A conversation with Ramin Zand, MD, Neurology Director of Clinical Stroke Operations, Northeastern Regional Stroke Director, Geisinger Health System, and Associate Professor of Neurology, University of Tennessee Health Science Center, and Vida Abedi, PhD, Research Scientist, Geisinger Health System, and Adjunct Professor, Virginia Tech, about using an artificial neural network to screen for stroke.

Interviewed by José G. Merino, MD, Associate Professor of Neurology, University of Maryland School of Medicine.

They will be discussing the paper “Novel Screening Tool for Stroke Using Artificial Neural Network,” published in the June issue of Stroke.

Dr. Merino: Could you please briefly summarize the key findings and put them in context of what was known before you did the study (i.e. an “elevator pitch” about your research)?

Vida Abedi, PhD

Vida Abedi, PhD

Drs. Zand and Abedi: We have developed ​a new computational method based on artificial intelligence to screen for the stroke in an emergency setting. Previous studies have shown that up to 25% of strokes can be initially misdiagnosed in the emergency department. The failure to recognize stroke in the emergency department is a missed opportunity for intervention. The goal of our study was to test if a supervised learning method could recognize and differentiate stroke from stroke mimics based on the patient demographics, risk factors, and certain clinical elements. Our results showed that in 6 out of the 10 data sets, the precision of our tool for the diagnosis of stroke was >90%. We believe that these methods can serve as a clinical decision support system and assist the emergency providers with early recognition of stroke.

Ramin Zand, MD

Ramin Zand, MD

Dr. Merino: Please explain in simple terms what a neural network is in the setting of your study? How does the new tool you propose (and the use of a network) differ from tools developed using standard regression methods (such as clinical decision rules)?

Drs. Zand and Abedi: Methods based on the artificial neural network can find patterns that are hidden in the data. Neural networks work in the same way as the human brain works. These methods are designed to find non-linear (non-obvious) relationships between various elements in the medical records. For example, even though in some cases it will be difficult to say why an individual is at a higher risk for a certain condition, these methods can still very reliably predict risks and identify patients prone to different diseases.

In recent years, prediction models using the artificial neural network and multivariable logistic regression analysis have been developed in health care research. The predictive ability of the artificial neural network model is comparable to that of the logistic regression model. However, in many cases, the artificial neural network models can outperform logistic models in both senses of discrimination and calibration. Neural networks can, in principle, model nonlinearities automatically. This must be explicitly modeled in linear regression. One of the caveats of artificial neural network is that it is prone to overlearning when the size of hidden neurons is too large and underlearning when the number of hidden neurons is too small. Neural networks can also be prone to overfitting when the training set is unbalanced and under/overrepresents certain features of the data.

Dr. Merino: How does the tool perform when compared with a neurologist in the ED? Are there implications for clinical practice now, or is there need for further validation? Can readers access the tool in the form of an online calculator?

Drs. Zand and Abedi: When we compare our results with a study in Houston, Texas (Houston Paramedic and Emergency Stroke Treatment and Outcomes Study) where the accuracy of trained paramedic prehospital diagnosis of stroke was 82%, we can say that artificial intelligence might be more precise in screening for stroke. Our analysis indicated that the average precision of neural networks for the diagnosis of stroke and stroke mimic was 85.2% (95% CI, 77.4–90.8) and 81.1% (95% CI, 73.4–87.1), respectively.

We believe, once these methods improve even further, they can be a great clinical decision support tool and assist with early recognition of stroke.

Dr. Merino: What is the next step in your line of research?

​ Drs. Zand and Abedi: The next step is to validate this method with a larger cohort, compare it to other machine learning methods from artificial intelligence, integrate it into the hospital electronic health record system and perform a full evaluation cycle.

Dr. Merino: Will there be a role for artificial learning to help clinicians make decisions about treatment in acute stroke?

​ Drs. Zand and Abedi: Of course! Physicians, including neurologists and neurointensivists, are often overwhelmed by the amount of clinical data. It can be very challenging to monitor, integrate and interpret all of the clinical and demographic data and integrate all the data elements in the real-time treatment decisions. Computers can assist health care providers in monitoring and integrating data. Artificial intelligence can help physicians to provide more individualized care and make more evidence-based medical decisions.