or resource allocation—the consequences become profound, even when the model performs well statistically.

The distinction matters particularly in geriatrics, where Deardorff emphasizes clinicians need awareness of two things: how an algorithm performs across different patient subgroups, including older adults specifically, and how to responsibly use its output in clinical decision-making.

Embedded and easy to trust

The integration of AI models directly into electronic health records makes them convenient to use but also easy to accept without scrutiny. Predictions appear alongside other clinical data, creating an illusion of equivalence that can obscure important limitations.

Deardorff has firsthand experience with this tension. Having developed multiple prediction models for older patient populations, he understands both their potential to help patients maintain independence and the risks when their outputs are misapplied or misunderstood.

The challenge extends beyond any single algorithm. As predictive models become standard tools across medical specialties, the medical community needs frameworks for evaluating not just whether a model works, but whether its use in specific clinical contexts serves patients well.

The path forward

For clinicians working with older adults, Deardorff's guidance suggests a more deliberate approach: understand the model's performance in relevant patient populations, recognize the difference between prediction and prescription, and match the use of algorithmic output to decisions where its limitations are acceptable.

The details were first reported by Katie Palmer at STAT.