FIU Business researcher developed an algorithm that can diagnose a stroke with 83 percent accuracy.

Researchers from FIU Business found that a machine learning (ML) algorithm that uses hospital data and social determinants of health data can help diagnose a stroke quickly – before the results of laboratory tests or diagnostic images are available – with 83 percent accuracy.

The ML algorithm that helps better diagnose strokes was developed using data of suspected stroke patients including social determinants of health (SDoH) - non-medical factors such as age, race, income and housing stability, access to transportation, and social isolation - that affect a wide range of health outcomes.

The algorithm “learns” and improves the more data it analyzes.

“Our algorithm can incorporate a lot of variables to analyze and interpret complex patterns, which will allow emergency department care teams to make better and faster decisions,” said Min Chen, associate professor of information systems and business analytics at FIU Business and one of the researchers. “This data-driven approach will help make it easier to identify stroke patients and ensure that they get the care they need in a timely manner.”

Stroke is among the most common and dangerous misdiagnosed medical conditions, currently used pre-hospital stroke scales miss approximately 30 percent of cases. Patients who are treated within an hour of the onset of symptoms have a greater chance of surviving and avoiding long-term brain damage. Data indicates that Blacks, Hispanics, women, older adults on Medicare and residents of rural areas are less likely to be diagnosed during this crucial window.

Published in the January 2023 issue of the Journal of Medical Internet Research, the study outlines how researchers developed the ML stroke prediction algorithm using the emergency department and hospitalization records from hospitals in the state of Florida from 2012 to 2014, merged with the SDoH data from the American Community Survey.

Their analysis included 143,203 hospital visits of unique patients. Patients who ended up being diagnosed with stroke tended to be older, have more chronic conditions, and have Medicare as the primary payer.

If a hospital is using the researchers’ ML algorithm, when a patient arrives with stroke or stroke-like symptoms, an automated, computer-assisted screening tool will quickly analyze all the patient’s information. If it predicts that the patient is at a high risk for stroke, a pop-up will be triggered to alert the emergency department team.

“We used an advanced method to explain the inner working of the complex stroke prediction algorithm, determine which factors are more important and make it actionable,” said Chen.

This study fills a critical gap in the current efforts to support stroke triage. Current ML methods developed to assist in detecting a stroke have focused on interpreting clinical notes and diagnostic imaging results, which may not be available when a patient arrives at a hospital, particularly in rural and underserved communities.

This type of technology is undergoing pilot testing in the ERs of several prominent healthcare systems. To facilitate the widespread implementation of this kind of ML-enabled decision support systems, it is important to have widespread health information exchange that enables patients' data to flow seamlessly across providers and also a standardized way to collect the granular SDoH data.

Chen conducted the study with Professors Xuan Tan of Clara University and Rema Padman of Carnegie Mellon University.