Stroke can be tricky to diagnose as patients don’t always present with classic symptoms, and other conditions can mimic it. Researchers have used existing data to develop a machine-learning model that accurately predicts stroke and may make diagnosis easier.
Diagnostic errors present a major public health problem and contribute to preventable patient harm and health overspending. Preventable deaths from stroke due to diagnostic errors are 30 times more common than misdiagnosed heart attacks.
Stroke can be particularly difficult to diagnose, as its signs and symptoms can be mimicked by other conditions such as seizures, migraines, psychiatric disorders, and drug and alcohol intoxication. Additionally, strokes can present with atypical symptoms. Approximately 25% of stroke sufferers don’t present with the usual speech problems, facial droop, and limb weakness, further complicating a health practitioner’s ability to make an accurate diagnosis.
Researchers from Carnegie Mellon, Florida International and Santa Clara Universities developed an automated screening tool using machine learning technology to take some of the guesswork out of diagnosing stroke.
“Machine learning methods have been used to help detect stroke by interpreting detailed data such as clinical notes and diagnostic imaging results,” said Rema Padman, corresponding author of the study. “But such information may not be readily available when patients are initially triaged in hospital emergency departments, especially in rural and underserved communities.”
To develop their stroke prediction algorithm, the researchers used more than 143,000 individual patient records from admissions to Florida acute care hospitals between 2012 and 2014. They also incorporated data from the American Community Survey conducted by the US Census Bureau, which included demographics such as age, gender, race, and existing medical conditions.
The machine learning model predicted stroke with 84% accuracy. It was also highly sensitive, outperforming existing diagnostic models, which tend to miss up to 30% of strokes.
“Existing models’ moderate sensitivity raises concerns that they miss a substantial percentage of people with stroke,” said Min Chen, lead author of the study. “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.”
The study’s findings suggest that this machine learning model can accurately predict the likelihood that a person has had, or is having, a stroke before obtaining confirmation through diagnostic imaging or laboratory tests.
“Because our model doesn’t require clinical notes or diagnostic test results, it might be particularly useful in addressing the misdiagnosis challenges when dealing with walk-in patients with stroke with milder and atypical symptoms,” said Xuan Tan, co-author of the study. “It could also be useful in low-volume or non-stroke centers’ emergency departments, where providers have limited daily exposure to stroke, and in rural areas with limited availability of sensitive diagnostic tools.”
But the researchers point out that their algorithm is not intended to be a stand-alone model; it should be used in conjunction with existing models of stroke diagnosis.
The researchers recommend that their stroke prediction algorithm be incorporated into an automated, computer-assisted screening tool accessible at the time of admission to hospital.
The study was published in the Journal of Medical Internet Research.
Source: Carnegie Mellon University
