Geisinger doctors see 13% boost in all-cause mortality predictions with help from new AI tool
A new deep learning tool based on hundreds of thousands of heart ultrasound images can help physicians predict patients’ one-year all-cause mortality, according to research published Monday.
Data scientists and cardiology experts at Geisinger Health System trained their convolutional neural network on echocardiogram videos from upwards of 34,000 patients. The tool beat out a number of cardiologist-based predictions of mortality and a machine learning model constructed from hand-picked variables.
Co-senior author of the paper Chris Haggerty, PhD, an assistant professor in Geisinger’s Department of Translational Data Science and Informatics, said this study is one of the largest medical imaging datasets ever published.
"We were excited to find that machine learning can leverage unstructured datasets such as medical images and videos to improve on a wide range of clinical prediction models," Haggerty said in a statement.
The team explained that imaging has become one of the most “data-rich” components of the electronic health record, with a sole heart ultrasound producing nearly 3,000 individual images. The average cardiologist or radiologist, however, has little time to look over all these pictures and analyze additional clinical data.
With this in mind, the Danville, Pennsylvania-based team trained their model using 812,278 echocardiogram videos gathered at their institution over the past decade.
When put to the test, their model’s predictions outperformed a handful of other metrics, including the Seattle Heart Failure score, widely used pooled cohort equations, and an AI tool combining 58 manually chosen variables from echo images and 100 clinical EHR data points.
And proving the notion that AI can augment human intelligence, cardiologists using the model improved their one-year all-cause mortality predictions by about 13%. The model bolstered their sensitivity while maintaining specificity.
With larger datasets and further testing, the team hopes to bring its machine learning platform to real-world clinical settings.
"Our goal is to develop computer algorithms to improve patient care," Alvaro Ulloa Cerna, PhD, senior data scientist at Geisinger, added Monday. "In this case, we're excited that our algorithm was able to help cardiologists improve their predictions about patients since decisions about treatment and interventions are based on these types of clinical predictions."
Read the entire study published in Nature Biomedical Engineering here.
Matt joined Chicago’s TriMed team in 2018 covering all areas of health imaging after two years reporting on the hospital field. He holds a bachelor’s in English from UIC, and enjoys a good cup of coffee and an interesting documentary.