Prediction of avatrombopag-induced thrombocytosis in pediatric immune thrombocytopenia: an AI-based real-world study Authors: Xi Lin, Yuntian Wang, Yunqi Zhu, Nan Wang, Jingjing Liu, Zhifa Wang, Yu Hu, Shuyue Dong, Hui Chen, Jinxi Meng, Jingyao Ma, Zhenping Chen, Wensheng Zhang, Yongqiang Tang, Runhui Wu, Xiaoling Cheng Published: 2026 ABSTRACT Avatrombopag (AVA), an oral thrombopoietin receptor agonist (TPO-RA), has demonstrated favorable efficacy in the treatment of pediatric immune thrombocytopenia (ITP). However, treatment-related thrombocytosis represents a clinically relevant adverse event that may compromise treatment safety and continuity. Currently, no validated tools are available to predict the risk of AVA-induced thrombocytosis before treatment initiation. In this real-world study, we aimed to develop and validate a predictive model for AVA-associated thrombocytosis in children with ITP. A total of 74 pediatric patients treated with AVA at the Hematology–Oncology Center of Beijing Children’s Hospital between July 2021 and January 2024 were included. We compared the proposed model with established classical machine learning baselines, including Logistic Regression (LR), Support Vector Machine (SVM), Multilayer Perceptron (MLP), Random Forest (RF), and XGBoost, as well as state-of-the-art deep learning models for tabular data, including TabPFN, FT-Transformer, and HyperTab. Among the evaluated models, the FT-Transformer achieved the best performance, with an accuracy of 0.785 ± 0.023 and an area under the receiver operating characteristic curve (AUC) of 0.851 ± 0.021. Model interpretability was enhanced using Shapley Additive Explanations (SHAP), enabling visualization of individual feature contributions to thrombocytosis risk. This AI-driven prediction model, grounded in real-world clinical data, demonstrates robust predictive performance and offers clinically interpretable insights. It provides a reliable reference for individualized risk assessment and supports safer, more precise use of AVA in pediatric ITP management. Introduction Immune thrombocytopenia (ITP) is the most common bleeding disorder among children, with an incidence rate ranging from 1.5 to 6.3 per 100,000 [1]. The disease is characterized by immune-mediated platelet destruction and impaired platelet production, leading to an increased risk of bleeding. Thrombopoietin-receptor agonists (TPO-RAs) stimulate megakaryopoiesis and platelet production and additionally exert immunomodulatory effects, including enhancement of regulatory T-cell function, thereby promoting immune tolerance. In children who fail to respond adequately to first-line therapies, the American Society of Hematology (ASH) guidelines recommend TPO-RAs as preferred second-line treatment options [2]. Large clinical trials have consistently demonstrated high response rates and favorable efficacy profiles of TPO-RAs in pediatric ITP [3]. Avatrombopag (AVA), a second-generation, orally administered small-molecule TPO-RA, was introduced in 2018. The clinical superiority of AVA in TPO-RA-resistant pediatric patients stems from its high-affinity targeting of the His499 transmembrane residue. This leads to a more comprehensive activation of downstream signaling than eltrombopag. By operating through a non-competitive mechanism, AVA exerts an additive synergism with endogenous TPO, providing the necessary stimulus to overcome refractory platelet insufficiency. The global phase 3b trial demonstrated that AVA significantly improved platelet response in pediatric ITP compared to placebo (81% vs. 0%), with no thromboembolic events(TEEs) reported in children [4]. Real-world studies further support these findings, showing an 89% response rate even in pediatric cohorts who failed multiple prior therapies [5]. While clinical trials generally report a favorable safety profile, age-related differences are notable. Adult studies show grade 3/4 adverse events in 18.8% of patients, including TEEs in 9.4% [6]. In contrast, pediatric cohorts demonstrate enhanced safety, with no reported TEEs or severe bleeding [4]. However, while AVA is often used as a salvage therapy—a setting that might theoretically moderate extreme responses—cases of rapid platelet recovery to supraphysiological levels and subsequent TEEs have still been documented [7]. Thrombocytosis has emerged as one of the most clinically concerning adverse events observed in routine practice [8]. Accumulating evidence suggests that AVA-associated thrombocytosis is not uncommon. In a multicenter observational cohort, 31 of 76 ITP patients experienced platelet counts ≥ 400 × 10⁹/L during AVA treatment [9]. Similarly, real-world data from the GEPTI study reported thrombocytosis in 14.9% (40/268) of the cohort [7]. Pharmacovigilance analyses of the U.S. FDA Adverse Event Reporting System (FAERS) database further revealed platelet count elevation in approximately 11.9% of AVA-treated patients, ranking it among the most frequently reported adverse events. [10] Importantly, excessive platelet elevation has been linked to clinically significant complications, including transient ischemic–like events requiring neuroimaging surveillance and acquired von Willebrand syndrome leading to severe bleeding manifestations such as hemothorax. [11]. To mitigate these risks, current protocols rely on response-guided titration and vigilant monitoring. In pediatric settings, weekly monitoring aims to maintain platelets within a 50–150 × 10⁹/L range [4]. Adult guidelines employ even more granular steps, where counts > 150 × 10⁹/L trigger dose reduction and levels > 250 × 10⁹/L necessitate treatment interruption to mitigate thromboembolic risks [12]. However, the monitoring framework remains inherently reactive. Dose adjustments often occur only after dangerous platelet overshoots have already manifested, rather than through proactive risk stratification. Consequently, there is an urgent need for predictive tools to identify high-risk patients before treatment initiation. To date, no validated predictive models exist to identify patients at high risk of AVA-induced thrombocytosis before treatment initiation. Such tools are urgently needed to support precision medicine and enhance the safety of TPO-RA therapy in pediatric ITP. Our previous investigations suggested that AVA-induced thrombocytosis may be influenced by concomitant infections, particularly viral infections such as COVID-19, through synergistic effects involving inflammatory cytokines (e.g., interleukin-6), endothelial injury, and increased von Willebrand factor (VWF) release [13]. Elevated thrombopoietin levels and pulmonary megakaryocytes have also been observed in ITP patients recovering from severe acute respiratory syndrome [14]. Furthermore, increased transforming growth factor-β (TGF-β) levels have been associated with relapse or poor response to eltrombopag therapy [15]. Demographic and treatment-related factors—including age, sex, body weight, and combined use of glucocorticoids—have been implicated in TPO-RA-associated thrombotic risk, yet systematic prediction of drug-induced thrombocytosis remains unexplored [16]. Recent advances in machine learning (ML) and deep learning have enabled the development of powerful predictive models capable of handling high-dimensional clinical data, modeling nonlinear relationships, and accommodating missing values. ML-based approaches have been successfully applied to predict adverse events across various clinical contexts, including mortality and bleeding after acute coronary syndrome [17] and cardiac toxicity associated with immune checkpoint inhibitors [18]. In pediatrics, ML models have been used to stratify the risk of high-dose methotrexate–induced hepatotoxicity [19]. Notably, emerging deep learning architectures for tabular data—such as Transformer-based and hypernetwork-based models—have challenged the assumption that large datasets are required for deep learning. By leveraging attention mechanisms and prior-informed learning strategies, these models demonstrate strong performance even in small-sample clinical settings. This makes them particularly well suited for real-world pediatric studies, where data availability is inherently limited. In this study, we therefore developed an AI-based predictive model for AVA-induced thrombocytosis in children with ITP, applying real-world clinical data on classical machine learning and advanced tabular deep learning algorithms. Our aim was to identify key risk factors, provide interpretable predictions, and support individualized risk assessment and treatment optimization in pediatric ITP management. Methods Study design and population This retrospective real-world study enrolled pediatric patients diagnosed with immune thrombocytopenia (ITP) who received AVA therapy at the Hematology–Oncology Center of Beijing Children’s Hospital between 2021 and January 2024. Only patients treated with AVA specifically for ITP were eligible for inclusion. After systematic screening of medical records, a total of 74 children were included in the final analysis, among whom 33 experienced thrombocytosis at least once during AVA treatment. This study received approval from the Beijing Children Hospital ethics committee ([2024]-Y-098-D), and informed consent was obtained from participants. Data collection and outcome definition All data was collected from medical records. The inclusion criteria for this study were children aged 1 to 16 years old with ITP and start using AVA for treatment. This study defines thrombocytosis as two consecutive platelet counts ≥ 400 × 10⁹/L, measured at 3-day intervals after initiation of AVA. The data collected can be categorized into three groups: demographics, laboratory test results, and history of previous ITP treatment. Variables with more than 50% missing values, as well as patients lacking two or more features, will be excluded from the analysis. The primary outcome of interest is the occurrence of thrombocytosis, which has been converted into a binary variable: “0” indicates that thrombocytosis did not occur, while “1” indicates that it did (Fig. 1b). Clinical variable definition ITP disease phase was classified according to disease duration as newly diagnosed (0–3 months), persistent (3–12 months), or chronic (> 12 months). Treatment response was assessed two months after AVA initiation. Complete response (CR) was defined as a PLT count ≥ 100 × 10⁹/L without bleeding; partial response (PR) as a PLT count between 30 and 100 × 10⁹/L, at least doubling the baseline value and without bleeding; and no response (NR) as a PLT count < 30 × 10⁹/L, failure to double baseline PLT, or presence of bleeding. Prior treatment was defined as any medication administered between the diagnosis of ITP and the initiation of AVA. Concomitant treatment was defined as any medication administered at or after the initiation of AVA. Bleeding severity was graded according to the 2021 Chinese Guidelines for the diagnosis and treatment of childhood ITP [20]. Missing values were recorded if patients were lost to follow-up. Model establishment Initially, 76 candidate features encompassing demographic, clinical, laboratory, treatment-related, and disease-course variables were considered. No separate feature selection procedure (e.g., RFE, LASSO, or univariate filtering) was performed in this study. The reduction from 76 candidate variables to 66 retained predictors reflected data cleaning and missingness-based preprocessing rather than outcome-driven feature selection. Given the limited sample size, model evaluation was performed using 5-fold cross-validation, where four folds were used for training and one for testing, and the final performance was averaged across all folds. To select optimal hyperparameters, a nested 3-fold cross-validation was applied within each training split, with one fold serving as the validation set during tuning. Classical machine learning models including Logistic Regression (LR), Support Vector Machine (SVM), Multilayer Perceptron (MLP), Random Forest (RF), and XGBoost, as well as three state-of-the-art neural network–based machine learning algorithms specifically designed for tabular data were implemented (Fig. 1d). TabPFN, FT-Transformer, and HyperTab are neural network–based models that learn feature representations and interactions in an end-to-end manner and are therefore well suited for tabular data–based machine learning prediction tasks. FT-Transformer: A simplified adaptation of the Transformer architecture tailored for tabular data, which has demonstrated competitive performance across a range of tabular learning tasks [21]. TabPFN: A probabilistic Transformer-based model tailored for tabular data. It performs fast, prior-informed predictions by approximating the posterior over functions and achieves strong performance on small tabular datasets without requiring extensive hyperparameter tuning [22]. HyperTab: A hypernetwork-based framework for tabular modeling that dynamically generates task-specific network weights conditioned on input features [23]. This design enables adaptive representation learning and has demonstrated strong performance on tabular benchmarks by capturing complex feature interactions efficiently. The three machine learning models were compared across multiple performance metrics, including receiver operating characteristic curve (AUC), accuracy, precision, and F1 score (Fig. 1e). Model interpretability To enhance clinical interpretability, feature importance was evaluated using model-based importance analysis to identify variables most strongly associated with thrombocytosis risk (Fig. 1f). Additionally, Shapley Additive Explanations (SHAP) were applied to the optimal model to quantify the contribution of each feature to individual predictions and to visualize the directionality (risk-enhancing or risk-reducing) of feature effects (Fig. 2). Statistical analysis Categorical variables are presented as counts and percentages, while continuous variables are expressed as mean ± standard deviation or median with interquartile range, as appropriate. Model performance was assessed using accuracy, precision, recall, F1 score, and area under AUC. Statistical analyses and model evaluations were conducted using Python-based machine learning libraries. Results Baseline characteristics and Baseline demographic and clinical characteristics of the 74 enrolled children are summarized in Table 1. At the initiation of AVA, 10 patients (13.5%) were classified as newly diagnosed ITP, 21 (28.4%) as persistent ITP, and 43 (58.1%) as chronic ITP. Table 1Baseline clinical characteristics at AVA initiation in children with ITPVariableOverall (n = 74)Thrombocytosis (n = 33)No thrombocytosis (n = 41)P value*Age, years, median (IQR)7.8(5.5–10.5)6.7(4.9–8.5)8.43(5.9–11.1)0.076Male, n (%)28(37.8)8(24.2)20(48.8)0.019Body weight, kg, median (IQR)25.0(16.4–39.2)23.0(16.0–29.5)28.0(17.5–41.2)0.093ITP phase, n (%)New-diagnose10(13.5)5(15.1)5(12.2)0.470Persistent21(28.4)7(21.2)14(34.1)Chronic43(58.1)21(63.6)22(53.7)Platelet count at AVA initiation (×10⁹/L), median (IQR)16(8.0–40.0)17(11.0–32.0)14(7.0–27.5)0.338Bleeding score at AVA initiation (median, range)0(0–3)0(0–2)0(0–3)0.095Initial AVA dose (mg/kg), median (IQR)0.7(0.6–0.9)0.8(0.6–0.9)0.7(0.6–0.9)0.103*P values were calculated by comparing the thrombocytosis and non-thrombocytosis groups Table 1Baseline clinical characteristics at AVA initiation in children with ITPVariableOverall (n = 74)Thrombocytosis (n = 33)No thrombocytosis (n = 41)P value*Age, years, median (IQR)7.8(5.5–10.5)6.7(4.9–8.5)8.43(5.9–11.1)0.076Male, n (%)28(37.8)8(24.2)20(48.8)0.019Body weight, kg, median (IQR)25.0(16.4–39.2)23.0(16.0–29.5)28.0(17.5–41.2)0.093ITP phase, n (%)New-diagnose10(13.5)5(15.1)5(12.2)0.470Persistent21(28.4)7(21.2)14(34.1)Chronic43(58.1)21(63.6)22(53.7)Platelet count at AVA initiation (×10⁹/L), median (IQR)16(8.0–40.0)17(11.0–32.0)14(7.0–27.5)0.338Bleeding score at AVA initiation (median, range)0(0–3)0(0–2)0(0–3)0.095Initial AVA dose (mg/kg), median (IQR)0.7(0.6–0.9)0.8(0.6–0.9)0.7(0.6–0.9)0.103*P values were calculated by comparing the thrombocytosis and non-thrombocytosis groups Baseline clinical characteristics at AVA initiation in children with ITP *P values were calculated by comparing the thrombocytosis and non-thrombocytosis groups Prior treatment history before AVA initiation is summarized in Table 2. Nearly all patients had previously received first-line therapy with glucocorticoids and/or intravenous immunoglobulin. A substantial proportion had been exposed to second-line treatments prior to AVA initiation: 59 patients (79.7%) had received rituximab, 53 (71.6%) had been treated with other thrombopoietin receptor agonists, and 9 (12.2%) had received additional immunosuppressive agents. During AVA treatment, concomitant therapies are detailed in Table 3. Table 2Prior treatment history before avatrombopag in children with ITPVariableOverall (n = 74)Thrombocytosis (n = 33)No thrombocytosis (n = 41)P value*Glucocorticoids, n (%)73(98.6)32(97.0)41(100.0)0.446Intravenous immunoglobulin, n (%)73(98.6)33(100.0)40(97.6)0.833TPO-RAs, n (%)53(71.6)21(63.6)32(78.0)0.238Rituximab, n (%)59(79.7)26(78.8)33(80.5)0.857Other immunosuppressants, n (%)9(12.2)2(6.0)7(17.1)0.139Splenectomy, n (%)000-*P values were calculated by comparing the thrombocytosis and non-thrombocytosis groups Table 2Prior treatment history before avatrombopag in children with ITPVariableOverall (n = 74)Thrombocytosis (n = 33)No thrombocytosis (n = 41)P value*Glucocorticoids, n (%)73(98.6)32(97.0)41(100.0)0.446Intravenous immunoglobulin, n (%)73(98.6)33(100.0)40(97.6)0.833TPO-RAs, n (%)53(71.6)21(63.6)32(78.0)0.238Rituximab, n (%)59(79.7)26(78.8)33(80.5)0.857Other immunosuppressants, n (%)9(12.2)2(6.0)7(17.1)0.139Splenectomy, n (%)000-*P values were calculated by comparing the thrombocytosis and non-thrombocytosis groups Prior treatment history before avatrombopag in children with ITP *P values were calculated by comparing the thrombocytosis and non-thrombocytosis groups Table 3Concomitant therapies during AVA treatment in children with ITPVariableOverall (n = 74)Thrombocytosis (n = 33)No thrombocytosis (n = 41)P value*Intravenous immunoglobulin, n (%)17(23.0)6(18.2)11(26.8)0.379Glucocorticoids, n (%)11(14.9)4(12.1)7(17.0)0.552Rituximab, n (%)31(42.0)17(51.5)14(34.1)0.132None, n(%)33(44.6)13(39.4)20(48.8)0.419*P values were calculated by comparing the thrombocytosis and non-thrombocytosis groups Table 3Concomitant therapies during AVA treatment in children with ITPVariableOverall (n = 74)Thrombocytosis (n = 33)No thrombocytosis (n = 41)P value*Intravenous immunoglobulin, n (%)17(23.0)6(18.2)11(26.8)0.379Glucocorticoids, n (%)11(14.9)4(12.1)7(17.0)0.552Rituximab, n (%)31(42.0)17(51.5)14(34.1)0.132None, n(%)33(44.6)13(39.4)20(48.8)0.419*P values were calculated by comparing the thrombocytosis and non-thrombocytosis groups Concomitant therapies during AVA treatment in children with ITP *P values were calculated by comparing the thrombocytosis and non-thrombocytosis groups Thrombocytosis occurred in 33 patients (44.6%). Apart from sex, baseline characteristics and concomitant therapies at AVA initiation were comparable between the two groups. Model performance and comparison Predictive models were constructed using 76 candidate features across eight machine learning algorithms. Model performance metrics are presented in Table 4. Among the eight approaches, the FT-Transformer consistently demonstrated superior predictive performance, achieving an accuracy of 0.785 ± 0.023 and an area under the receiver operating characteristic curve (AUC) of 0.851 ± 0.021. Table 4Performance comparison of predictive models built by various machine learning algorithmsaccuracyprecisionrecallF1AUCLogistic Regression0.539 ± 00.452 ± 00.487 ± 00.454 ± 00.583 ± 0Support Vector Machine0.568 ± 00.526 ± 00.667 ± 00.548 ± 00.545 ± 0Random Forest0.563 ± 0.0210.55 ± 0.0680.476 ± 0.050.457 ± 0.0390.678 ± 0.019XGBoost0.569 ± 00.589 ± 00.513 ± 00.485 ± 00.671 ± 0Multilayer Perceptron0.62 ± 0.0230.59 ± 0.0250.477 ± 0.0150.509 ± 0.0210.638 ± 0.004TabPFN0.592 ± 0.0110.614 ± 0.0210.456 ± 0.0290.474 ± 0.0190.559 ± 0.004FT-Transformer 0.785 ± 0.023 0.766 ± 0.031 0.768 ± 0.034 0.751 ± 0.023 0.851 ± 0.021 HyperTab0.615 ± 0.0260.62 ± 0.0480.555 ± 0.0540.559 ± 0.0330.6 ± 0.021 Table 4Performance comparison of predictive models built by various machine learning algorithmsaccuracyprecisionrecallF1AUCLogistic Regression0.539 ± 00.452 ± 00.487 ± 00.454 ± 00.583 ± 0Support Vector Machine0.568 ± 00.526 ± 00.667 ± 00.548 ± 00.545 ± 0Random Forest0.563 ± 0.0210.55 ± 0.0680.476 ± 0.050.457 ± 0.0390.678 ± 0.019XGBoost0.569 ± 00.589 ± 00.513 ± 00.485 ± 00.671 ± 0Multilayer Perceptron0.62 ± 0.0230.59 ± 0.0250.477 ± 0.0150.509 ± 0.0210.638 ± 0.004TabPFN0.592 ± 0.0110.614 ± 0.0210.456 ± 0.0290.474 ± 0.0190.559 ± 0.004FT-Transformer 0.785 ± 0.023 0.766 ± 0.031 0.768 ± 0.034 0.751 ± 0.023 0.851 ± 0.021 HyperTab0.615 ± 0.0260.62 ± 0.0480.555 ± 0.0540.559 ± 0.0330.6 ± 0.021 Performance comparison of predictive models built by various machine learning algorithms Feature importance and SHAP analysis To enhance interpretability, SHAP were applied to the FT-Transformer model to quantify the contribution of individual features to thrombocytosis risk (Fig. 2). The SHAP summary plot illustrates both the magnitude and direction of feature effects across the cohort. Five variables emerged as the most influential predictors: sex, ITP disease phase, bone marrow megakaryocyte count, prior exposure to TPO receptor agonists, and concomitant use of rituximab. Male sex and longer disease duration were predominantly associated with positive SHAP values, indicating an increased predicted risk of thrombocytosis. Similarly, higher megakaryocyte counts in bone marrow smears and concomitant rituximab therapy consistently shifted predictions toward higher risk. In contrast, prior treatment with other TPO-RAs was mainly associated with negative SHAP values, suggesting a relative risk-reducing effect in patients previously exposed to this drug class. Collectively, these findings indicate that the model not only identifies clinically meaningful predictors but also distinguishes whether specific feature values contribute to increased or decreased thrombocytosis risk, thereby improving clinical interpretability. Discussion In this real-world study, we developed and validated an artificial intelligence(AI)–based model to predict AVA-induced thrombocytosis in children with immune thrombocytopenia. The FT-Transformer model demonstrated robust predictive performance, achieving an area under the curve (AUC) of 0.851 and an accuracy approaching 0.80. To our knowledge, this is the first study to apply AI for individualized risk prediction of AVA-associated thrombocytosis in pediatric ITP. By integrating AI-assisted analysis with predictive modeling, clinicians can estimate the likelihood of thrombocytosis based on both historical and longitudinal data. Anticipating such enables more informed clinical decision-making, including proactive dose titration and optimization of treatment strategies, ultimately improving safety and efficacy in pediatric populations where clinical experience remains limited. In our study, the most influential features identified by the model—sex, disease duration, megakaryocyte activity, and prior or concomitant therapies—are biologically plausible and consistent with established mechanisms of platelet production and treatment response in ITP. Notably, sex emerged as the most influential predictor, with male patients exhibiting a higher predicted risk of thrombocytosis. This finding aligns with previous observations that sex-related differences influence ITP phenotype and treatment response [24]. Bone marrow megakaryocyte count was another strong risk-enhancing feature. Higher baseline megakaryocyte reserves predict enhanced responses to TPO-RAs, whereas reduced megakaryocyte numbers indicate poor responsiveness. [25, 26] In addition, prior TPO-RA exposure was associated with lower thrombocytosis risk, whereas naïve patients may be more susceptible to excessive platelet increases after starting AVA. possibly due to heightened receptor sensitivity or lack of physiological adaptation. [27] Concomitant rituximab also increased risk, likely due to synergistic effects with TPO-RAs on platelet recovery [28]. This association may partly reflect cumulative treatment exposure rather than disease duration alone in chronic ITP. [29] While these risk factors are well-recognized, their combined and dynamic effects remain difficult to quantify in real-world settings using conventional methods. Our model addresses this by integrating multidimensional clinical and laboratory variables to generate interpretable predictions via SHAP analysis. The potential of such AI-driven frameworks is further exemplified by recent advances in hemophilia A, where machine learning models—incorporating large language models—have surpassed traditional pharmacokinetic tools in predictive accuracy [30]. Regarding treatment timing, prior therapies were defined cumulatively from ITP diagnosis to avatrombopag initiation. Although arbitrary time-window cutoffs were not applied, the potential for confounding variability was mitigated by the structured, stepwise treatment strategy employed in our cohort [31]. Most patients followed a predefined treatment protocol, progressing sequentially from corticosteroids to rituximab and then to TPO-RAs. This protocol-driven escalation, supported by consistent longitudinal monitoring, allowed us to incorporate prior therapies as key covariates to reflect the cumulative treatment burden before AVA initiation. Such a research design inherently mirrors the complexity of real-world clinical practice. Although our model offers significant robust advantages, its role should be viewed as a supportive decision-making tool rather than a definitive treatment algorithm for individualized management. This study demonstrates how integrating these variables into a quantifiable framework can assist in individualized risk stratification. Early identification of high-risk patients through this clinical adjunct could inform more vigilant monitoring and facilitate preemptive dose adjustments, helping to maintain platelet counts within a safe therapeutic window. Ultimately, such a proactive approach aims to refine the management of pediatric ITP by providing clinicians with actionable insights. Several limitations should be acknowledged. First, this was a single-center study with a relatively small sample size, reflecting the rarity of pediatric ITP and the real-world nature of the cohort. Second, external validation in independent, multi-institutional cohorts is currently lacking. Due to the requirement for standardized baseline laboratory assessments and the increasing use of AVA outside our center, the number of eligible patients with complete data remains limited, posing challenges for short-term external validation. To address these issues, we have implemented an electronic follow-up platform to enhance data granularity and are planning future multicenter studies to further validate and refine the model. In conclusion, this study demonstrates the feasibility of applying AI to predict AVA-induced thrombocytosis in children with ITP. By leveraging real-world data within a framework that icombines classical machine learning with interpretable deep learning, our approach provides actionable, clinically relevant insights. Rather than serving as a standalone decision-making algorithm, the model is intended as a clinical adjunct to support individualized risk assessment and optimize the safety and precision of TPO-RA therapy in pediatric ITP. Fig. 1Workflow of machine learning model development and interpretation for thrombocytosis prediction. The original dataset consisted of 74 patients with 76 candidate variables and a binary outcome of thrombocytosis (A). After data preprocessing and normalization, 66 variables were retained for subsequent analysis (B). The dataset was then partitioned into training and testing sets for model development and evaluation (C). Supervised machine learning models, including Logistic Regression, Support Vector Machine, Random Forest, XGBoost, Multilayer Perceptron, TabPFN, FT-Transformer, and HyperTab, were trained using the training data to generate thrombocytosis predictions (D). Model performance was assessed using the area under the receiver operating characteristic curve (AUC), accuracy, precision, and F1 score (E). Finally, model-based feature importance analysis was performed to identify key clinical and laboratory features associated with thrombocytosis risk. (F) Workflow of machine learning model development and interpretation for thrombocytosis prediction. The original dataset consisted of 74 patients with 76 candidate variables and a binary outcome of thrombocytosis (A). After data preprocessing and normalization, 66 variables were retained for subsequent analysis (B). The dataset was then partitioned into training and testing sets for model development and evaluation (C). Supervised machine learning models, including Logistic Regression, Support Vector Machine, Random Forest, XGBoost, Multilayer Perceptron, TabPFN, FT-Transformer, and HyperTab, were trained using the training data to generate thrombocytosis predictions (D). Model performance was assessed using the area under the receiver operating characteristic curve (AUC), accuracy, precision, and F1 score (E). Finally, model-based feature importance analysis was performed to identify key clinical and laboratory features associated with thrombocytosis risk. (F) Fig. 2The SHAP summary plot illustrates the impact of individual features on model predictions. Each row represents a single feature, and each dot corresponds to an individual patient. Dot color indicates the feature value, with red representing higher values and blue representing lower values. The position of each dot along the x-axis reflects the direction and magnitude of the feature’s contribution to the predicted risk. A predominance of red dots on the right side indicates that higher feature values are associated with an increased predicted risk, whereas a rightward shift of blue dots suggests that lower feature values contribute to a higher risk The SHAP summary plot illustrates the impact of individual features on model predictions. Each row represents a single feature, and each dot corresponds to an individual patient. Dot color indicates the feature value, with red representing higher values and blue representing lower values. The position of each dot along the x-axis reflects the direction and magnitude of the feature’s contribution to the predicted risk. A predominance of red dots on the right side indicates that higher feature values are associated with an increased predicted risk, whereas a rightward shift of blue dots suggests that lower feature values contribute to a higher risk