Researchers in a study used machine-learning methods and patient data upon hospital arrival to develop a model that predicts strokes with greater accuracy than current models.
The study, published earlier this year in the Journal of Medical Internet Research, could help address diagnostic errors and delayed stroke diagnoses that can be fatal, according to the investigators.
Stroke is among the most dangerous and commonly misdiagnosed medical conditions. Black and Hispanic people, women, Medicare patients, and rural residents are all less likely to be diagnosed in time for effective treatment. Numerous conditions also mimic stroke, including seizures, migraines, and alcohol intoxication.
Preventable stroke deaths due to diagnostic errors occur about 30 times more often than myocardial infarction deaths. An automated screening tool to analyze available data and suggest a stroke diagnosis may help address this problem.
Artificial intelligence and machine learning can identify hidden insights from large volumes of data and generate predictions. Machine-learning methods have previously been used to help detect stroke by interpreting clinical notes and imaging results. But such information may not be readily available when patients are initially triaged in hospital emergency departments, especially in rural and underserved communities.
The study researchers sought to develop a stroke prediction algorithm based on data available upon patient admission. Social determinants of health (SDoH) in predicting strokes, including the conditions people are born into, grow up in, reside in, and age in, were also assessed. They examined more than 143,000 individual patient visits to Florida acute care hospitals from 2012 to 2014, along with SDoH data from the U.S. Census’s American Community Survey. Variables routinely collected by healthcare providers upon hospital entry, including age, gender, race, ethnicity, chronic conditions, and primary insurance provider, were also incorporated.
Their model showed 84% precision in predicting strokes and higher sensitivity than existing scales, which miss up to 30% of strokes. Therefore, it may potentially predict the likelihood of stroke at the time of hospital presentation, prior to obtaining diagnostic imaging or laboratory test results.
The researchers said their model might be particularly useful in addressing the potential misdiagnoses of walk-in stroke patients with milder and/or atypical symptoms in low-volume emergency departments where providers have limited daily exposure to stroke, and in rural areas with limited access to sensitive diagnostic tools or incomplete data-gathering capabilities.
Since their study was retrospective, confirming stroke cases relied on International Classification of Diseases codes rather than patient records. The researchers recommend using their algorithm as a model, rather than a standard, to complement existing stroke scoring systems used in hospitals.
“Existing models’ moderate sensitivity raises concerns that they miss a substantial percentage of people with stroke,” Min Chen, a co-author and Florida International University associate professor, said in a statement. “In hospitals with a shortage of medical resources and clinical staff, our algorithm can supplement current models to help quickly prioritize patients for appropriate intervention.”