Integrating AI Pipelines Into The Diagnosis Of Cardiomyopathies

Ashley P. Akerman¹, Will Hawkes¹, Jeremy Slivnick², Jorge Oliveira¹, Gary Woodward¹, Izhan Hamza³, Viral K. Desai³, Martha Grogan³, Christopher G. Scott³, Halley N. Davison³, Juan Cotella², Matthew Maurer⁴, Stephen Helmke⁴, Marielle Scherrer-Crosbie⁵, Marwa Soltani⁵, Akash Goyal⁶, Karolina M. Zareba⁶, Richard Cheng⁷, James N. Kirkpatrick⁷, Tetsuji Kitano⁸, Masaaki Takeuchi⁸, Viviane Tiemi Hotta⁹, Marcelo Luiz Campos Vieira⁹, Pablo Elissamburu¹⁰, Ricardo E. Ronderos¹⁰, Aldo Prado¹¹, Efstratios Koutroumpakis¹², Anita Deswal¹², Amit Pursnani¹³, Nitasha Sarswat², Amit M. Patel¹⁴, Karima Addetia², Frederick L. Ruberg¹⁵, Michael Randazzo², Federico M. Asch¹⁶, Sarah Cuddy¹⁷, Roberto M. Lang², Patricia A. Pellikka³, Ross Upton¹

¹Ultromics, Ltd., Oxford, UK; ²University of Chicago, Chicago, IL; ³Mayo Clinic, Rochester, MN; ⁴Columbia University, New York, NY; ⁵University of Pennsylvania, Philadelphia, PA; ⁶Ohio State University, Columbus, OH; ⁷University of Washington, Seattle, WA; ⁸Hospital of University of Occupational and Environmental Health, Kitakyushu, Japan; ⁹Heart Institute (InCor), Sao Paolo, Brazil; 10ICBA, Buenos Aires, Argentina; ¹¹Centro Privado de Cardiología, Tucuman, Argentina; ¹²University of Texas MD Anderson Cancer Center, Houston, Texas; ¹³NorthShore, Evanston, IL; ¹⁴University of Virginia Medical Center, Charlottesville, VA; ¹⁵Boston University Chobanian & Avedisian School of Medicine, Boston, MA; ¹⁶MedStar Health Research Institute, Washington, DC;
¹⁷Brigham & Women’s Hospital

See full-poster

Background

The differentiation between heart failure with preserved ejection fration (HFpEF) and cardiac amyloidosis (CA) is critical for correct clinical management. Artificial intelligence (AI) has demonstrated promise in aiding clinicians to make complex decisions, but understanding how such models might be deployed in the real-world is important.

Methods

Two artificial intelligence models have been developed to aid in the screening for HFpEF (EchoGo Heart Failure) and CA (EchoGo Amyloidosis). Three-dimensional convolutional neural networks were independently trained to identify HFpEF1 and CA2 on apical four chamber videos from transthoracic echocardiograms. Both models return a classification of positive or negative for disease, and EchoGo Heart Failure is supplemented with a continuous estimate of disease probability. Patients from an external validation study, collected from 17 clinical sites, was used to test the implementation of the two models. The output of the two AI models separately and combined, was tested to understand how they might be implemented in the real world for maximum utility. Model performance (classification) is presented under different implementation scenarios.

Results

From a larger external validation dataset, 1270 patients with HFpEF or cardiac amyloidosis were identified (Table 1). The two models demonstrated high accuracy in the populations they were developed (Table 2). The two models were utilized together to examine how funneling of potential HFpEF patients into cardiac amyloidosis screening might be used (Figure 1 and Table 2). In a workflow where a positive output by either model would initiate further diagnostic testing for cardiac amyloidosis, 91.5% (88.4, 94.0%) of cardiac amyloidosis patients would be correctly identified (sensitivity). More sophisticated workflows incorporating rule-in and rule-out probability thresholds for entry into the funnel were examined (Table 2).

Discussion and Conclusion

Two recently developed models can be utilized independently and in combination to assist in detection of otherwise complex clinical diagnoses. Ongoing prospective and retrospective research will identify the optimal approach to provide maximum clinical utility.