Use Healthcare AI for Patients First: Califf
FDA commissioner Robert Califf says he is concerned that the main use of artificial intelligence (AI) algorithms in healthcare will be “decisions that optimize the bottom line rather than optimizing the longevity and well-being of patients.” Speaking 3/5 to the Coalition for Health AI, of which FDA is a member, Califf said that focusing AI on decisions that improve corporate earnings “is counter to the mission of FDA, where effectiveness means an improvement in a health outcome.”
Referring to the fact that the U.S. lags behind other high-income countries in health outcomes despite spending $4.5 trillion per year on healthcare, Califf said the combination of the primacy of finance over clinical outcomes and the optimization of those finances in a balkanized manner creates a problem that may explain our poor health status amid financial success. “The compelling need and advantage to sharing data and automating access to analysis could mean that AI could disrupt these unhealthy attributes,” he said.
“Imagine an AI system that is connected by a data infrastructure that enables a continuous learning loop between current practice, decisions, and outcomes,” Califf concluded. “At its best, AI could radically change our understanding of what practices, organizational and clinical, would lead to the best outcomes for patients and clinicians, and then the payment system could be aligned to reward best practices and outcomes. Broad education and collaboration will be needed to ensure that these challenges are met.”
Califf said he is concerned that health systems don’t have the infrastructure and tools to make the most important determinations about whether an AI application is “effective” for health outcomes. He said to know whether an algorithm of any kind is truly effective for health, one needs two conditions to be supported with a functional infrastructure:
- monitoring the algorithm and testing its operating characteristics over time; and
- following-up on the population to whom the algorithm is applied, at least in a valid sample, so the monitoring of the algorithm is based on a valid inference for its use in a particular population or clinical circumstance.