180 - PREDICTING SHOCK AND LEFT VENTRICULAR DYSFUNCTION IN PATIENTS ADMITTED WITH KAWASAKI DISEASE AND MISC USING TRANSFER MACHINE LEARNING

Background: Patients hospitalized with acute KD are at risk of developing shock and left ventricular (LV) dysfunction, both of which are rare but can have serious consequences. Given the limited number of patients affected, creating accurate prediction models has been challenging. Since those complications are much more prevalent in patients with MIS-C, we hypothesized that a modified transfer machine learning process could help in this context.

Methods: We performed a deep phenotypic substudy through the International KD Registry (IKDR) which enrolled patients with acute KD and COVID-associated MIS-C hospitalized between 2020 and 2023. IKDR patients not enrolled in the substudy were used for external validation. We trained 6 machine learning algorithms (logistic regression, support vector, random forest, XGBoost, light GBM and neural network) to separately predict LV dysfunction (ejection fraction below 55%) and shock. A total of 93 clinical and laboratory features were available as potential predictors; feature selection was performed using a tree-based optimization algorithm. Model calibration was performed separately for KD and MIS-C. Model performance was evaluated using the area under the receiving operating curve (AUC) and model calibration expressed as the average absolute observed/expected (O/E) ratio over 10 deciles of risk in the validation cohort.

Results: A total of 929 patients were enrolled and 2,977 patients were available for the external validation cohort. Prevalence of LV dysfunction was lower for KD (7.5%) vs. MISC (40.2%) (p=0.001). Models’ AUC in the external validation cohort was between 0.76-0.78 for all methods with XGBoost having the best calibration for both KD (±3.6% O/E ratio) and MIS-C (±12.1% O/E ratio). Shock was less prevalent in KD (2.9%) vs. MIS-C (32.6%) (p=0.001). AUCs in the external validation cohort was similar between methods (0.83-0.84). XGBoost calibration showed an O/E ratio of ±1.4% for KD patients and ±5.9% for MIS-C. Lower platelets, alkaline phosphatase, and albumin; and higher d-dimer, ferritin and creatinine were associated with higher risk prediction in both models. Higher CRP was associated with increased risk of LV dysfunction while higher neutrophil/lymphocyte ratio was associated with increased risk of shock.

Conclusion: We demonstrated that training joint prediction models for KD and MIS-C patients allowed for the algorithm to accurately learn the relationship between features and risk, while separate calibration accounted for the difference in prevalence between groups. This study demonstrates that a modified transfer machine learning approach could be used to train prediction models in the context of rare diseases.