Echocardiography and Artificial Intelligence in the Cardiac Amyloidosis Referral Pathway
A. P. Akerman¹, W. Hawkes¹, J. A. Slivnick², G. Woodward¹, C. G. Scott³, M. S. Maurer⁴, S. Cuddy⁵, J. B. Strom⁶, J. O’Driscoll⁷, R. M. Lang², P. A. Pellikka³, R. Upton¹
¹ Ultromics Ltd, Oxford, United Kingdom, ² University of Chicago, Chicago, IL, ³ Mayo Clinic, Rochester, MN, ⁴ Columbia University, New York, NY, ⁵ Brigham and Women’s Hospital, Boston, MA, ⁶ Beth Israel Deaconess Medical Centre, Boston, MA, ⁷ University of Leicester, Leicester, United Kingdom
Purpose
Echocardiography is critical in identification of patients at risk of cardiac amyloidosis (CA). Low cost and high accessibility ensure that patients can be screened in high volumes, and appropriately referred for confirmatory testing, monitoring, and treatment. However, specific data and clear recommendations on how exactly such tools could integrate into clinical practice, and the impact for patient management is required. This study therefore aimed to aimed to examine how echocardiography and artificial intelligence (AI) may be utilized in the CA diagnostic pathway.
Methods
Retrospective, multi-site data comprising 4255 patients without CA, and 560 patients with CA was collected for external validation of an AI model for screening patients for CA1 (AI- CA; EchoGo Amyloidosis, Ultromics Ltd). Using this data, clinical decision making was modelled (decision curve analysis) under two real-world scenarios; (1) patients considered high risk for CA (e.g., heart failure) have an echocardiogram to assess who should have a more in-depth “second read” by a clinician; and (2) patients who have already been pre-screened for key selection criteria (older, heart failure, structural remodelling) have an echocardiogram to assess who should be referred for confirmatory CA testing (PYP). Modelling compared the clinical utility of making decisions to refer for a second read or confirmatory testing based on an echocardiographic red flag for CA (increased wall thickness) or an AI-indicated presence of heart failure with preserved ejection fraction2 (AI-HF; EchoGo Heart Failure, Ultromics Ltd), compared with an AI indication of CA, or the AI-CA in combination with increased wall thickness or AI-HF.
Results
Patient characteristics are highlighted in Table 1. The modelled clinical decision-making
scenarios are presented in Figure 1. Standardized net benefit represents the proportion
of patients with CA who would be correctly referred, and net reduction in interventions
represents the number of incorrect referrals avoided without missing a patient with CA. Threshold probability (x-axis), represents the preference of the clinician to be more
concerned with missing the disease (lower threshold) than the risk of the decision (higher
threshold). Using wall thickness alone to decide which patients should be sent for a
second read or confirmatory testing would result in up to 65% correct referrals, whereas
incorporating AI-CA information to wall thickness or AI-HF results in up to 77% correct
referrals, similar to decisions based on AI-CA alone (~80%; panel A and B). Similarly,
compared with referring for a second read or confirmatory testing based on wall thickness
alone, unnecessary referrals are reduced by up to 18% and 14% (respectively) when AI-
CA information is combined with either wall thickness or AI-HF.
Conclusion
Integrating AI tools into clinical decision-making demonstrated increased clinical utility compared with using only traditional echocardiographic red flags for CA, resulting in more patients being correctly managed in the CA diagnostic pathway. This strategy of combining sources of information has the potential to increase the number of patients with CA being detected and reduce the number of incorrect referrals for follow-up assessment.