Surgical AI adapts to changing patients

New adaptive model improves surgical predictions across evolving patient populations

Beth Miller 03.03.2026

The COVID-19 pandemic changed many things about our lives, including computer models. Chenyang Lu’s lab found that their model performed better to predict the outcome of patients who had pancreatic surgery before the COVID-19 pandemic than after, leading them to change the model. (Credit: iStock photo)

Facebook Twitter Linkedin Email

Washington University in St. Louis researchers and clinicians have been incorporating data from Fitbit wristbands into machine learning models that could predict surgical outcomes, pain after surgery and potential mental health issues, among other uses. While working with clinicians to predict pancreatic surgery outcomes, researchers at Washington University in St. Louis encountered an unexpected factor that changed their prediction model: the COVID-19 pandemic.

The pandemic disrupted daily life, from school and work to medical procedures, including pancreatic surgery, which can be complex with risks for complications and a challenging recovery. While it may be difficult for clinicians to determine who might be a good candidate for such surgery and who has a good prognosis for recovery, they may get some insight from using a model that analyzes patient data from their electronic health record as well as from a Fitbit wristband, a previous study published in 2021 revealed.

In a follow-up study, Jingwen Zhang, who earned a doctorate in computer science from the McKelvey School of Engineering in 2024 in the lab of Chenyang Lu, the Fullgraf Professor in the Department of Computer Science & Engineering and director of the AI for Health Institute, found that their model performed better to predict the outcome of patients who had pancreatic surgery before the COVID-19 pandemic than after, leading Zhang to develop a novel solution in collaboration with a multidisciplinary team of AI researchers and surgeons. Results of the research were published online in ACM Transactions on Computing for Healthcare Jan. 19, 2026.

“We realized patients were not the same before the pandemic and after, leading to cohort variability,” Zhang said. “For example, the rate of readmission and severe complications dropped in patients who had surgery after the pandemic.”

To tackle this challenge, Zhang created a new model called Adaptive Mixture of Experts (AdaMoE). It blends data from Fitbit devices, such as sleep, heart rate and activity levels, with clinical information such as blood tests and medical history, automatically adjusting how much each source matters for each patient.

“The relative predictive value of traditional clinical data and wearable signals varies substantially across patients,” Zhang said. “For some individuals, clinical variables are more informative, while for others, wearable data provide stronger predictive signals. We showed that by dynamically integrating these two data sources at the patient level, we can maintain predictive performance even as patient populations change.”

Lu said model performance can deteriorate after deployment in real-world health care systems, a growing concern as AI becomes more embedded in clinical practice.

“In practice, we typically design and train a model using historical data and then deploy it in a hospital system,” Lu said. “At first, it performs well. But over time, performance can decline as conditions change. Patient characteristics evolve, clinical workflows shift and broader societal and environmental factors influence care. COVID provided a powerful stress test, revealing how dramatically patient data can change in a short period of time.

“But the implications of our work extend far beyond the pandemic,” Lu continued. “We demonstrate that AI models can be designed to adapt to changing patient populations, an important step toward trustworthy, long-term deployment of AI in health care.”

Zhang J, Wang R, Xu Z, Liu H, Rodriguez J, Cos H, Srivastava R, Raper L, Sanford D, Hammill C, Lu C. Addressing cohort variability with adaptive fusion of wearable and clinical data: A case study in predicting pancreatic surgery outcomes. ACM Transactions in Computing for Healthcare. Published online Jan. 19, 2026. DOI: 10.1145/3788681

This research was supported by funding from the Fullgraf Foundation, the Foundation for Barnes Jewish Hospital and the BJC Health Systems Innovation Lab.