Development of a Machine Learning Algorithm for Differential Diagnosis Between Primary Immune Thrombocytopenia and Connective Tissue Disease‐Related Thrombocytopenia in Pediatric Patients

Vollständiger Abstract

Worum geht es in dieser Arbeit?

ABSTRACT Background To develop a machine learning model for early differentiation of primary immune thrombocytopenia (pITP) from connective tissue disease‐related thrombocytopenia (CTD‐TP) in children presenting with thrombocytopenia. Method A retrospective study was conducted on 387 newly diagnosed children with thrombocytopenia. All patients were clinically diagnosed and divided into a training set and a test set in a 7:3 ratio. Six machine learning algorithms, including XGboost, RF, SVM, LR, GBDT, BPNN, were used to establish differential diagnostic models for pITP and CTD‐TP. The model performance was evaluated by accuracy, precision, recall, F1 score, and area under the receiver operating characteristic curve (AUC). Explain the importance and contribution of model feature variables through the Shapley Additive Explanations (SHAP) method. Result A total of 387 patients were included in the study, including 347 cases of pITP and 40 cases of CTD‐TP. Based on the above evaluation indicators, the BPNN model had the best performance, with an F1 score of 0.95, an accuracy rate of 94.87%, and an AUC of 0.9738. According to the SHAP value, the top 10 influencing factors were selected as age, erythrocyte sedimentation rate, antinuclear antibodies, IgM, Complement C3, T + B + NK%, B cells, body weight, complement C4, lymphocyte count. Conclusion The successful construction of the pITP and CTD‐TP identification models shows that BPNN has the best performance and can provide clinical strategies for early disease identification based on the importance of its variables.