Geisinger Runner-Up in National AI Health Outcomes Challenge Geisinger has been named runner-up out of more than 300 entries in the Centers for Medicare & Medicaid Services (CMS) Artificial Intelligence Health Outcomes Challenge. Geisinger partnered with Medial EarlySign, a leader in machine learning-based solutions to aid in early detection and prevention of high-burden diseases, to use artificial intelligence (AI) to predict unplanned hospital admissions, readmissions occurring soon after hospital discharge, healthcare-associated complications, and mortality. The two entities collaborated to develop models that predict the risk of these outcomes using Medicare administrative claims data and created novel visualizations to explain the results in a clinician-friendly manner, a key component of AI implementation. “We are honored to be recognized as a national leader in using artificial intelligence to improve health outcomes,” said David Vawdrey, Geisinger’s chief data informatics officer. “The opportunity to participate in the CMS competition has significantly broadened our capabilities to design and implement predictive models, which will ultimately help prevent unnecessary hospitalizations and complications and reduce healthcare costs.” Geisinger and EarlySign’s shared vision of innovation and their collective focus on patient-centered care garnered recognition by CMS for “consistent strong performance across all competition elements while generating the best prediction accuracy results.” Their ability to successfully communicate predictions to clinicians, known as AI explainability, was a key factor in their selection as runner-up. “This achievement demonstrates the synergistic relationship Geisinger and EarlySign have in the journey to provide better care for patients,” said Ori Geva, co-founder and chief executive officer of Medial EarlySign. “This recognition is another validation that successful clinical AI solutions require deep understanding of clinical workflow, and expertise in clinical machine learning and clinical data.” The CMS AI Health Outcomes Challenge launched in 2019 with more than 300 entities proposing AI solutions for predicting patient health outcomes. Submissions aimed to forecast a variety of outcomes, including unplanned admissions related to heart failure, pneumonia, chronic obstructive pulmonary disease, and various other high-risk conditions; and adverse events such as hospital-acquired infections, sepsis, and respiratory failure. Geisinger was chosen as one of seven finalists in November 2020. To select the winner and runner-up, CMS conducted a rigorous evaluation process, supported by a team of AI scientists. Clinicians from the American Academy of Family Physicians, a CMS partner in the AI Challenge, reviewed and scored the models’ explainability. Submissions were reviewed and winners selected by a panel of CMS senior leadership. For more information on Geisinger’s work with artificial intelligence and machine learning, visit geisinger.org/innovation-steele-institute/innovative-partners/ai-and-deep-learning-lab.
Researchers Find AI Can Predict New Atrial Fibrillation, Stroke Risk A team of scientists from Geisinger and Tempus have found that artificial intelligence can predict risk of new atrial fibrillation (AF) and AF-related stroke. Atrial fibrillation is the most common cardiac arrhythmia and is associated with numerous health risks, including stroke and death. The study, published in Circulation, used electrical signals from the heart—measured from a 12-lead electrocardiogram (ECG)—to identify patients who are likely to develop AF, including those at risk for AF-related stroke. “Each year, over 300 million ECGs are performed in the U.S. to identify cardiac abnormalities within an episode of care. However, these tests cannot generally detect future potential for negative events like atrial fibrillation or stroke,” said Joel Dudley, chief scientific officer at Tempus. “This critical work stems from our major investments in cardiology to generate algorithms that make existing cardiology tests, such as ECGs, smarter and capable of predicting future clinical events. Our goal is to enable clinicians to act earlier in the course of disease.” To develop their model, the team of data scientists and medical researchers used 1.6 million ECGs from 430,000 patients over 35 years of patient care at Geisinger. These data were used to train a deep neural network—a specialized class of artificial intelligence—to predict, among patients without a previous history of AF, who would develop AF within 12 months. The neural network performance exceeded that of current clinical models for predicting AF risk. Furthermore, 62% of patients without known AF who experienced an AF-related stroke within three years were identified as high risk by the model before the stroke occurred. “Not only can we now predict who is at risk of developing atrial fibrillation, but this work shows that the high risk prediction precedes many AF-related strokes,” said Brandon Fornwalt, M.D., Ph.D., co-senior author and chair of Geisinger’s Department of Translational Data Science and Informatics. “With that kind of information, we can change the way these patients are screened and treated, potentially preventing such severe outcomes. This is huge for patients.” Geisinger and Tempus continue to work together to advance precision medicine using practical applications of artificial intelligence. Funding for this project was provided by Geisinger Clinic and Tempus. Geisinger has an exciting research environment with more than 50 full-time research faculty and more than 30 clinician scientists. Areas of expertise include precision health, genomics, informatics, data science, implementation science, outcomes research, health services research, bioethics and clinical trials.
