Design of a prediction system for intravenous immunoglobulin resistance in hospitalized Kawasaki disease patients with artificial intelligence based on machine learning

Machine Learning (ML) algorithms are widely used as a prediction tool in medicine to support clinical decisions. In this study, the machine learning methods are used to investigate IVIG resistance in Kawasaki disease in a referral center of Iran. This is a retrospective cross-sectional study conducted among 1312 hospitalized patients with primary diagnosis of Kawasaki disease. Data were collected from clinical and laboratory records which contains about patients’ medical history, echocardiographic findings, and laboratory results. Independent variables which are statistically significant in the univariate logistic regression were included in training phase of machine learning models. Due to class imbalance between resistant and non-resistant groups, the SMOTE-NC (Synthetic Minority Oversampling Technique) method was applied to augment the resistant group along with a weighting technique to increase model sensitivity towards resistant cases. Three machine learning algorithms were trained: Logistic Regression, Support Vector Machine (SVM), and Multi-Layer Perceptron (MLP). All analyses were performed using Python (version 3.11) with the Scikit-learn, Pandas, and NumPy libraries. 494 patients were analyzed and divided into two groups, 117 patients were resistant and 377 were non-resistant to therapeutic administration. Clinical features such as duration of illness (Odds Ratio = 1.07, CI = 1.02–1.13), maximum temperature (Odds Ratio = 1.80, CI = 1.26–2.56) and non-purulent conjunctivitis (Odds Ratio = 1.87, CI = 1.01–3.47) were significantly associated with an increased likelihood of resistance. In terms of echocardiographic findings, a Z-Score calculation between 2.5 and 5 (Odds Ratio = 4.68, CI = 2.22–9.88) and a Z-Score ≥ 5 (Odds Ratio = 20.30, CI = 7.84–52.53) increased the likelihood of resistance compared to a Z-Score < 2. In laboratory findings, the WBC count (Odds Ratio = 1.05, CI = 1.01–1.10) increased the chance of resistance. Non-purulent conjunctivitis, prolonged fever, elevated white blood cell count, lower hemoglobin level, and higher coronary artery Z- scores were identified as predictive factors for treatment-resistant Kawasaki disease in Iranian children. Although, the model showed moderate predictive performance in internal cross-validation; external validation in independent Iranian cohorts is needed before clinical implementation. The machine learning model might demonstrate reliable predictive accuracy.