Researchers from Emory University and collaborating institutions reported on Jun. 8 that a flaw exists in many peer-reviewed studies using reinforcement learning, an artificial intelligence method, as a theoretical guide for sepsis treatment. The findings were published in the journal npj Digital Medicine.
Shengpu Tang, assistant professor of computer science at Emory University, said that while AI is boosting positive outcomes in healthcare, deployment of such tools should proceed thoughtfully and at a measured pace. Tang and colleagues found that a commonly used technique for preprocessing and indexing data related to sepsis treatment can cause slight time misalignment. This leads the AI agent to sometimes use future events to predict past states, which may remain hidden if both training and testing data are misaligned the same way.
The researchers demonstrated through simulation experiments that flawed systems could recommend overtreatment or undertreatment in nearly half of patient states. "We found that the large majority of the papers that use reinforcement learning to analyze sepsis treatment over the last decade made this time-misalignment mistake—including our own work," Tang said.
Tang's team developed a simple workaround by shifting the action index backward by one time step to achieve correct temporal alignment. Their simulations showed that addressing this flaw resulted in an 8-10% decrease in patient mortality rates. "We hope this work serves as a wake-up call and a roadmap for building safer, more reliable reinforcement-learning models for the clinical bedside," Tang said.
The paper was co-authored by Sonali Parbhoo from Imperial College London, Jenna Wiens from the University of Michigan, and Jiayu Yao, formerly at Columbia University. The researchers noted concern that similar flaws may occur across other reinforcement-learning applications beyond healthcare settings.
