Research Projects

Machine learning for improved hospital efficiency and patient flow

To address unprecedented hospital bed and staffing shortages, Mayo Clinic has implemented several novel processes to ensure that people arriving in our emergency departments are matched with the best combination of inpatient and outpatient care. Our team has been studying the use of machine learning to optimize hospital resource distribution and facilitate timely emergency department disposition. Accordingly, we have developed machine learning models and used them in practice. Through this work, our team aims to ensure that people who come to a Mayo Clinic emergency department are matched with the best inpatient and outpatient resources for their needs, as efficiently as possible. We are currently assessing the impact of one machine learning tool in practice through a randomized controlled trial.

Collaborating faculty: Alexander J. Ryu, Ph.D., Shant Ayanian, M.D., Ray Qian, M.D., Heather A. Heaton, M.D., M.S., Riddhi S. Parikh, M.B.B.S., Sagar Dugani, M.D., Ph.D., Derick D. Jones, M.D., M.B.A., Daniel Chiang, M.D.

Relevant publications:
Ryu AJ, Romero-Brufau S, Quian R, Heaton HA, Nestler DM, Ayanian S, Kingsley TC. Assessing the generalizability of a clinical machine learning model across multiple emergency departments. Mayo Clinic Proceedings. 2022; doi:10.1016/j.mayocpiqo.2022.03.003.

Ryu AJ, Ayanian S, Quian R, Core MA, Heaton HA, Lamb MW, Parikh RS, Boyum JP, Garza EL, Condon JL, Peters SG. A clinician's guide to running custom machine-learning models in an electronic health record environment. Mayo Clinic Proceedings. 2023; doi:10.1016/j.mayocp.2022.11.019.

Frameworks for the assessment and mitigation of bias in artificial intelligence

As the adoption of machine learning increases in healthcare, our group believes that it will be important to understand whether these tools are performing equitably for people of all backgrounds. Our team is exploring appropriate metrics for assessing the equitability of machine learning model performance with the goal of developing methods for mitigating any bias that is detected. We also seek to apply these insights to machine learning models we develop and deploy.

Collaborating faculty: Gabriel O. Demuth, Ph.D., Curtis B. Storlie, Ph.D., Alexander J. Ryu, Ph.D., Shant Ayanian, M.D., Ray Qian, M.D., Sean R. Legler, M.D., Thomas C. Kingsley, M.D., M.P.H

Machine learning to understand genome-phenotype associations

The goal of this project is to combine existing genomic databases with clinical, text and imaging data to infer associations with therapeutics, diagnostics and outcomes. The project will help streamline the use of specific diagnostic or therapeutic approaches to improve outcomes. It personalizes medicine and helps push the limits of multimodal data analytics in the omics era. A preliminary study predicting the most appropriate initial treatment in rheumatoid arthritis is currently underway.

Collaborating faculty: Elena Myasoedova, M.D., Ph.D., Thomas C. Kingsley, M.D., M.P.H., William A. Faubion Jr., M.D., Guangchao Sun, Ph.D., Shant Ayanian, M.D.

Natural language processing for medical record summarization

A large proportion of critical information in medical records exists in text formats. Reviewing this text in anticipation of providing clinical care can be exceedingly time-consuming. In this work, we explore methods for incorporating natural language processing into clinical workflows to improve the efficiency and effectiveness of chart review. Through this work, we seek to equip clinicians with tools that enable thorough and efficient chart review and improve decision-making and enable clinicians to spend more time with the people in their care.

Collaborating faculty: Alexander J. Ryu, Ph.D., Kevin J. Peterson

Predicting delirium in the hospital

The study focuses on people who are hospitalized and aims to predict those who will develop delirium in-house. The model, which has been studied and peer-reviewed, will be implemented in EPIC and help streamline predictability and downstream workflow. Early prediction of delirium allows for early intervention and possible prevention of the condition in people who are hospitalized.

Collaborating faculty: Sandeep Pagali, Alexander Ryu, Shant Ayanian