P1652: MACHINE-LEARNING-BASED MORTALITY PREDICTION OF ICH IN ADULTS WITH ITP: A NATIONWIDE REPRESENTATIVE MULTICENTRE STUDY

• DOI: 10.1097/01.hs9.0000849464.21167.c9
• Publication Date: 2022-06-24T08:02:19Z
• Authors: X.-H. Zhang; S. Chong; P. Zhao; R.-B. Huang; H. Zhou; J.-N. Zhang; Q.-S. Huang; H.-X. Fu; Q. Jiang; H. Jiang; J. Lu; X.-J. Huang
• Keywords: AdaBoost; Gradient boosting; Boosting

Background: ITP is more likely to be persistent or chronic in adults, whereas it a self-limiting disease children. Although mortality patient with only slightly higher than the general population, severe bleeding events such as intracranial haemorrhage (ICH) are often considered associated poor prognosis. Bleeding one of important clinical outcomes patients ITP. According previous studies, ICH occurs approximately 1% patients, rate 24.0%-31.2%, and fatal complication (Blood, 2009; Platelets, 2021) However, large-scale studies on risk factors still lacking. Aims: To identify 30-day for adults. develop validate machine learning model predict mortality. establish an application prediction. Methods: A national real-world study adult was conducted using data from 27 centres. The characteristics these were summarized. In addition, we identified by least absolute shrinkage selection operator (Lasso) regression training cohort 16 We developed 10 models algorithms including support vector (SVM), k-nearest neighbour (kNN), logistic regression, linear discriminant analysis (LDA), decision tree, random forest, gradient boosted tree (GBDT), adaptive boosting (AdaBoost), extreme (XGBoost), light (LGBM). then evaluated performance metrics receiver operating characteristic (ROC) area under curve (AUC), accuracy, sensitivity/recall, specificity, positive predictive value (PPV)/precision, negative (NPV), F1 score 10-fold cross-validation independent external test 11 other geographically separate selected best-performing further established adults Results: 33.8% 142 Ninety (65.7%) had platelet count /L less at time ICH. parenchyma most commonly affected region (N=75, 58.14%). Intraparenchymal haemorrhage, ICH, coexistence infection, prior events, absence head trauma potential death, adjusted parameter (λ) 1 standard error minimum criterion: 0.076 lasso regression. internal validation, SVM exhibited optimal AUC 0.879 ± 0.145 0.748 0.183. also sensitivity 0.600 0.667. Therefore, corresponding (47.94.162.105:8080/ich/) users specific process depicted Figure 1. Image:Summary/Conclusion: rare life-threatening event patients. algorithm validated its prediction