177 - DEVELOPMENT AND DEPLOYMENT OF MISKD: A CLINICAL DECISION SUPPORT SYSTEM FOR THE DIAGNOSIS AND MANAGEMENT OF PATIENTS ALONG THE KD TO MISC CLINICAL SPECTRUM USING AI-POWERED PREDICTIVE ALGORITHMS

Background: In the absence of a definitive diagnosis tool for KD and with substantial clinical overlap with other conditions including COVID-associated MISC; there are limited tools available to personalize the management of these patients.

Methods: Deep phenotyping was performed for 944 patients with KD, MISC or acute COVID infection without either KD/MISC recruited in 18 sites in 5 countries. Machine learning algorithms, both unsupervised (gaussian mixture model) and supervised (XGBoost) were used to train 12 prediction models from over 400 clinical features. These models are used in the diagnosis, treatment and management of patients along this clinical spectrum. External validation (or cross-validation for the treatment algorithms) was performed on the remainder of patients enrolled in the International Kawasaki Disease Registry (N=2,000) to evaluate the performance of the prediction models. Prediction models were integrated in a cloud-based clinical decision support system named MISKD which is integrated but independent of electronic health records.

Results: The diagnosis algorithm identified 5 phenotypic clusters in this patient population, 2 with a KD-like profile, 2 with a MISC-like profile and 1 with a profile consistent with neither condition (COVID without MIS-C/KD). Clinical misclassification of patients was common (Figure) with 8.6% of patients labelled as MISC having a phenotypic profile consistent with KD and 10.6% vice versa. In the validation cohort, the diagnostic algorithm for MISC and KD had AUCs of 0.85 and 0.88, respectively, with 85.5% of patients classifiable with >80% certainty. Algorithms (N=4) to predict response to IVIG and methylprednisone, both for primary (AUC 0.87 and 0.78 for IVIG and methylprednisone, respectively) and rescue therapy (AUC 0.91 and 0.94, respectively), showed excellent performance. Excellent performance (AUC 0.92) was shown in an algorithm to predict a refractory response to IVIG (no measurable change in fever during IVIG therapy). Algorithms for ICU admission, respiratory failure, shock, and reduced ejection fraction had AUCs of 0.88, 0.96, 0.84 and 0.74, respectively. Model performance was well balanced between patients with a KD vs. MISC phenotypic profile. All 12 models were integrated in a prototype of MISKD and successfully implemented for all 944 patients. A functional version of the clinical decision support system is currently under development.

Conclusions: A clinical decision support system incorporating multiple machine learning algorithms can be used to diagnose, personalize treatment choice and provide risk-based management for patients with KD. Informatic tools such as MISKD will be necessary to bridge the implementation gap for clinical prediction models.