Project Summary

Electronic medical records are now near universally used in hospitals and outpatient practices in the United States. This presents an unrivalled opportunity to leverage their data for clinical research that can improve patient care, help develop new treatments and determine which of the established treatments are more helpful to patients.

However, a number of barriers to effective use of electronic medical records data for research and patient care remain. This study will address two of these barriers: (1) analysis of longitudinal (over time) patient data and (2) utilization of data “locked” in narrative documents, such as provider notes.

Longitudinal data that is accumulated for the same patient over time (e.g., multiple blood sugar measurements) presents a unique opportunity to improve understanding of the dynamic of the patient’s condition: is it getting better, deteriorating or remaining stable? However, most commonly used analytical methods cannot take advantage of this important data aspect. This study will test whether novel techniques—neural networks or deep learning—can improve our ability to analyze changes in a patient’s measurements over time.

One potential downside of neural networks is that it is not always clear how they arrived at their results, a so-called “black box.” To avoid this, the study will implement an explainable artificial intelligence approach that will present neural network-derived results in a way that is clear to clinicians and their patients, facilitating trust and acceptance of the scientific findings.

The research team will build and evaluate the above technologies on the example of diabetes mellitus, a condition that affects millions of Americans, leads to multiple complications and carries costs of more than $350 billion annually. Specifically, the team will use these technologies to identify patients with diabetes at high risk for treatment failure in both out- and inpatient settings: (a) failure to sufficiently lower blood sugar over 12 months and (b) readmission to the hospital within 30 days of being discharged after the first hospitalization. Being able to identify patients at high risk for these adverse events early on will make it possible to devote additional resources (e.g., phone calls from a pharmacist or a nurse visit at home after hospital discharge) to help prevent these treatment failures and improve patients’ clinical outcomes.

The study team will develop these analytical technologies using electronic medical record data from Mass General Brigham, a large hospital and outpatient practice network in Eastern Massachusetts. They will then evaluate how well the technology performs using electronic medical record data from Johns Hopkins Health System in Baltimore, Maryland, to ensure that it works well on data from different organizations. The team will also compare these technologies to a technology currently used for this purpose, called logistic regression, to make sure they represent an improvement.

The team is a mix of patients, caregivers, health care professionals and researchers. The project’s goal is to work together to place better tools in the hands of everyone who needs to analyze electronic medical record data to improve patient care and the outcomes patients care about.

*Methods to Support Innovative Research on AI and Large Language Models Supplement
This study received supplemental funding to build on existing PCORI-funded comparative clinical effectiveness research (CER) methods studies to improve understanding of emerging innovations in large language models (LLMs).

Project Information

Alexander Turchin, M.D., MS
Partners Healthcare Brigham and Women's Hospital

Key Dates

36 months
March 2023


State State The state where the project originates, or where the primary institution or organization is located. View Glossary
Last updated: April 12, 2024