Abstract 4146712: Prediction of Thrombosis in Coronary Artery Aneurysms Following Kawasaki Disease Using a Deep Learning Survival Approach

Background: Prediction of thrombosis of coronary artery (CA) aneurysms in children with Kawasaki disease (KD) is challenging both because risk is determined by the interplay of complex anatomical, physiological and pharmacological factors and because of the rarity of events. Methods: We used data from 483 patients with at least one giant CA aneurysm (z>10) and enrolled in the International Kawasaki Disease Registry to train and evaluate the performance of a deep learning survival (DeepSurv) algorithm predicting CA thrombosis over time. Outcomes were predicted separately for each CA branch. Each patient’s total follow-up duration was divided into epochs; new epochs were created with any new echocardiogram, change in thromboprophylaxis (either antiplatelet or anticoagulation) or outcome. Data (absolute aneurysm size, z-score and architecture) from the most recent echocardiogram were used at the start of each new epoch. Censoring was done on the last day of the epoch or date of last follow-up for the last epoch. Three-fold cross-validation was used to estimate model performance. SHAP analysis was used to calculate variable importance. Results: Average duration of follow-up was 4.2 years per patient; 87 (18%) patients had thrombosis during the follow-up period. The prediction model for thrombosis identified a high-risk group of epochs (28%). The high-risk epochs had a substantially greater incidence rate (5.1; 95%CI: 3.8-6.7) vs. 0.5 (95%CI: 0.4-0.7) events/100 patient-years, p<0.001) and rate ratio (9.8; 95%CI: 6.3-15.4, p<0.001). SHAP analysis identified critical features in predicting thrombosis, including larger aneurysm size, irregular aneurysm shape, recent change in thromboprophylaxis and, for those events diagnosed in the acute phase of the disease, immunosuppression and degree of inflammation. Events missed by the prediction algorithm were most often associated with rapid changes in CA aneurysm size during the acute phase of KD. Conclusion: A deep learning survival approach can be used to combine multiple dimensions of risk into a comprehensive, time-dependent, prediction model for thrombosis. This model identified a group of situations for which patients may be at substantially higher risk of developing a clot and where enhanced thromboprophylaxis should be considered. Particular attention should be paid to patients with large, irregular aneurysms, especially in hyperinflammatory contexts and when changing thromboprophylaxis regimen.