Few predicted in the mid-’90s that the deep learning revolution was coming.
But Geisinger, a healthcare system serving people across Pennsylvania and New Jersey, made an early bet on digital health records. And that gave it a major advantage when implementing AI two decades later.
Geisinger adopted electronic health records in 1996 and began digitally storing medical images around 2001. It’s saved digital records for nearly 2 million patients since then. And as a regional healthcare provider, Geisinger also has a fairly stable patient population.
This means it has longitudinal data for a large set of patients — a dataset ripe for deep learning, said Brandon Fornwalt, associate professor and founding chair at Geisinger’s Department of Imaging Science and Innovation.
“It allows us to do things other places can’t do,” he said.
During a talk Tuesday at the GPU Technology Conference, in San Jose, Fornwalt and Aalpen Patel, chair of radiology at Geisinger, shared how the healthcare system is using the NVIDIA DGX-1 within its clinical network to power multiple AI solutions that improve patient care.
“Machine learning is going to help us figure out what we can do earlier in a patient’s life to let patients live happier, longer and healthier lives,” Fornwalt said.
AI for the Head and the Heart
When a patient has an injury or a stroke that causes bleeding in the brain, rapid diagnosis and treatment are essential to minimize brain damage. A radiologist may take five minutes or less to read an individual patient’s CT scan. But these experts have prioritized worklists — inpatients and acute cases in the emergency room are typically read much faster than outpatients.
However, Patel and Fornwalt explained, there are always critical cases hiding within the outpatient scans. Deployed since January 2017, a deep learning model trained on NVIDIA GPUs reads all head CT scans at Geisinger and automatically re-prioritizes the radiology worklist within seconds, moving potential acute cases up on a radiologist’s list. This re-prioritization resulted in an astounding 96 percent reduction in time to diagnosis.
To improve the care of patients with heart disease, Geisinger is developing deep learning models to analyze data from electrocardiograms and echocardiograms, two of the most common tests to monitor the heart’s function. Harnessing 2 million EKG records from its databases, it’s building neural networks to predict future cardiac events in patients.
And for echocardiograms — ultrasounds of the heart — the team is training a neural network to find patterns that can predict specific health outcomes. Both the EKG and echocardiogram models are being developed and tested on a DGX-1, and Geisinger will soon be installing an NVIDIA DGX-2 to speed up its pipeline.
Geisinger is also using machine learning to optimize resource allocation for patients with heart failure. Its models found an association between heart failure patients who did not receive certain evidence-based therapies, such as an annual flu shot, and the resulting increased risk of death or hospitalization.
Teams of pharmacists and other healthcare providers at Geisinger are now using these insights to target patients who will benefit most from increased resources. Using the NVIDIA RAPIDS data science libraries to analyze nearly 1 million sets of data from heart failure patient encounters, the team saw a threefold speedup using a single GPU compared to a 52-core CPU server.
“When you have those big, tabular datasets with almost a million rows and let’s say 100 columns, that’s what RAPIDS is really helping with,” Fornwalt said.
Patel sees AI as a key tool to address the growing shortage of physicians, especially as the U.S. population grows older and life expectancy rises.
“To even allow us to take care of patients, machine learning is no longer a choice,” he said. “It’s going to be necessary for our survival.”
Main image show a radiologist reviewing CT scans from a patient with a brain injury. U.S. Air Force photo by Senior Airman Julianne Showalter, licensed under public domain.