Model using machine learning techniques and patient data at hospital arrival predicts stroke more accurately

Stroke is one of the most dangerous and commonly misdiagnosed medical conditions. Black and Hispanic people, women, Medicare seniors, and people in rural areas are less likely to be diagnosed in time for treatment to work. Using methods of learning and data available when patients were admitted to the hospital, we developed a model that predicts stroke more accurately than current models.

The study, by researchers at Carnegie Mellon University (CMU), Florida International University (FIU), and Santa Clara University (SCU), Journal of Medical Internet Research.

Diagnosis errors are a major public health problem, with preventable stroke deaths from such errors occurring more than 30 times more frequently than deaths from myocardial infarction. Diagnosing stroke is difficult because there are many conditions that mimic stroke, including seizures, migraines, and alcoholism. These issues can lead to delays and exacerbate health problems.

Automated screening tools that analyze available data and suggest stroke diagnoses have great potential to help address this problem. Scientists are using artificial intelligence and machine learning to identify hidden insights in large amounts of data and generate new patient predictions.

“Machine learning techniques have been used to aid in stroke detection by interpreting detailed data such as clinical records and imaging findings,” said co-author of the study, Heinz College, CMU. Rema Padman, Director of Management Science and Health Informatics, explains. “However, when patients are first triaged in hospital emergency departments, such information may not be readily available, especially in rural and underserved areas.”

Padman and her colleagues sought to develop a stroke prediction algorithm based on widely available data during patient admissions. They also assessed the added value of social determinants of health (SDoH) in predicting stroke. These include the circumstances in which people are born, grow up, live and age, and the factors that contribute to these circumstances.

Their study looked at more than 143,000 patient visits seen at acute care hospitals in Florida between 2012 and 2014. Researchers also looked at her SDoH data from the US Census’ American Community Survey. Their model includes data on basic demographics (age, gender, race, ethnicity), number of chronic conditions, and primary payers (Medicare, Medicaid, or Individuals), as well as the number of health care providers and payers on admission. It has built-in variables that you collect periodically. insurance).

The researchers’ model was accurate (84% predictive accuracy for stroke), sensitive, and outperformed existing scales (which tend to miss up to 30% of strokes). This model could be used to predict the likelihood of a patient’s condition to be stroke at hospital presentation based on the patient’s demographics and social determinants of health available at admission, prior to obtaining an imaging study. Suggested. Or laboratory test results, says the authors.

“The moderate sensitivity of existing models raises concerns that they miss a significant proportion of stroke patients,” said co-author of the study, associate professor of information systems and business analytics in the FIU School of Management. Min Chen explains.

“In hospitals where medical resources and clinical staff are scarce, our algorithms will complement current models to help quickly prioritize patients for appropriate intervention.”

“Because our model does not require clinical records or diagnostic test results, it may be particularly useful in addressing the challenge of misdiagnosis when working with walk-in stroke patients with mild and atypical presentations.” is an analyst at the Leavey School of Business and another co-author of the study.

“Emergency departments in small or non-stroke centers where health care providers have limited daily exposure to stroke, and rural areas with limited access to sensitive diagnostic tools, may also be useful.”

Among the limitations of this study, the authors noted that because their study was retrospective, confirmation of stroke cases relied on the International Classification of Diseases Code and did not require review of patient records. They further caution that their algorithm should not be viewed as the gold standard for stroke diagnosis, but rather as a model to complement existing stroke scoring systems used in hospitals. Finally, their findings are limited by the social determinants of health variables available in administrative data.