Prediction of chronicity in pediatric immune thrombocytopenia based on developed and internally validated machine learning models.

BACKGROUND: Primary immune thrombocytopenia (ITP) typically remits spontaneously, 20%-30% of children progress to chronic ITP. Early prediction of chronicity in ITP may facilitate in timely intervention and improve prognosis. METHODS: In our current study, we collected clinical features such as platelet-specific antibodies and used machine learning (ML) to predict chronic progression of ITP. ML models were applied to data from our hospital. Model performance was evaluated using accuracy, precision, sensitivity, specificity, F1 score, and area under the receiver operating characteristic (ROC) curve to assess the binary classification models performance. RESULTS: A total of 156 patients were enrolled for ITP classification prediction model construction, including 68 chronic patients and 88 non-chronic patients. All models performed well, with AUC values ranging from 0.833 to 0.864. The ET model was selected for predictive model construction due to its highest AUROC score and interpretability. Then, the ET model identified occult disease course, age, and platelet-specific antibodies as significant predictors. The absence of an occult disease course decreased the probability of chronic ITP, while older age increased it. It's worth noting when platelet-specific antibodies are negative, patients are less likely to develop chronic ITP. CONCLUSIONS: In conclusion, the ET model achieved high prediction accuracy for ITP chronicity by using clinical parameters, especially platelet-specific antibodies. Despite the limited sample size, this study suggests occult disease course, age, and platelet-specific antibodies are early predictors of chronic ITP in children, meriting confirmed in future larger cohort multicenter study.