Establishment of bleeding risk assessment model for children with primary immune thrombocytopenia based on machine learning algorithms

Establishment of bleeding risk assessment model for children with primary immune thrombocytopenia based on machine learning algorithms | Authorea try { document.documentElement.classList.add('js'); } catch (e) { } var _gaq = _gaq || []; _gaq.push(['_setAccount', 'G-8VDV14Y67G']); _gaq.push(['_trackPageview']); (function() { var ga = document.createElement('script'); ga.type = 'text/javascript'; ga.async = true; ga.src = ('https:' == document.location.protocol ? 'https://ssl' : 'http://www') + '.google-analytics.com/ga.js'; var s = document.getElementsByTagName('script')[0]; s.parentNode.insertBefore(ga, s); })(); Skip to main content Preprints Collections Wiley Open Research IET Open Research Ecological Society of Japan All Collections About About Authorea FAQs Contact Us Quick Search anywhere Search for preprint articles, keywords, etc. Search Search ADVANCED SEARCH SCROLL This is a preprint and has not been peer reviewed. Data may be preliminary. 27 January 2025 V1 Latest version Share on Establishment of bleeding risk assessment model for children with primary immune thrombocytopenia based on machine learning algorithms Authors : Zhi-Cong Li , Shi-Qi Tong , Xin-Ya Jiang , Lu-Lu Xu , and Wei-Wei Tong [email protected] Authors Info & Affiliations https://doi.org/10.22541/au.173797928.84432957/v1 201 views 106 downloads Contents Abstract Information & Authors Metrics & Citations View Options References Figures Tables Media Share Abstract Background: Severe bleeding events caused by Primary immune thrombocytopenia (ITP) in children can be life-threatening or lead to long-term complications. This study aimed to establish a risk assessment model for hemorrhage and provide valuable insights for clinical diagnosis and treatment planning in these children. Procedure: This retrospective study reviewed 500 children with ITP, who were categorized into low-bleeding risk and high-bleeding risk groups. Collected data included disease characteristics and laboratory test, compared the differences of variables between groups. Cohort were randomly split at a 7:3 ratio for training set for model development and test set for model validation, employed six type of machine learning algorithms for bleeding risk assessment model construction and evaluated models based on the ROC curve. Finally employed alignment diagram to visually represent the risk assessment model derived from the optimal algorithm. Results: Age, fever, cytomegalovirus infection, neutrophil percentage, erythrocyte, platelet, activated partial thromboplastin time, aspartate aminotransferase, creatine kinase, creatine kinase MB isoenzyme, urea, creatinine and cystatin C were significant difference between low-bleeding risk and high-bleeding risk groups ( p <0.05). Bleeding risk assessment model include platelet, erythrocyte, creatine kinase, urea, age and cytomegalovirus infection variables presented the best performance among the six prediction models (AUC: 0.815) based on binary logistic regression algorithm. Conclusions: A bleeding risk assessment model for children with primary immune thrombocytopenia which variables include platelet, erythrocyte, etc, based on binary logistic regression algorithm is established and demonstrated the best performance among the evaluated machine learning algorithms. Establishment of bleeding risk assessment model for children with primary immune thrombocytopenia based on machine learning algorithms Zhi-Cong Li 1,2 , Shi-Qi Tong 3 , Xin-Ya Jiang 1 , Lu-Lu Xu 1 , Wei-Wei Tong 1 * 1, Department of Laboratory Medicine, Shengjing Hospital of China Medical University, Shenyang, 110004, P. R. China 2, Department of Clinical Laboratory, The Second Affiliated Hospital of Guangxi Medical University, Nanning, 530007, P. R. China 3, Clinical Medicine, The first Affiliated Hospital, Dalian Medical University, Dalian, 116000, P. R. China * Correspondence author: Tel: 86-24-96615-72122 Email: [email protected] Keywords: primary immune thrombocytopenia; machine learning algorithm; bleeding risk assessment model. Word Count for: Abstract: 241 Main Text: 3392 The number of Tables: 4 The number of Figures: 2 Short running title: “Establishment of bleeding risk model for children with ITP” Papers that have not been previously published as meeting abstracts. Abbreviations key Abbreviation Full term or phrase ALT alanine aminotransferase APTT activated partial thromboplastin time AST aspartate aminotransferase AUC area under ROC curve BN bayesian Network CK creatine kinase CK-MB creatine kinase MB isoenzyme DT decision tree ITP primary immune thrombocytopenia KNN k-nearest neighbor LR logistic regression NN neural network SVM support vector machine Abstract Background: Severe bleeding events caused by Primary immune thrombocytopenia (ITP) in children can be life-threatening or lead to long-term complications. This study aimed to establish a risk assessment model for hemorrhage and provide valuable insights for clinical diagnosis and treatment planning in these children. Procedure: This retrospective study reviewed 500 children with ITP, who were categorized into low-bleeding risk and high-bleeding risk groups. Collected data included disease characteristics and laboratory test, compared the differences of variables between groups. Cohort were randomly split at a 7:3 ratio for training set for model development and test set for model validation, employed six type of machine learning algorithms for bleeding risk assessment model construction and evaluated models based on the ROC curve. Finally employed alignment diagram to visually represent the risk assessment model derived from the optimal algorithm. Results: Age, fever, cytomegalovirus infection, neutrophil percentage, erythrocyte, platelet, activated partial thromboplastin time, aspartate aminotransferase, creatine kinase, creatine kinase MB isoenzyme, urea, creatinine and cystatin C were significant difference between low-bleeding risk and high-bleeding risk groups ( p <0.05). Bleeding risk assessment model include platelet, erythrocyte, creatine kinase, urea, age and cytomegalovirus infection variables presented the best performance among the six prediction models (AUC: 0.815) based on binary logistic regression algorithm. Conclusions: A bleeding risk assessment model for children with primary immune thrombocytopenia which variables include platelet, erythrocyte, etc, based on binary logistic regression algorithm is established and demonstrated the best performance among the evaluated machine learning algorithms. Introduction Primary immune thrombocytopenia (ITP) is a haematological autoimmune disease characterised by decreased platelet counts and increased risk of bleeding. ITP can occur at any age but exhibits peak incidence in children and older adults [1] . A prospective study in Norway reported a childhood ITP incidence of 5.3 per 100,000, highlighting its prevalence in pediatric settings [2] . It reported that ITP pathogenesis involves the production of autoantibodies that target platelets, leading to their destruction or impaired production, ultimately causing bleeding [3] . The primary clinical manifestation of childhood ITP is bleeding, it can range from skin and mucosal bleeding to more serious visceral bleeding. Typically, children with ITP do not present with hepatosplenomegaly or lymphadenectasis, no deformities in bones or joints. Non-specific symptoms like fatigue and exhaustion may also occur，and some children and their families may have a decreased quality of life due to the disease. While most children with ITP experience mild bleeding, a smaller proportion can develop severe bleeding events that can be life-threatening or cause lasting complications [4] . A systematic review reported cerebral hemorrhage risk of 0.4% and severe bleeding risk of 20.2% from other sites with ITP in children [5] . Although platelet counts below 20×10 9 /L are associated with an increased risk of severe bleeding, the risk is not entirely eliminated at higher platelet counts [5] . Other factors potentially influencing bleeding risk include patient age, illness duration, and lack of treatment [5-7] . Accurate bleeding assessment is vital for improving the management and treatment outcomes of ITP patients. While An Zhuoyu et al. developed a machine-learning model for predicting severe bleeding in adult ITP patients [8] , no such effective tool exists for children. Given the potential severity and lasting consequences of severe bleeding in children with ITP, establishing a risk assessment model is crucial. This model would allow for early prediction of severe bleeding risk, facilitating more informed clinical decisions regarding treatment or follow-up strategies. As the application of machine learning in medicine developed rapidly, particularly for analyzing disease risk factors and establishing predictive models, which has a significant impact on diagnosis, treatment and prognosis of the disease. Common machine learning algorithms used for predicting categorical patient outcomes include decision trees (DT), logistic regression (LR), Bayesian networks (BN), support vector machines (SVM), K-nearest neighbors (KNN), neural networks (NN). This study proposes utilizing these machine learning algorithms to establish a risk assessment model for severe bleeding in children with ITP. Such a model could provide valuable insights for clinicians, aiding in treatment decision-making. 2. Methods 2.1 Study Design This retrospective review involves a cohort of 500 ITP pediatric patients, categorized into high risk group and low risk group, wihch were treated within the Division of Pediatric Hematology at Shengjing Hospital, China Medical University between January 2020 and June 2023. Study population comprised all participants identified using institutional administrative data, with inclusion criteria based on billing codes for ITP. This study was conducted in accordance with established medical ethics standards and received approval from the Hospital Ethics Committee (Ethics Number: 2024PS153K). Given the retrospective nature of the study, informed consent was not required from participants. 2.2 Inclusion Criteria Patients of ages ranging from birth to 14 years, who have been diagnosed with ITP based on the Guideline for the Diagnosis and Treatment of ITP for Chinese Children [9] , and meanwhile without any other bleeding disorders or hormones, gamma globulin, thrombopoietin receptor agonists, rituximab treatment before admission. Classification of bleeding risk categories were defined five grades based on guideline. Grade 0 indicated no bleeding, while Grade 1 encompassed minor bleeding like a small number of petechiae (≤100) and (or) up to 5 small ecchymosis (≤3cm diameter) without mucosal bleeding, Grade 2 represented mild bleeding with more petechiae (>100) or larger ecchymoses (>5cm and >3cm diameter) but no mucosal bleeding, Grade 3 meaned moderate bleeding with obvious mucosal bleeding and adverse impact on children’s life, Grade 4 were severe mucosal bleeding accompanied by a hemoglobin drop exceeding 20 g/L or suspected visceral bleeding. Treatment regimen for pediatric patients experiencing Grades 0-2 bleeding with minimal life disruption were closely observed and followed up, while pediatric patients of Grades 3-4 required immediate treatment. So low bleeding risk group included Grade 0-2 and high bleeding risk group covered Grade 3 and Grade 4. 2.3 Data Collection Patient data encompassed demographics (gender, age), potential triggers, bleeding symptoms, medical history (including fever, cough, diarrhea, and other respiratory/digestive tract infections within 3 weeks prior to illness onset and vaccination history within 2 weeks), duration of hormone, gamma globulin, or thrombopoietin receptor agonist therapy after hospitalization, bleeding signs and symptoms before and during admission, corresponding examination results, and other relevant information. Additionally, all laboratory test results from the outpatient clinic or upon admission were collected, including complete blood count (white blood cell count, neutrophil and lymphocyte percentages, hemoglobin, erythrocyte and platelet counts, etc.), coagulation function test (prothrombin time, activated partial thromboplastin time, fibrinogen, thrombin clotting time, D-dimer), liver function test (albumin, aspartate aminotransferase, alanine aminotransferase, bilirubin), renal function test, cardiac enzymes, C-reactive protein, and etiologic tests. 2.4 Machine learning algorithms Cohort of 500 children with ITP were randomly split at a 7:3 ratio for training set (n=347; low bleeding risk group, n=271, high bleeding risk group, n=76) for model development and test set (n=153; low bleeding risk group, n=117, high bleeding risk group, n=36) for model validation. Difference variables analyzed with univariate analysis between the high and low-risk groups required further multicollinearity test to mitigate the potential influence of multicollinearity on the model’s performance, and the criteria for multicollinearity were tolerance threshold of less than 0.1 and variance inflation factor (VIF) exceeding 10 [10] . Establishment of risk assessment model for hemorrhage with primary immune thrombocytopenia in children based on machine learning algorithms Bleeding risk assessment models for children with ITP were established based on six machine learning algorithms, which included binary logistic regression model, decision tree algorithm model, bayesian network model, support vector machine model， K-nearest neighbor classification model and neural network model. Following model development, performance of each algorithm was evaluated using a confusion matrix for both the training and test sets. This allowed us to calculate accuracy, precision, recall, F1 score, and AUC for each model in both data sets. Additionally, ROC curves were generated to visually compare predictive abilities of models, and the model with the largest AUC in the test set was considered the optimal performer [11, 12] . Finally An alignment diagram was employed to visually represent the risk assessment model derived from the optimal algorithm. 2.5 Statistics Statistical analysis was performed using IBM SPSS Statistics 27 (IBM Corp., Armonk, NY). Categorical variables were compared using the chi-square test or Fisher’s exact test, continuous data with normal distribution were analyzed by the t-test and presented as mean ± standard deviation (x±SD). For non-normally distributed continuous data, the Mann-Whitney U test was employed, with results expressed as median (quartile range) [M(Q 1 , Q 3 )]. P <0.05 was considered statistically significant. Machine learning model development, ROC curve generation, and model validation were conducted using IBM SPSS Modeler 18.0 (IBM Corp., Armonk, NY). Alignment diagrams were created with R Studio (R Studio Team, 2020). 3. Results 3.1 Clinical and Laboratory data comparison This study enrolled total of 500 children with ITP, categorized into high-risk group (n=112) and low-risk group (n=388). The high-risk group exhibited a significantly lower proportion of fever symptoms than the low-risk group ( P <0.05). No significant differences were observed between the groups regarding gender, cough symptoms, diarrhea symptoms, or vaccination history prior to ITP onset. Cytomegalovirus (CMV) infection was significantly less frequent in the high-risk group compared to the low-risk group (P<0.05). In contrast, no significant differences were observed in the infection rates of other viruses or bacteria commonly associated with childhood illnesses, including Mycoplasma pneumoniae, herpes simplex virus, Epstein-Barr virus, parainfluenza virus, Human Parvovirus B19, Helicobacter pylori, and COVID-19. Comparison of categorical variables are presented in Table 1. Analysis of continuous variables revealed significant differences between the high and low-risk groups. Compared to the low-risk group, the high-risk group exhibited lower levels of hemoglobin, erythrocyte count, hematocrit, lymphocyte percentage, and platelet count ( P <0.05). Additionally, activated partial thromboplastin time (APTT) was significantly prolonged in the high-risk group ( P <0.05). Conversely, age and neutrophil percentage were significantly higher in the high-risk group ( P <0.05). No significant differences were observed in white blood cell count, mean corpuscular volume, mean hemoglobin content, mean hemoglobin concentration, red blood cell distribution width, prothrombin time, fibrinogen, thrombin coagulation time and D-dimer levels. As to liver and kidney function tests, children in the high-risk group exhibited lower levels of aspartate aminotransferase (AST), alanine aminotransferase (ALT), creatine kinase (CK), creatine kinase MB isoenzyme (CK-MB) and cystatin C ( P <0.05). Conversely, urea and creatinine levels were significantly higher in the high-risk group ( P <0.05). No significant differences were observed in albumin, Gamma-glutamyltransferase, total bilirubin, conjugated bilirubin and C-reactive protein levels. Comparison of continuous variables are presented in Table 2. Multicollinearity test was performed to address potential multicollinearity among independent variables with statistically significant differences in univariate analysis. Firstly, variables of neutrophil percentage and lymphocyte percentage were identified as having multicollinearity, additionally, multicollinearity was showed among variables of hemoglobin, erythrocyte count, and hematocrit and also beween variables of AST and ALT. 3.2 Establishment of bleeding risk model Following the exclusion of multicollinearity, 13 independent variables were identified as significantly different between the high and low-risk groups. These variables were included as initial inputs for all machine learning models: age, fever, neutrophil percentage, erythrocyte count, platelet count, APTT, AST, CK, CK-MB, urea, creatinine, cystatin C, and cytomegalovirus infection status. Backward stepwise regression within the Binary Logistic Regression framework identified the final model after eight iterations. This model included the following independent variables: age, CK, urea, cytomegalovirus infection status, platelet count, and erythrocyte count. All other independent variables were excluded due to a lack of statistical significance. The model demonstrated statistical significance (Omnibus test χ² = 75.669, P < 0.001). The maximum log-likelihood was 289.142. Goodness-of-fit statistics indicated an acceptable model fit, with a COX&Snell R² of 0.196, a Nagelkerke R² of 0.301, and a non-significant Hosmer-Lemeshow test ( P = 0.588). Detailed results are presented in Table 3. According to the results of logistic regression modeling, the equation for predicting the probability of high bleeding risk in children with ITP is shown below: The probability of high bleeding risk of ITP child was calculated by P=1/(1+e Y ), Y= - (-1.364 × CMV infection + 0.122 × Age -1.254 × Erythrocyte count - 0.057 × Platelet count - 0.007 × Creatine kinase + 0.17 × Urea + 3.387). The Decision Tree C5.0 algorithm was used to conduct further feature selection during model building. The final model primarily relied on APTT, neutrophil percentage, platelet count, CK, and CK-MB. The Bayesian Network Model ranked the importance of the included independent variables as follows (in descending order): AST, erythrocyte count, cytomegalovirus infection status, creatine kinase, platelet count, urea level, fever, age, cystatin C, APTT, CK-MB, neutrophil percentage, and creatinine. Feature importance analysis of the SVM model revealed the following ranking of independent variables: neutrophil percentage, age, platelet count, CK, erythrocyte count, APTT, CK-MB, urea, cystatin C, AST, cytomegalovirus infection, and fever. KNN Model emerged more effective predictive power when the optimal number of neighbors ( k ) was 3. Total of 13 independent variables were incorporated into the model, including age, fever status, neutrophil percentage, erythrocyte count, platelet count, APTT, AST, CK, CK-MB, urea, creatinine, cystatin C, and cytomegalovirus infection status. The Neural Network model employed in this study incorporated the following independent variables, ranked by their relative importance: erythrocyte count, urea, AST, platelet count, APTT, creatine kinase, neutrophil percentage, CK-MB, age, creatinine, cystatin C, cytomegalovirus infection status, and presence of fever. The model architecture consisted of an input layer, a single hidden layer with six neurons, and an output layer. 3.3 Validation of bleeding risk model Among the six machine learning models evaluated for predicting bleeding risk in the test set, the Binary Logistic Regression model achieved the highest accuracy and the largest AUC. Consequently, the Binary Logistic Regression model was selected as the final model for predicting bleeding risk in pediatric ITP. Table 4 summarizes the performance metrics (accuracy, precision, recall, F1-score, and AUC) of all six models on the validation set. The ROC curves for each model’s performance in the test set are visually compared in Figure 1. 3.4 Alignment diagram of bleeding risk model The logistic regression model was used to generate an alignment diagram, the results of which are presented in Figure 2. This application method alignment diagram functioned as follows: examination results for children with ITP were plotted on the corresponding item lines in the column line graph. The upper point for each item’s score was then identified, and all item scores were summed to obtain a Total Point. Finally, the corresponding scores for the identified risks were displayed as the probability of children having a high risk of hemorrhage. 4. Discussion ITP is a common hemorrhagic disease in children characterized by thrombocytopenia. The risk of bleeding in patients with ITP is a critical prognostic factor, severe bleeding events can be life-threatening or lead to long-term complications. According to grade of bleeding risk of ITP, different treatment schemes should be adopted to avoid the adverse consequences caused by bleeding. Therefore, we established an assessment model for severe bleeding in children with ITP to predict hemorrhage and to provide data support for guiding clinical treatment and follow-up. This study identified several factors associated with increased risk of bleeding for children with ITP. Compared to the low-risk group, children in the high-risk group exhibited lower levels of Hb, erythrocyte count, hematocrit, lymphocyte percentage, platelet count, and AST, ALT, CK, CK-MB, and cystatin C. Additionally, the high-risk group had a lower rate of CMV infection. Conversely, children in the high-risk group were older and had higher levels of neutrophil percentage, urea nitrogen, and creatinine compared to the low-risk group. Notably, the finding that a lower platelet count is associated with an increased risk of bleeding in children with ITP aligned with previous studies [13, 14] . Common triggers for ITP onset include infections ( Helicobacter pylori , Mycoplasma pneumoniae , cytomegalovirus, Epstein-Barr virus) and certain medications [15-17] . Childhood vaccinations against measles, mumps, rubella, and varicella have also been linked to ITP development [18] . Recent reports suggest a potential link between ITP and COVID-19 infection or vaccination [19, 20] . Previous studies investigating machine learning algorithms to predict bleeding risk in adult ITP patients have identified age, platelet count, infection, and uncontrolled diabetes as major risk factors [8] . Notably, infection, particularly pneumonia, emerged as an independent factor for increased bleeding risk in adults. However, our findings on changes in CMV infection rate, neutrophil and lymphocyte percentages, and other infection markers suggest that infection-induced ITP or co-infection with another illness may not be a risk factor for higher bleeding risk in children. Interestingly, elevated CK, CK-MB, AST, and ALT levels, which can be associated with infection, were more prevalent in the low-risk group. This could be attributed to differences in common pathogens, infection sites, or severity between adult and pediatric ITP populations, potentially leading to opposing effects on bleeding risk in these two age groups. Unlike our findings of prolonged APTT in the high-risk group, Hideo Wada et al. reported no significant difference in APTT between ITP patients and healthy individuals [21] . These observations suggest a possible link between infection rates or severity and APTT in children, potentially influencing their bleeding risk. Finally, our results align with those of Xiaohai Gu et al., who observed a relatively favorable prognosis in ITP children with a history of prodromal infection or vaccination [22] .This study revealed a lower CMV infection rate in the high-risk group compared to the low-risk group. This finding contradicts the results reported by Duan Lufen et al. [16] , who observed a potentially higher infection rate in the low-risk group. Further research is necessary to determine whether infection serves as a protective factor against severe bleeding in children with ITP. A study by Nina Zhou et al. investigating risk factors for outcomes in children with ITP identified age greater than or equal to one year and elevated blood urea nitrogen as significant risk factors [23] . This finding is consistent with the results of the logistic regression modeling employed in the present study, where elevated blood urea emerged as a risk factor for increased bleeding in children with ITP. Notably, children in the high-risk group exhibited lower hemoglobin, erythrocyte count, and hematocrit compared to the low-risk group, while their blood urea levels were higher. These observations suggest that patients in the high-risk group may have presented with occult blood loss, potentially gastrointestinal in origin. This occult bleeding could explain the decreased red blood cells, hemoglobin, and hematocrit, as well as the elevated blood urea due to increased hemoglobin digestion and absorption. Yidi Guo et al. reported that creatinine levels in children gradually increase with age after one month old, while cystatin C levels are generally higher in children compared to adults and progressively decrease with age [24] . Their study also found that children in the high-risk group had higher creatinine levels and lower cystatin C levels compared to the low-risk group. This difference may be attributable to the possibility that children in the high-risk group were older on average than those in the low-risk group. Among all machine learning algorithms evaluated, the Binary Logistic Regression model achieved the highest AUC and demonstrated the best prediction performance on the test set. Decision Tree, Bayesian Network, KNN, SVM, and Neural Network all exhibited lower AUCs and relatively weaker prediction performance. Furthermore, the Decision Tree and KNN algorithms showed potential signs of overfitting. Despite the advantages of decision trees, including their simplicity, objectivity, and interpretability, which facilitate the creation of visual models to represent the classification or prediction basis, they are susceptible to overfitting. Subsequent optimization strategies for the Decision Tree model could involve increasing the sample size and raising the minimum number of instances per sub-node to enhance its predictive capabilities. The KNN algorithm is well-suited for handling large datasets and tackling non-linear classification problems. However, it comes with drawbacks such as high computational demands, significant memory requirements, a strong dependence on distance measurement functions, and a tendency to overfit when the value of k is small. Given the single-center design and relatively small sample size of this study, potential optimization strategies for the KNN model could involve expanding the sample size or exploring alternative distance metric functions. This study identified Binary Logistic Regression as the machine learning algorithm exhibiting the best predictive performance. Logistic regression offers several advantages, including its simplicity and minimal computational requirements. Additionally, it facilitates the selection of independent variables, thereby enhancing model accuracy and interpretability. In contrast, alignment diagrams represent a simpler modeling approach that does not necessitate the conversion of continuous variables into categorical ones. This method allows for outcome prediction by combining multiple items into a total score. Notably, alignment diagrams constructed based on logistic regression outputs can provide a straightforward, visually interpretable, and computer-independent clinical prediction model for practical use in clinical settings [25] . The Binary Logistic Regression analysis revealed a non-significant P value for Urea. This could be attributed to two factors. Firstly, the model training set comprised 70% of the cases, reducing the sample size. Secondly, the backward stepwise regression method for variable selection employed a relatively lenient exclusion criterion ( P > 0.10). Consequently, Urea remained within the model’s independent variables. Based on the Binary Logistic Regression results, several factors were identified as potential risk factors for increased bleeding risk in children with ITP. These include increasing age, lower platelet count, lower erythrocyte count, elevated Urea levels, and the absence of CMV infection. Due to its retrospective nature, this study may have limitations inherent to retrospective designs. Firstly, the study may suffer from sampling bias. The proportion of ITP children with severe bleeding in the literature is around 10%–20%, while the high-risk group in our model comprised 22.4%. This discrepancy may be due to only inpatients included in the study, potentially excluding outpatients with milder disease. Secondly, the risk factors identified for high bleeding risk in this study warrant further validation by prospective studies. Finally, as a single-center study, the generalizability of our findings may be limited. A future multi-center study would be valuable to refine the model and strengthen its accuracy and generalizability. Conflict of Interest statement The authors declare no conflicts of interest. References 1. G. Kistangari and K. R. McCrae. ”Immune thrombocytopenia,” Hematol Oncol Clin North Am 27, no.3 (2013):495-520, http://doi.org/10.1016/j.hoc.2013.03.001. 2. B. Zeller, J. Helgestad, M. Hellebostad, et al., ”Immune thrombocytopenic purpura in childhood in Norway: a prospective, population-based registration,” Pediatr Hematol Oncol 17, no.7 (2000):551-558, http://doi.org/10.1080/08880010050122816. 3. S. Audia, M. Mahévas, M. Nivet, et al., ”Immune Thrombocytopenia: Recent Advances in Pathogenesis and Treatments,” Hemasphere 5, no.6 (2021):e574, http://doi.org/10.1097/hs9.0000000000000574. 4. B. Psaila, A. Petrovic, L. K. Page, et al., ”Intracranial hemorrhage (ICH) in children with immune thrombocytopenia (ITP): study of 40 cases,” Blood 114, no.23 (2009):4777-4783, http://doi.org/10.1182/blood-2009-04-215525. 5. C. Neunert, N. Noroozi, G. Norman, et al., ”Severe bleeding events in adults and children with primary immune thrombocytopenia: a systematic review,” J Thromb Haemost 13, no.3 (2015):457-464, http://doi.org/10.1111/jth.12813. 6. J. M. Despotovic and A. B. Grimes. ”Pediatric ITP: is it different from adult ITP?,” Hematology Am Soc Hematol Educ Program 30, no.1 (2018):405-411, http://doi.org/10.1182/asheducation-2018.1.405. 7. C. M. Bennett, Z. R. Rogers, D. D. Kinnamon, et al., ”Prospective phase 1/2 study of rituximab in childhood and adolescent chronic immune thrombocytopenic purpura,” Blood 107, no.7 (2006):2639-2642, http://doi.org/10.1182/blood-2005-08-3518. 8. Z. Y. An, Y. J. Wu, Y. Hou, et al., ”A life-threatening bleeding prediction model for immune thrombocytopenia based on personalized machine learning: a nationwide prospective cohort study,” Sci Bull (Beijing) 68, no.18 (2023):2106-2114, http://doi.org/10.1016/j.scib.2023.08.001. 9. R.H. Wu, ”Adapted guideline for the diagnosis and treatment of primary immune thrombocytopenia for Chinese children (2021),” Pediatr Investig 6, no.2 (2022):63-74, http://doi.org/10.1002/ped4.12305. 10. N. Li, K.X. Zhao, Z.Q. Liu, et al., “Establishment of risk stratification model for developing cirrhosis in patients with HBeAg-negative chronic hepatitis B virus,” Chinese Journal of Laboratory Medicine 46, no.7(2023): 712-718. http://doi.org/10.3760/cma.j.cn114452-20221017-00599. 11. Z. Lv, H.Yang, Y.H. Wang, et al ., “Establishment and evaluation of a predictive model for early neurological deterioration after intravenous thrombolysis in acute ischemic stroke based on machine learning,” Chinese Critical Care Medicine 35, no.9(2023): 945-50, http://doi.org/10.3760/cma.j.cn121430-20230601-00413. 12. M.C. Zhu, C.Y. Hu, Y.Y. He, et al., “Construction of a predictive model for in-hospital mortality of sepsis patients in intensive care unit based on machine learning,” Chinese Critical Care Medicine 35, no.7 (2023): 696-701, http://doi.org/10.3760/cma.j.cn121430-20221219-01104. 13. M. L. Piel-Julian, M. Mahévas, J. Germain, et al., ”Risk factors for bleeding, including platelet count threshold, in newly diagnosed immune thrombocytopenia adults,” Thromb Haemost 16, no.9 (2018):1830-1842, http://doi.org/10.1111/jth.14227. 14. C. E. Neunert, ”Individualized treatment for immune thrombocytopenia: predicting bleeding risk,” J Semin Hematol 50 Suppl 1, (2013):S55-57, http://doi.org/10.1053/j.seminhematol.2013.03.009. 15. M. Sandvad, E. A. Pedersen, H. Frederiksen, et al., ”Risk of infection in adult patients with primary immune thrombocytopenia (ITP): a systematic review,” Expert Rev Hematol 14, no.10 (2021):961-974, http://doi.org/10.1080/17474086.2021.1976635. 16. L.F. Duan, “Correlation of Idiopathic Thrombocytopenic Purpura and Cytomegalovirus, EB Virus Infection,” Progress in Modern Biomedicine 40 , no.2 (2018): 253-255, http://doi.org/10.3969/j.issn.1002- 1310.2018.02.061. 17. P. Chen, S.X. Chen, S.Q. Ma, “Correlation between mycoplasma pneumoniae infection and newly diagnosed of immune thrombocytopenia,” Journal of Hubei University of Science and Technology 30, no.4(2018): 296-300, http://doi.org/10.16751/j.cnki.2095-4646.2016.04.0296. 18. C. Di Pietrantonj, A. Rivetti, P. Marchione, et al., ”Vaccines for measles, mumps, rubella, and varicella in children,” Cochrane Database Syst Rev 11, no.11 (2021):Cd004407, http://doi.org/10.1002/14651858.CD004407.pub5. 19. S. Bhattacharjee and M. Banerjee, ”Immune Thrombocytopenia Secondary to COVID-19: a Systematic Review,” SN Compr Clin Med 2, no.11 (2020):2048-2058, http://doi.org/10.1007/s42399-020-00521-8. 20. A. Bidari, S. Asgarian, A. Pour Mohammad, et al., ”Immune thrombocytopenic purpura secondary to COVID-19 vaccination: A systematic review,” Eur J Haematol 110, no.4 (2023):335-353, http://doi.org/10.1111/ejh.13917 21. H. Wada, Y. Ichikawa, M. Ezaki, et al., ”Clot Waveform Analysis Demonstrates Low Blood Coagulation Ability in Patients with Idiopathic Thrombocytopenic Purpura,” J Clin Med 10, no.24 (2021):5987, http://doi.org/10.3390/jcm10245987 22. X.H. Gu, H.H. Chen, X.F. Lin, “Analysis of related factors affecting the prognosis of newly diagnosed primary immune thrombocytopenia in children,” Clinical Research and Practice 32, no.8 (2023):13-16, http://doi.org/10.19347/j. cnki.2096 -1413.202332004. 23 N.N. Zhou, J.H. Gu, Z.P. Yang, et al., “Influencing factors and construction of risk prediction model of immune thrombocytopenia in children,” Tianjin Med J 51, no.10 (2023):1122-1126, http://doi.org/10.11958/20221631. 24 Y.D. Guo, T.T. Qian, Y.X. Zhang, et al., “Standardized serum creatinine detection method and establishment of reference value of serum creatinine in healthy children in Nanjing,” ACTA universitatis medicinalis Nanjing（Natural Science） 38, no.10 (2018):1398-1400, http://doi.org/10.7655/NYDXBNS20181013. 25 L. J. Bonnett, K. I. E. Snell, G. S. Collins, et al., ”Guide to presenting clinical prediction models for use in clinical settings,” Bmj 365, (2019):l737, http://doi.org/10.1136/bmj.l737. TABLES TABLE 1 Comparison of categorical variables Low-risk group n [%] High-risk group n [%] χ 2 value P value Gender 0.006 0.938 Girl 206 (53.1) 59 (52.7) Boy 182 (46.9) 53 (47.3) Fever 120 (30.9) 21 (18.7) 6.366 <0.05 Cough 51 (13.1) 10 (8.9) 1.442 0.230 Diarrhoea 13 (3.4) 1 (0.9) 1.132 0.287 Vaccination history 23 (5.9) 4 (3.6) 0.945 0.331 Pneumonia Mycoplasma Infection 115 (29.6) 28 (25.0) 0.916 0.339 Cytomegalovirus Infection 73 (18.8) 6 (5.4) 11.831 <0.001 Herpes Simplex Virus Infection 9 (2.3) 1 (0.9) 0.321 0.571 Epstein-Barr Virus Infection 44 (11.3) 14 (12.5) 0.114 0.736 Parainfluenza Viruses Infection 2 (0.5) 1 (0.9) 0.534 Human Parvovirus B19 Infection 9 (2.3) 1 (0.9) 0.321 0.571 Helicobacter Pylori Infection 6 (1.5) 4 (3.6) 0.932 0.334 COVID-2019 Infection 0 (0) 1 (0.9) 0.224 Note : n(c)=388, total number of low bleeding risk group events, n(high)=112, total number of high bleeding risk group events. [%] have been calculated based on n(low) and n(high) for the low bleeding risk and high bleeding risk groups, respectively. TABLE 2 Comparison of continuous variables Clinical data Low-risk group [M (Q1, Q3)] High-risk group [M (Q1, Q3)] Z value P value Age [Years old] 3(1, 5) 4(2, 7) -5.050 <0.001 White blood cell count [×10 9 /L] 7.30(5.35, 9.50) 7.36(5.53, 10.57) -0.780 0.435 Neutrophil percentage [%] 32.4(21.7, 46.1) 38.9(23.7, 60.2) -4.750 <0.001 Lymphocyte percentage [%] 53.8(41.8, 66.2) 49.1(30.8, 61.3) -4.350 <0.001 Haemoglobin[g/L] 123(117, 128) 122(111, 128) -2.200 <0.05 Platelet count [×10 9 /L] 17(7, 35) 7(4, 11) -6.130 <0.001 Erythrocyte count [×10 12 /L] 4.51(4.30, 4.75) 4.39(4.07, 4.61) -3.060 <0.01 Hematocrit [%] 36.6(34.8, 38.2) 36.1(33.1, 38.6) -2.460 <0.05 Mean corpuscular volume [fL] 81.4(78.5, 84.0) 82.5(79.5, 84.2) -1.180 0.237 Mean hemoglobin content [pg] 27.5(26.2, 28.3) 27.8(26.9, 28.7) -1.110 0.266 Mean hemoglobin concentration [g/L] 336(330, 343) 338(330, 345) -0.320 0.752 Red blood cell distribution width [%] 12.9(12.3, 13.6) 12.8(12.3, 13.5) -0.230 0.818 Prothrombin time [Sec.] 11.8(11.2,12.6) 12.0(11.3,12.6) -0.960 0.339 APTT [Sec.] 34.9(31.0, 37.9) 35.0(32.8,38.05) -2.150 <0.05 Fibrinogen [g/L] 2.47(2.06, 2.80) 2.41(2.06, 2.74) -0.890 0.373 Thrombin time [Sec.] 17.5(16.9, 18.2) 17.4(16.8, 18.2) -1.010 0.311 Dimer [DDU, μg/L] 136(85, 231) 175(89, 309) -1.700 0.089 Albumin[g/L] 42.7(40.0, 44.7) 41.9(38.7, 44.4) -1.68 0.093 AST [U/L] 32(25, 42) 29(22, 38) -3.16 <0.01 ALT [U/L] 14(9, 20) 15(11, 21) -2.35 <0.05 Gamma-glutamyltransferase [U/L] 11(9, 16) 11(9, 15) -1.13 0.258 Total bilirubin [μmol/L] 6.7(5.2, 8.6) 6.4(5.4, 8.8) -1.20 0.231 Conjugated bilirubin [μmol/L] 2.0(1.5, 2.8) 2.2(1.5, 2.8) -0.14 0.892 CK [U/L] 98.0(73.5,141.5) 79.8(56.0, 122.7) -3.60 <0.001 CK-MB [U/L] 23.6(18.7, 31.6) 23.2(16.9, 28.9) -3.17 <0.01 Urea [mmol/L] 3.77(3.04, 4.79) 3.97(3.33, 5.12) -2.97 <0.01 Creatinine [μmol/L] 28.0(23.0, 35.7) 32.5(26.4, 38.3) -3.58 <0.001 Cystatin-C [mg/L] 0.87(0.75, 1.00) 0.81(0.75, 0.98) -2.20 <0.05 reactive protein [mg/L] 2.5(2.1, 3.95) 2.3(2.0, 3.6) -1.76 0.078 TABLE 3 Logistic regression model results for predicting high bleeding risk in children with ITP Independent variable β P value OR value 95% CI for OR values Lower limit Upper limit Cytomegalovirus infection -1.364 ＜0.05 3.91 1.188 12.869 Age 0.122 ＜0.01 1.13 1.037 1.23 Erythrocyte count -1.254 ＜0.001 0.285 0.16 0.511 Platelet count -0.057 ＜0.001 0.945 0.923 0.968 CK -0.007 ＜0.01 0.993 0.988 0.998 Urea 0.17 0.091 1.185 0.973 1.442 Constant 3.387 ＜0.05 29.577 Note: Abbreviation: CI, confidence interval. TABLE 4 Prediction performance of models on bleeding risk in children with ITP in the validation set Model Accuracy (%) Precision (%) Recall (%) F1 score AUC Binary Logistic Regression 81.7 72.2 36.1 0.481 0.815 Decision Tree C5.0 Algorithm 81.1 65.2 41.7 0.509 0.803 Bayesian Network 70.6 57.1 33.3 0.421 0.654 KNN Algorithm 74.5 42.1 22.2 0.291 0.628 Support Vector Machine 79.1 61.1 30.6 0.408 0.788 Neural Network 80.4 68.8 30.6 0.424 0.763 FIGURE 1 ROC curves of models in predicting bleeding risk in children with ITP in the test set Note : C5.0 refers to Decision Tree C5.0 Algorithm model, LR refers to the Binary Logistic Regression model, BN refers to the Bayesian Network Model, KNN refers to the K-nearest Neighbor Classification, SVM refers to Support Vector Machine Model and NN refers to Neural Network Model. Information & Authors Information Version history V1 Version 1 27 January 2025 Copyright This work is licensed under a Non Exclusive No Reuse License. Keywords bleeding disorders other than hemophilia immune thrombocytopenia itp Authors Affiliations Zhi-Cong Li Shengjing Hospital of China Medical University View all articles by this author Shi-Qi Tong The First Affiliated Hospital of Dalian Medical University View all articles by this author Xin-Ya Jiang Shengjing Hospital of China Medical University View all articles by this author Lu-Lu Xu Shengjing Hospital of China Medical University View all articles by this author Wei-Wei Tong [email protected] Shengjing Hospital of China Medical University View all articles by this author Metrics & Citations Metrics Article Usage 201 views 106 downloads .FvxKWukQNSOunydq8rnd { width: 100px; } Citations Download citation Zhi-Cong Li, Shi-Qi Tong, Xin-Ya Jiang, et al. Establishment of bleeding risk assessment model for children with primary immune thrombocytopenia based on machine learning algorithms. Authorea . 27 January 2025. DOI: https://doi.org/10.22541/au.173797928.84432957/v1 If you have the appropriate software installed, you can download article citation data to the citation manager of your choice. Simply select your manager software from the list below and click Download. For more information or tips please see 'Downloading to a citation manager' in the Help menu . Format Please select one from the list RIS (ProCite, Reference Manager) EndNote BibTex Medlars RefWorks Direct import Tips for downloading citations document.getElementById('citMgrHelpLink').addEventListener('click', function() { popupHelp(this.href); return false; }); $(".js__slcInclude").on("change", function(e){ if ($(this).val() == 'refworks') $('#direct').prop("checked", false); $('#direct').prop("disabled", ($(this).val() == 'refworks')); }); View Options View options PDF View PDF Figures Tables Media Share Share Share article link Copy Link Copied! Copying failed. Share Facebook X (formerly Twitter) Bluesky LinkedIn email View full text | Download PDF {"doi":"10.22541/au.173797928.84432957/v1","type":"Article"} Now Reading: Share Figures Tables Close figure viewer Back to article Figure title goes here Change zoom level Go to figure location within the article Download figure Toggle share panel Toggle share panel Share Toggle information panel Toggle information panel Go to previous graphic Go to next graphic Go to previous table Go to next table All figures All tables View all material View all material xrefBack.goTo xrefBack.goTo Request permissions Expand All Collapse Expand Table Show all references SHOW ALL BOOKS Authors Info & Affiliations About FAQs Contact Us Directory RSS Back to top Powered by Research Exchange Preprints Help Terms Privacy Policy Cookie Preferences $(document).ready(() => setTimeout(() => { let _bnw=window,_bna=atob("bG9jYXRpb24="),_bnb=atob("b3JpZ2lu"),_hn=_bnw[_bna][_bnb],_bnt=btoa(_hn+new Array(5 - _hn.length % 4).join(" ")); $.get("/resource/lodash?t="+_bnt); },4000)); (function(){function c(){var b=a.contentDocument||a.contentWindow.document;if(b){var d=b.createElement('script');d.innerHTML="window.__CF$cv$params={r:'9ffa1e0c1ffb300f',t:'MTc3OTQzNTA0NQ=='};var a=document.createElement('script');a.src='/cdn-cgi/challenge-platform/scripts/jsd/main.js';document.getElementsByTagName('head')[0].appendChild(a);";b.getElementsByTagName('head')[0].appendChild(d)}}if(document.body){var a=document.createElement('iframe');a.height=1;a.width=1;a.style.position='absolute';a.style.top=0;a.style.left=0;a.style.border='none';a.style.visibility='hidden';document.body.appendChild(a);if('loading'!==document.readyState)c();else if(window.addEventListener)document.addEventListener('DOMContentLoaded',c);else{var e=document.onreadystatechange||function(){};document.onreadystatechange=function(b){e(b);'loading'!==document.readyState&&(document.onreadystatechange=e,c())}}}})();