Abstract
Introduction:
Cancer therapy-induced thrombocytopenia (CTIT) is a common hematologic toxicity associated with anti-tumor treatment. Recombinant human interleukin-11 (rhIL-11), as a thrombopoietic agent, is widely used in clinical practice. However, its efficacy and safety exhibit substantial individual variability, and reliable tools for individualized prediction are currently unavailable.
Methods:
Based on real-world data from 12,431 CTIT patients treated with rhIL-11 across 58 Chinese centers (July 2023-June 2024), this study retrospectively collected comprehensive information on demographics, clinical characteristics, and treatment regimens. The Least Absolute Shrinkage and Selection Operator (LASSO) regression was applied to identify key clinical features. Six machine learning algorithms --logistic regression (LR), decision tree (DT), random forest (RF), support vector machine (SVM), extreme gradient boosting (XGBoost), and lightweight gradient boosting machine (LightGBM) --were used to construct separate models for predicting platelet response (efficacy) and adverse events (safety) following rhIL-11 treatment. The dataset was randomly divided into training and test sets in a 7:3 ratio, and an independent external validation set was used to assess model generalizability. SHapley Additive exPlanations (SHAP) analysis was used to provide visual explanations of model feature importance.
Results:
The RF model demonstrated superior performance in efficacy prediction, with an AUC of 0.812 (95% CI: 0.797-0.828) in the test set and 0.761 (95% CI: 0.717-0.805) in the external validation set; for safety prediction, the RF model also performed best, with an AUC of 0.796 (95% CI: 0.778-0.814) in the test set and 0.739 (95% CI: 0.695-83), in the external validation set. However, the relatively low specificity of RF models in predicting efficacy and safety limited their potential for practical clinical application. SHAP analysis revealed that the predominant factors influencing efficacy were chemotherapy, prior anti-tumor therapy, history of grade III myelosuppression, targeted therapy, and duration of rhIL-11 treatment; whereas key predictors of safety outcomes included immunotherapy, dosing frequency, chemotherapy, and CTIT grades.
Discussion:
This study validated that machine learning models (particularly the RF model combined with SHAP analysis) exhibit highly sensitive preliminary screening performance, while enhancing model interpretability. These findings may facilitate the effective identification of patients most likely to benefit from rhIL-11 and provide references for personalized, precision treatment decisions in CTIT management.
1 Introduction
Cancer therapy-induced thrombocytopenia (CTIT) is a common hematologic toxicity associated with anti-tumor treatment, which can lead to treatment delay, interruption, dose reduction, and even bleeding events, thereby compromising therapeutic efficacy and long-term survival (; ). CTIT is defined as a peripheral blood platelet (PLT) count below 100 Ć 109/L attributable to anti-tumor therapies (such as chemotherapy, radiotherapy, targeted therapy, and immunotherapy), with severity graded from I to IV according to the Common Terminology Criteria for Adverse Events (CTCAE) version 5.0 (; ). The incidence and severity of CTIT vary with tumor types, anti-tumor treatment regimens, comorbidities, and extent of bone marrow involvement (). Current treatment strategies for CTIT primarily include PLT transfusion and thrombopoietic agents, the latter represented by recombinant human interleukin-11 (rhIL-11), recombinant human thrombopoietin (rh-TPO), and thrombopoietin receptor agonists (TPO-RAs) (). Although PLT transfusion is the most rapid, direct, and effective intervention for CTIT, its routine use is limited by blood product shortages, short duration of PLT count maintenance, and the risks of transfusion reactions and infections. Consequently, it is typically reserved for patients with PLT counts ā¤10 Ć 109/L (; ). RhIL-11 is one of the thrombopoietic agents approved by both the United States Food and Drug Administration (FDA) and the National Medical Products Administration (NMPA) of China for the treatment of CTIT. As a pleiotropic cytokine, it stimulates the proliferation and differentiation of hematopoietic stem cells and megakaryocyte precursors, while also promoting megakaryocyte maturation, thereby accelerating PLT production (; ; ). Although multiple studies, including our previous work, have confirmed the overall efficacy and safety of rhIL-11 in treating CTIT (; ), considerable heterogeneity remains in its efficacy and safety across different real-world clinical settings. Therefore, developing personalized and precise treatment strategies for patients receiving rhIL-11 has become an urgent priority in the clinical management of CTIT.
In recent years, significant advances have been made in research on rhIL-11 for CTIT treatment. A cross-sectional study involving 1,437 Chinese patients with CTIT reported that 27% opted for monotherapy with rhIL-11 (). Several clinical studies have further revealed the heterogeneity in the efficacy of rhIL-11, which is closely associated with tumor type, chemotherapy regimen, baseline PLT level, and individual patient characteristics (; ). However, effective tools for predicting treatment response to rhIL-11 remain unavailable, and clinical decision-making still relies primarily on empirical judgment, lacking sufficient precision. Therefore, identifying the key factors that influence the efficacy and safety of rhIL-11 and developing personalized predictive models have become major priorities in current research.
Machine learning has been increasingly applied in the medical field and has demonstrated notable advantages in the screening, diagnosis, and prognosis assessment of various diseases, particularly in predictive research related to malignant tumors (). The technology can extract and integrate clinically relevant information from large-scale datasets to support clinical decision-making (). With the help of machine learning, it is expected that personalized and precise therapeutic decisions will become more achievable. However, despite their high predictive capabilities, machine learning models are often limited by their inherent āblack boxā nature, which makes it difficult to elucidate the specific contributions of each variable to the predictive results, thereby undermining their credibility and acceptability in clinical practice (). The SHapley Additive Explanations (SHAP) method combines optimal credit allocation with local interpretability, intuitively illustrating the importance of each variable and thus generating more interpretable model outputs ().
Therefore, this study aims to develop a machine learning-based predictive model using real-world data to identify CTIT patients likely to benefit from rhIL-11 treatment in terms of efficacy and safety, thereby providing decision support for personalized and precise clinical management. Additionally, SHAP analysis will be applied to enhance model interpretability and facilitate the identification of high-risk patients, thereby promoting early intervention in clinical practice.
2 Materials and methods
2.1 Study population
We retrospectively analyzed clinical data from 12,431 patients with malignant tumors who received rhIL-11 treatment for CTIT across 58 medical centers in China between July 2023 and June 2024. The data encompassed baseline demographic characteristics, medical history, tumor types, anti-tumor treatment regimens, rhIL-11 administration details, efficacy evaluations, and adverse events (AEs). This study was approved by the ethics committees of all participating centers and conducted in accordance with the principles of the Declaration of Helsinki (2013 revision). Given the retrospective design of the study, all data were de-identified prior to analysis, and the requirement for informed consent was waived by the ethics committees.
Inclusion Criteria: (1) Diagnosis of malignant tumor confirmed by pathology or cytology; (2) Receipt of anti-tumor treatments such as chemotherapy, radiotherapy, targeted therapy, or immunotherapy; (3) Development of thrombocytopenia (PLT count ā¤100 Ć 109/L) following treatment and subsequent treatment with rhIL-11; (4) Regular monitoring of PLT counts during treatment.
Exclusion Criteria: (1) Thrombocytopenia caused by non-anti-tumor treatment factors (e.g., chronic liver disease or infection); (2) Use of other PLT-elevating drugs; (3) Poor adherence to rhIL-11 treatment or discontinuous use; (4) Incomplete records or missing critical information.
2.2 Data extraction and preprocessing
We extracted relevant clinical characteristics from the data, including gender, age, weight, history of comorbidities (hypertension, coronary heart disease, diabetes, infectious diseases, history of surgery or trauma, and history of COVID-19 infection), history of prior myelosuppression (grades I to IV), tumor types (hematologic malignancy, lung cancer, gastric cancer, colorectal cancer, hepatic carcinoma, esophageal cancer, uterine malignancy, prostate cancer, breast cancer, pancreatic cancer, thyroid cancer, head and neck malignancy, urothelial cancer, and others), bone marrow involvement, history of prior anti-tumor therapy, history of bleeding, current chemotherapy, radiotherapy, targeted therapy, immunotherapy, surgery, chemotherapy regimens (platinum monotherapy, taxane monotherapy, anthracycline monotherapy, fluorouracil monotherapy, gemcitabine-based combination, platinum-based combination, anthracycline-based combination, taxane-based combination, and others), radiotherapy regimens (radiotherapy alone, concurrent chemoradiotherapy, and sequential chemoradiotherapy), baseline CTIT grades (grades I to IV), total days of rhIL-11 prescription, and dosing frequency (first-time, multiple sessions). The Least Absolute Shrinkage and Selection Operator (LASSO) regression was employed to identify key features while discarding irrelevant data. Multicategorical variables were included in the analysis after dummy variable transformation.
2.3 Treatment regimen
RhIL-11 was administered subcutaneously at a dose of 25ā50 μg/kg/day until the PLT count reached ā„100 Ć 109/L or increased by ā„ 50 Ć 109/L from the baseline level.
2.4 Outcome evaluation metrics
The primary endpoint of this study is defined as the response rate, measured as the incidence of PLT count returning to 100 Ć 109/L within 1Ā month of the first administration of rhIL-11; the secondary endpoint is the incidence of treatment-related AEs. AEs were classified and assessed according to the CTCAE v5.0, and mainly included fatigue, edema, gastrointestinal reactions, fever, rash, and atrial fibrillation, with severity graded from I to V.
2.5 Construction, validation, and evaluation of machine learning models
We randomly selected a medical centerās data as the external validation set, while the data from the remaining centers were randomly divided into training and testing sets in a 7:3 ratio (all data from each center was allocated to one set without internal splitting). A total of six machine learning models were constructed in this study: Extreme Gradient Boosting (XGBoost), Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), Lightweight Gradient Boosting Machine (LightGBM), and Decision Tree (DT). Five-fold cross-validation was used to perform hyperparameter tuning for the six models. The test set was not used during the model tuning phase but was reserved for model evaluation only after training was completed. The key hyperparameters for each model are detailed in Supplementary Tables S1 and S2.
To assess model performance, we compared various evaluation metrics, including the area under the ROC curve (AUC), accuracy, specificity, precision, recall, and F1-score, with the AUC serving as the primary evaluation metric to identify the machine learning model with the best predictive performance. Calibration curves were used to assess the consistency between actual and predicted outcomes, while decision curve analysis (DCA) was used to evaluate the net clinical benefit of each model.
2.6 Interpretability analysis
SHAP analysis was performed on the optimal predictive model to assess its interpretability. Based on Shapley values derived from game theory, SHAP calculates the contribution of each feature to the prediction results (). The importance of global features is ranked and presented in bar charts, highlighting the average influence of each feature on the modelās output. Additionally, waterfall plots and force plots were employed to interpret individual samples, revealing the direction and intensity of each featureās influence on the prediction results, which helps in understanding the modelās decision-making process.
2.7 Statistical analysis
Statistical analysis and data visualization were performed using SPSS v29, R v4.5.0, and Origin. Continuous variables were presented as means and standard deviations, with group comparisons conducted using the Wilcoxon test or the KruskalāWallis test. Categorical variables were presented as frequencies and percentages, and group comparisons were performed using the Ļ2 test or Fisherās exact test. The following R packages were used in this study for data preparation, analysis, and visualization, including: ādplyrā (v1.1.4), ādata.tableā (v1.17.4), ācaretā (v7.0ā1), āglmnetā (v4.1ā9), āreadxlā (v1.4.5), āopenxlsxā (v4.2.8), āMatrixā (v1.7ā3), āxgboostā (v1.7.11.1), āe1071ā (v1.7ā16), ārandomForestā (v4.7ā1.2), ālightgbmā (v4.6.0), ātreeā (v1.0ā44), ārpartā (v4.1.24), āpartykitā (v1.2ā24), āpROCā (v1.18.5), āggplot2ā (v3.5.2), āggpubrā (v0.6.0), ārmsā (v8.0ā0), ādcurvesā (v0.5.0), āggbeeswarmā (v0.7.2), āshapvizā (v0.9.7), āfastshapā (v0.1.1). All statistical tests were two-tailed, and a P-value of <0.05 was considered statistically significant.
3 Results
3.1 Baseline information and clinical characteristics
Based on the inclusion and exclusion criteria, 12,431 CTIT patients treated with rhIL-11 were included in the analysis cohort. Among them, 7,653 were allocated to the training set, 3,274 to the test set, and 1,504 to the external validation set. Table 1 presents the baseline characteristics of the patients, including demographic details, medical history, tumor types, anti-tumor treatment regimens, and medication status. In this study, there were 1,463 patients with grade IV CTIT, distributed as follows: 666 in the training set, 632 in the test set, and 165 in the external validation set. Among all participants, 8,177 received chemotherapy, including 5,439 in the training set, 1,524 in the test set, and 1,214 in the external validation set.
TABLE 1
| Variables | Categories | Training set (n = 7653) | Test set (n = 3274) | External validation set (n = 1504) | P-value |
|---|---|---|---|---|---|
| Gender (%) | male | 4316 (56.4) | 1804 (55.1) | 847 (56.3) | 0.446 |
| female | 3337 (43.6) | 1470 (44.9) | 657 (43.7) | ||
| Age (%) | <60Y | 3823 (50.0) | 1993 (60.9) | 914 (60.8) | <0.001 |
| ā„60Y | 3830 (50.0) | 1281 (39.1) | 590 (39.2) | ||
| Weight | Median | 62Ā kg | 61Ā kg | 60Ā kg | ā |
| mean ± SD | 62.05 ± 9.73 | 61.51 ± 8.59 | 60.00 ± 8.02 | <0.001 | |
| Medical history (%) | no | 6791 (88.7) | 2857 (87.3) | 1454 (96.7) | <0.001 |
| yes | 862 (11.3) | 417 (12.7) | 50 (3.3) | ||
| Prior myelosuppression (%) | no | 6238 (81.5) | 2297 (70.2) | 1227 (81.6) | <0.001 |
| Grade I | 149 (1.9) | 38 (1.2) | 21 (1.4) | ||
| Grade II | 555 (7.3) | 240 (7.3) | 83 (5.5) | ||
| Grade III | 447 (5.8) | 339 (10.4) | 64 (4.3) | ||
| Grade IV | 264 (3.4) | 360 (11.0) | 109 (7.2) | ||
| Tumor type (%) | Hematologic malignancy | 1958 (25.6) | 1235 (37.7) | 492 (32.7) | <0.001 |
| Solid tumor | 5695 (74.4) | 2039 (62.3) | 1012 (67.3) | ||
| Lung cancer | 1308 (17.1) | 421 (12.9) | 142 (9.4) | ||
| Gastric cancer | 810 (10.6) | 335 (10.2) | 158 (10.5) | ||
| Colorectal cancer | 955 (12.5) | 278 (8.5) | 176 (11.7) | ||
| Hepatic carcinoma | 658 (8.6) | 200 (6.1) | 211 (14.0) | ||
| Esophageal cancer | 275 (3.6) | 165 (5.0) | 26 (1.7) | ||
| Uterine malignancy | 351 (4.6) | 144 (4.4) | 61 (4.1) | ||
| Prostate cancer | 60 (0.8) | 10 (0.3) | 7 (0.5) | ||
| Breast cancer | 408 (5.3) | 219 (6.7) | 37 (2.5) | ||
| Pancreatic cancer | 171 (2.2) | 34 (1.0) | 7 (0.5) | ||
| Thyroid cancer | 38 (0.5) | 27 (0.8) | 1 (0.1) | ||
| Head and neck malignancy | 145 (1.9) | 69 (2.1) | 118 (7.8) | ||
| Urothelial cancer | 57 (0.7) | 14 (0.4) | 9 (0.6) | ||
| Other | 459 (6.0) | 123 (3.8) | 59 (3.9) | ||
| Bone marrow involvement (%) | no | 6694 (87.5) | 3016 (92.1) | 1290 (85.8) | <0.001 |
| yes | 959 (12.5) | 258 (7.9) | 214 (14.2) | ||
| Prior anti-tumor therapy (%) | no | 6795 (70.4) | 1922 (58.7) | 926 (61.6) | <0.001 |
| yes | 2858 (29.6) | 1352 (41.3) | 578 (38.4) | ||
| Current surgery (%) | no | 7350 (96.0) | 3137 (95.8) | 1425 (94.7) | 0.072 |
| yes | 303 (4.0) | 137 (4.2) | 79 (5.3) | ||
| Current chemotherapy (%) | no | 2304 (30.1) | 1750 (53.5) | 290 (19.3) | <0.001 |
| yes | 5349 (69.9) | 1524 (46.5) | 1214 (80.7) | ||
| Current radiotherapy (%) | no | 6564 (85.8) | 2865 (87.5) | 1465 (97.4) | <0.001 |
| yes | 1089 (14.2) | 409 (12.5) | 39 (2.6) | ||
| Current targeted therapy (%) | no | 6781 (88.6) | 2583 (78.9) | 1348 (89.6) | <0.001 |
| yes | 872 (11.4) | 691 (21.1) | 156 (10.4) | ||
| Current immunotherapy (%) | no | 6700 (87.5) | 3007 (91.8) | 1398 (93.0) | <0.001 |
| yes | 953 (12.5) | 267 (8.2) | 106 (7.0) | ||
| Baseline CTIT grade (%) | Grade I | 584 (7.6) | 176 (5.4) | 162 (10.8) | <0.001 |
| Grade II | 2718 (35.5) | 830 (25.4) | 560 (37.2) | ||
| Grade III | 3685 (48.2) | 1636 (50.0) | 617 (41.0) | ||
| Grade IV | 666 (8.7) | 632 (19.3) | 165 (11.0) | ||
| Dosing frequency (%) | First-time | 5005 (65.4) | 2161 (66.0) | 985 (65.5) | 0.828 |
| Multiple sessions | 2648 (34.6) | 1113 (34.0) | 519 (34.5) | ||
| Concurrent myelostimulatory drugs* (%) | no | 6645 (86.8) | 2837 (86.7) | 1320 (87.8) | 0.549 |
| yes | 1008 (13.2) | 437 (13.3) | 184 (12.2) | ||
| Bleeding history (%) | no | 7402 (96.7) | 3210 (98.0) | 1375 (91.4) | <0.001 |
| yes | 251 (3.3) | 64 (2.0) | 129 (8.6) |
Baseline characteristics of patients.
Myelostimulatory medicines include recombinant human granulocyte colony-stimulating factor and erythropoietin.
3.2 Construction and evaluation of efficacy prediction models
3.2.1 Feature selection
The LASSO regression method was employed to identify key features, with iterative analysis conducted through 10-fold cross-validation. By selecting the lambda value corresponding to one standard error, 14 features closely associated with the therapeutic efficacy of rhIL-11 were identified from the training set: age, history of comorbidities, history of prior myelosuppression, tumor types, bone marrow involvement, history of prior anti-tumor therapy, history of bleeding, current chemotherapy, radiotherapy, targeted therapy, chemotherapy regimens, radiotherapy regimens, baseline CTIT grades, and total days of rhIL-11 prescription (Figure 1).
FIGURE 1
3.2.2 Evaluation and comparison of model performance
Based on 14 key features, six machine learning models were constructed to predict the clinical efficacy of rhIL-11 treatment in patients with CTIT. After tuning the hyperparameters, each model was trained using the entire training set and subsequently evaluated in the test set. The discriminatory performance of each model in the test set is shown in Figure 2A. Among these models, the RF model exhibited the best performance, with an AUC of 0.812 (95% CI: 0.797ā0.828). The remaining models also demonstrated good predictive ability, ranked in descending order of performance as follows: LightGBM (AUC = 0.778, 95% CI: 0.762ā0.794), XGBoost (AUC = 0.772, 95% CI: 0.755ā0.788), SVM (AUC = 0.758, 95% CI: 0.740ā0.775), DT (AUC = 0.680, 95% CI: 0.661ā0.699), and LR (AUC = 0.672, 95% CI: 0.653ā0.692). To comprehensively evaluate model performance, we calculated accuracy, specificity, precision, recall, and F1-score (Supplementary Table S3). Comparative analysis demonstrated that the RF model outperformed other models across most metrics, with the highest accuracy (0.771), precision (0.784), recall (0.914), and F1-score (0.844), but relatively low specificity (0.477) (Supplementary Figure S1).
FIGURE 2
In terms of clinical applicability, decision curve analysis showed that all models demonstrated robust net benefit across most threshold probability ranges, with the RF model exhibiting the highest net benefit (Figure 2B). Furthermore, the calibration curve for the RF model demonstrated a good fit, indicating a high degree of consistency between the modelās predicted probabilities and actual outcomes, which further confirmed the reliability of its predictions (Figures 2C,D).
3.2.3 Model performance in external validation set
The RF model also demonstrated favorable predictive performance for efficacy outcomes in the external validation set, with an AUC of 0.761 (95% CI: 0.717ā0.805) (Figure 3).
FIGURE 3
3.3 Construction and evaluation of safety prediction models
3.3.1 Feature selection
We used the LASSO regression method to identify key features and performed iterative analysis using 10-fold cross-validation. By selecting the lambda value corresponding to one standard error, we identified 14 features in the training set closely related to the safety of rhIL-11 treatment: age, weight, history of comorbidities, history of prior myelosuppression, tumor types, bone marrow involvement, history of prior anti-tumor therapy, current surgery, chemotherapy, targeted therapy, immunotherapy, chemotherapy regimens, dosing frequency, and baseline CTIT grades (Figure 4).
FIGURE 4
3.3.2 Evaluation and comparison of model performance
Leveraging the 14 selected features, we constructed six machine learning models to predict the occurrence of AEs in CTIT patients treated with rhIL-11. The discriminatory performance of each model in the test set is shown in Figure 5A, with the RF model exhibiting superior performance and achieving an AUC of 0.796 (95% CI: 0.778ā0.814). The remaining models also demonstrated certain predictive capabilities, ranked in descending order of performance as follows: XGBoost (AUC = 0.742, 95% CI: 0.722ā0.762), LightGBM (AUC = 0.740, 95% CI: 0.720ā0.761), SVM (AUC = 0.711, 95% CI: 0.688ā0.733), LR (AUC = 0.697, 95% CI: 0.676ā0.719), and DT (AUC = 0.667, 95% CI: 0.645ā0.689). To comprehensively evaluate model performance, we calculated accuracy, specificity, precision, recall, and F1-score (Supplementary Table S4). Overall, the RF model performed best among all evaluated models, with the highest accuracy (0.821), precision (0.826), recall (0.968), and F1-score (0.891), although its specificity was relatively low (0.357) (Supplementary Figure S2).
FIGURE 5
With regard to clinical utility, decision curve analysis showed that models such as RF, LightGBM, and XGBoost demonstrated robust net benefit across most threshold probability ranges, with the RF model exhibiting the highest net benefit (Figure 5B). In addition, the calibration curve for the RF model demonstrated a good fit, indicating strong consistency between the predicted probability of AEs and the actual incidence rate, providing important insights into the modelās predictive reliability (Figures 5C,D).
3.3.3 Model performance in external validation set
In terms of safety prediction, the RF model continued to demonstrate good predictive performance in the external validation set, with an AUC of 0.739 (95% CI: 0.695ā0.783), further confirming the modelās robustness (Figure 6).
FIGURE 6
3.4 Interpretability analysis of the efficacy and safety prediction models
3.4.1 The efficacy prediction model
To improve the interpretability of the RF model, SHAP analysis was performed to characterize the contribution of features to the prediction results at both the global and individual levels. In terms of treatment efficacy prediction, the higher the SHAP value of each feature, the greater its influence on PLT recovery following rhIL-11 treatment. As shown in Figure 7A, the global feature importance was ranked in descending order by median SHAP values of each variable in the RF model. The five features most strongly associated with efficacy prediction were: current chemotherapy, history of prior anti-tumor therapy, history of prior grade III myelosuppression, current targeted therapy, and total days of rhIL-11 prescription. The SHAP beeswarm plot further illustrated the direction and contribution of these features in the predictive model output, with yellow representing high feature values and purple representing low feature values (Figure 7B).
FIGURE 7
At the individual level, local explanation quantified and visualized the contribution of each feature to the prediction of a specific sample. This approach not only explained the nonlinear relationship between features and prediction outcomes but also provided traceable evidence for personalized prediction, thereby enhancing the credibility of model-based decision-making in clinical applications (; ). Through a randomly selected patient as an example, the SHAP waterfall plot demonstrated that current chemotherapy (+0.141), a platinum-based chemotherapy regimen (+0.136), and gastric cancer (+0.107) were the major positive factors to PLT recovery in this patient (Figure 7C). In this plot, yellow indicates a positive contribution, purple indicates a negative contribution, and the length of each bar represents the magnitude of the contribution. E [f(x)] represents the SHAP baseline value, namely, the average model prediction, whereas f(x) refers to the SHAP output for the individual sample, corresponding to the actual predicted probability for that patient (). By accumulating SHAP values, the plot provides an intuitive view of how the final prediction is generated, aiding in a deeper understanding of the decision-making logic of the machine learning models. Additionally, the single-sample force plot (Figure 7D) further illustrates the cumulative formation process underlying the efficacy prediction for this patient.
3.4.2 The safety prediction model
In terms of safety prediction, the higher the SHAP value of a feature, the greater its influence on the occurrence of adverse reactions to rhIL-11 treatment. The global importance of features in the RF model is shown in Figure 8A, where the featuresā contribution to the model is sorted in descending order by median SHAP value. The top five features most strongly associated with safety prediction were: current immunotherapy, dosing frequency, chemotherapy, grade II CTIT, and weight. The SHAP beeswarm plot visualizes the direction and extent of these features in the predictive model, with yellow representing high feature values and purple representing low feature values (Figure 8B).
FIGURE 8
At the individual level, by randomly selecting a patient sample for local interpretation, the SHAP waterfall plot illustrated the contribution of each feature to the prediction of adverse reactions for that patient. Among these, factors such as current immunotherapy (+0.286) and targeted therapy (+0.161) were the primary positive predictors of adverse reactions, while grade II CTIT and weight exhibited negative contributions (Figure 8C). Furthermore, the single-sample force plot (Figure 8D) provided a more intuitive visualization of the cumulative formation process of the safety prediction results for this patient.
4 Discussion
RhIL-11 is one of the key hematopoietic agents widely used in the treatment of CTIT and has been in clinical practice for over 20Ā years (; ). However, heterogeneity in real-world efficacy and the risk of adverse reactions remain critical challenges, complicating clinical decision-making and limiting the personalized application of this agent (; ). Previous studies have predominantly focused on comparisons between rhIL-11 and other hematopoietic agents or pharmacoeconomic evaluations. Nevertheless, there is a lack of predictive tools to guide individualized treatment strategies that can adequately balance therapeutic benefits and risks (; ). Based on real-world data from 12,431 patients across 58 medical centers, we developed and validated predictive models for both the clinical efficacy and safety of rhIL-11 in CTIT management. This work not only confirmed the clinical effectiveness of rhIL-11 in the large-scale population but also identified key factors influencing treatment outcomes by employing interpretable machine learning methods, providing crucial guidance for personalized clinical treatment decisions.
Among the models evaluated for predicting therapeutic efficacy, the RF model demonstrated superior performance on the test set (AUC = 0.812), consistently outperforming other models in terms of accuracy, precision, recall, and F1-score. Furthermore, it also maintained good predictive performance on the external validation set (AUC = 0.761), demonstrating its generalizability. SHAP analysis further identified the main features influencing the efficacy of rhIL-11, including current chemotherapy, history of prior anti-tumor therapy, previous grade III myelosuppression, current targeted therapy, and total days of rhIL-11 prescription. Chemotherapy primarily induces chemotherapy-induced thrombocytopenia (CIT) by damaging and suppressing hematopoietic stem cells and megakaryocytes (; ). RhIL-11 acts on the aforementioned target cells to promote PLT production (). Notably, our findings indicated that patients with a history of prior anti-tumor therapy responded better to rhIL-11. This suggests that the bone marrow microenvironment shaped by previous treatment may be associated with responsiveness to subsequent hematopoietic agents. The underlying mechanism may be that repeated anti-tumor treatments select for stem cells with stronger anti-apoptotic capacity and relatively preserved hematopoietic reserve, which are more readily driven to proliferate or mature by hematopoietic agents during PLT depletion. Previous studies have indicated that a history of prior myelosuppression is a significant risk factor for the development of CTIT (), consistent with our findings and further quantifying the impact of this factor on the efficacy of rhIL-11. Of note, total days of rhIL-11 prescription, as a modifiable factor, provides a potential entry point for optimizing clinical dosing regimens and maximizing therapeutic outcomes.
For safety prediction, the RF model also achieved the best performance, with an AUC of 0.796 in the test set, and its accuracy, precision, recall, and F1-score reached 0.821, 0.826, 0.968, and 0.891, respectively. SHAP analysis revealed that current immunotherapy, dosing frequency, current chemotherapy, CTIT grades, weight and age are the key features for predicting AEs. In particular, the association between immunotherapy and the risk of AEs warrants special attention. In recent years, with the widespread adoption of immunotherapy, immune checkpoint inhibitor-induced thrombocytopenia (ICI-IT) has attracted increasing attention in clinical practice. ICI-IT is commonly observed in melanoma and non-small cell lung cancer. Although relatively uncommon, ICI-IT can be life-threatening, and its pathogenesis involves T-cell-mediated immune responses and autoantibody production (; ; ; ). Furthermore, in addition to its hematopoietic-promoting activity, rhIL-11 also exerts immunomodulatory effects involved in the regulation of inflammation and immune responses. Consequently, in the context of immunotherapy, the safety management of rhIL-11 treatment must take into account more complex drug interactions to avoid an increased risk of AEs. Preclinical studies of rhIL-11 have shown that repeated administration results in cumulative toxic effects, yet the overall safety profile remains manageable (), which is consistent with the results of this study. Recent chemotherapy exposure often causes multi-organ damage involving the heart, kidneys, liver, and bone marrow (; ; ; ), leading to reduced functional reserve and compensatory capacity. This subsequently decreases the bodyās tolerance to rhIL-11 and increases the risk of adverse reactions. Given that the current chemotherapy is associated with both efficacy and an increased risk of AEs, the benefit-risk ratio must be carefully weighed when using rhIL-11 to treat CTIT. The dual influence of age and weight further highlights the importance of individualized dosing, suggesting that adjusting the dose based on body surface area may be an effective strategy for optimizing the risk-benefit ratio.
In terms of the modeling strategy, this study established efficacy as the primary endpoint and safety as the secondary endpoint, constructing separate predictive models for each. This design enhances the specificity of the features for each endpoint, reduces the risk of overfitting associated with joint modeling, and better aligns with the needs of multidimensional clinical assessment. Although this strategy cannot automatically capture the intrinsic relationships between endpoints, combining it with SHAP analysis can elucidate these interconnections, thereby providing a basis for clinically precise treatment benefit-risk balancing.
Notably, although the RF model performs well overall in predicting both efficacy and safety, its specificity remains relatively low. Based on this, the RF model tends to yield a high false-positive rate, limiting its accuracy in directly guiding personalized treatment. The relatively low specificity may be attributed to several factors, including insufficient feature discriminative power, suboptimal threshold selection, and inherent limitations of the study design and dataset. Negative samples may exhibit high clinical heterogeneity, with a more dispersed feature distribution, making it difficult for the model to learn a consistent classification pattern. In addition, the default probability threshold (0.5) currently used in the model may not be the optimal choice; adjusting this threshold appropriately is likely to achieve a better balance between recall and specificity. Furthermore, given the retrospective nature of the data, inherent biases and missing variables may have constrained the modelās ability to capture detailed nuances of real-world clinical decision-making. Future research could further enhance the modelās overall discriminatory ability through optimizing feature selection, adjusting the probability threshold, and conducting multi-center validation studies, thereby better meeting practical clinical needs.
Furthermore, while selecting an external validation set on a per-center basis approximates real-world cross-center heterogeneity, it also results in significant differences among the training, testing, and external validation sets across multiple key clinical characteristics, thereby introducing systematic selection bias. This baseline discrepancy may lead to an overestimation or underestimation of model performance in external validation. The external validation performance metrics should be interpreted as conservative estimates, reflecting both generalization error and distributional shift, and should not be directly equated with expected real-world performance in unobserved clinical settings. Therefore, the interpretation of external validation results requires careful consideration of the potential impact of baseline differences on performance estimates. Nevertheless, the model still maintains above-average discriminative performance on the external validation set despite these imbalances. It is recommended that future studies incorporate center effects as covariates into the modeling algorithm or optimizing data partitioning strategies to obtain more reliable assessments of generalizability.
This study has several strengths. Firstly, the large sample size and multi-center data substantially enhance the generalizability and extrapolation of the results, while the diversity of the study population enables the model to capture broader clinical characteristics. Secondly, the application of multiple machine learning algorithms overcomes the limitations of the linear assumptions inherent in traditional statistical models, better reflecting the complex non-linear relationships found in clinical data. Moreover, the integration of SHAP explainability analysis quantifies the magnitude and direction of the effects of key features on the predictive results, offering clinicians an intuitive basis for result interpretation and identification of potential intervention targets.
However, some limitations should be acknowledged. The retrospective design may introduce selection bias. Certain clinical variables (such as PLT counts prior to the next treatment cycle and laboratory parameters) were not included in the analysis, which may result in the omission of some key variables. In addition, the documentation of AEs relies on real-world case data, which may be subject to potential under-reporting or incomplete documentation. In the future, further large-scale, prospective studies should be conducted to validate the modelās accuracy and its generalizability to other populations. Clinical variables from a wider range of dimensions should be incorporated to continuously improve the modelās predictive accuracy, and the feasibility of integrating it into clinical decision support systems should be explored, such as developing web-based calculators or integrating it into hospital information systems to enable real-time treatment recommendations and risk assessment.
In summary, machine learning-based predictive models developed and validated in this study help identify patient subgroups likely to benefit from rhIL-11 treatment for CTIT, as well as potential safety risks. The RF model demonstrates reasonable predictive performance and generalizability, suggesting its potential as a preliminary screening tool with high clinical sensitivity. The key characteristics further identified by SHAP analysis (such as anti-tumor therapy regimens, CTIT grades, and history of prior myelosuppression) serve as an evidence-based foundation for developing stratified treatment strategies and identifying high-risk populations. With the integration of multi-omics data and the advancement of artificial intelligence technologies, we look forward to developing more precise predictive models to drive the transition of CTIT treatment towards precision medicine.
Statements
Data availability statement
The raw data supporting the conclusions of this article are not publicly available due to ethical and privacy restrictions related to clinical patient data, but are available from the corresponding authors upon reasonable request and with appropriate institutional approval.
Ethics statement
The studies involving humans were approved by Ethics Committee of Renmin Hospital of Wuhan University. The studies were conducted in accordance with the local legislation and institutional requirements. The ethics committee/institutional review board waived the requirement of written informed consent for participation from the participants or the participantsā legal guardians/next of kin because Given that this study is retrospective in design, all data were de-identified prior to analysis, and the ethics committee approved an exemption from obtaining informed consent from patients.
Author contributions
YiW: Methodology, Validation, Investigation, Writing ā review and editing, Formal Analysis, Writing ā original draft. JX: Writing ā review and editing, Data curation, Writing ā original draft, Investigation. YD: Writing ā original draft, Writing ā review and editing, Data curation, Conceptualization. ZL: Conceptualization, Writing ā review and editing. SY: Writing ā review and editing. YuW: Data curation, Writing ā review and editing. SQ: Writing ā review and editing. XL: Writing ā review and editing. JM: Writing ā review and editing. DT: Writing ā review and editing. YY: Project administration, Methodology, Supervision, Funding acquisition, Resources, Writing ā review and editing.
Funding
The author(s) declared that financial support was received for this work and/or its publication. This work was supported by Beijing Health Alliance Charitable Foundation (No. QSMYZL001) and Beijing Health Alliance Charitable Foundation (No. JZT-rh11-2024).
Conflict of interest
The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Generative AI statement
The author(s) declared that generative AI was not used in the creation of this manuscript.
Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.
Publisherās note
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.
Supplementary material
The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fcell.2026.1882398/full#supplementary-material
References
1
BelleV.PapantonisI. (2021). Principles and practice of explainable machine learning. Front. Big Data4, 688969. 10.3389/fdata.2021.688969
2
CapobiancoE. (2022). High-dimensional role of AI and machine learning in cancer research. Br. J. Cancer126, 523ā532. 10.1038/s41416-021-01689-z
3
ChenM.LiL.XiaQ.ChenX.LiaoZ.WangC.et al (2024). A real-world observation on thrombopoietic agents for patients with cancer treatment-induced thrombocytopenia in China: a multicenter, cross-sectional study. Cancer130, 1524ā1538. 10.1002/cncr.35292
4
DingL.PengL.XuZ.CuiZ.WangZ. (2026). Interpretable machine-learning prediction of severe myelosuppression in colorectal cancer patients receiving chemotherapy using XGBoost and SHAP: a retrospective study with a web-based calculator. Front. Oncol.16, 1785146. 10.3389/fonc.2026.1785146
5
GaoA.ZhangL.ZhongD. (2023). Chemotherapy-induced thrombocytopenia: literature review. Discov. Oncol.14, 10. 10.1007/s12672-023-00616-3
6
GongF. M.LiuF. Y.MaX.MaS. T.XiaoH. T.JiangG.et al (2024). Effectiveness and economic evaluation of rhTPO and rhIL-11 in the treatment of cancer therapy induced thrombocytopenia based on real-world research. Front. Pharmacol.15, 1288964. 10.3389/fphar.2024.1288964
7
GuanC.GongA.ZhaoY.YinC.GengL.LiuL.et al (2024). Interpretable machine learning model for new-onset atrial fibrillation prediction in critically ill patients: a multi-center study. Crit. Care28, 349. 10.1186/s13054-024-05138-0
8
Guideline Working Committee of Chinese Society of Clinical Oncology (2025). Guidelines of Chinese Society of Clinical Oncology (CSCO): Cancer Therapy Induced Thrombocytopenia. Beijing: People's Medical Publishing House.
9
HumbrechtC.KientzD.GachetC. (2018). Platelet transfusion: current challenges. Transfus. Clin. Biol.25, 151ā164. 10.1016/j.tracli.2018.06.004
10
JotatsuT.OdaK.YamaguchiY.NoguchiS.KawanamiT.KidoT.et al (2018). Immune-mediated thrombocytopenia and hypothyroidism in a lung cancer patient treated with nivolumab. Immunotherapy10, 85ā91. 10.2217/imt-2017-0100
11
KrollM. H.Rojas-HernandezC.YeeC. (2022). Hematologic complications of immune checkpoint inhibitors. Blood139, 3594ā3604. 10.1182/blood.2020009016
12
KudererN. M.DesaiA.LustbergM. B.LymanG. H. (2022). Mitigating acute chemotherapy-associated adverse events in patients with cancer. Nat. Rev. Clin. Oncol.19, 681ā697. 10.1038/s41571-022-00685-3
13
KuterD. J. (2022). Treatment of chemotherapy-induced thrombocytopenia in patients with non-hematologic malignancies. Haematologica107, 1243ā1263. 10.3324/haematol.2021.279512
14
LiX.DvornekN. C.ZhouY.ZhuangJ.VentolaP.DuncanJ. S. (2019). Efficient interpretation of deep learning models using graph structure and cooperative game theory: application to ASD biomarker discovery. Inf. Process Med. Imaging11492, 718ā730. 10.1007/978-3-030-20351-1_56
15
LiY.DingN.LiuX.SiQ.WangY.LiC.et al (2026). The disrupted mitochondrial quality control network: a unifying mechanism and therapeutic target for chemotherapy-induced multi-organ toxicity. Biol. (Basel)15, 230. 10.3390/biology15030230
16
LiuN. W.HuangX.LiuS.LiuW. J.WangH.WangW. D.et al (2019). Elevated BNP caused by recombinant human interleukin-11 treatment in patients with chemotherapy-induced thrombocytopenia. Support Care Cancer27, 4293ā4298. 10.1007/s00520-019-04734-z
17
LiuZ.WangY.YanJ.LiuJ.ChenB.ZhangL.et al (2020). Efficacy and safety of recombinant human interleukin-11 in the treatment of acute leukaemia patients with chemotherapy-induced thrombocytopenia: a systematic review and meta-analysis. J. Eval. Clin. Pract.26, 262ā271. 10.1111/jep.13152
18
LundbergS. M.ErionG.ChenH.DegraveA.PrutkinJ. M.NairB.et al (2020). From local explanations to global understanding with explainable AI for trees. Nat. Mach. Intell.2, 56ā67. 10.1038/s42256-019-0138-9
19
LustbergM. B.KudererN. M.DesaiA.BergerotC.LymanG. H. (2023). Mitigating long-term and delayed adverse events associated with cancer treatment: implications for survivorship. Nat. Rev. Clin. Oncol.20, 527ā542. 10.1038/s41571-023-00776-9
20
MeiJ.JiaoF.LiY.CuiJ.YangH.WangL. (2025). Application of thrombopoietic agents in cancer therapy-induced thrombocytopenia: a comprehensive review. Blood Rev.70, 101257. 10.1016/j.blre.2025.101257
21
MooreD. C.ElmesJ. B.ArnallJ. R.StrasselS. A.PatelJ. N. (2024). PD-1/PD-L1 inhibitor-induced immune thrombocytopenia: a pharmacovigilance study and systematic review. Int. Immunopharmacol.129, 111606. 10.1016/j.intimp.2024.111606
22
National Cancer Institute (2017). āCommon terminology criteria for adverse events,ā in Version 5.0. Available online at: https://ctep.cancer.gov/protocoldevelopment/electronic_applications/ctc.htm#ctc_50 (Accessed June 13, 2026).
23
SchifferC. A.BohlkeK.DelaneyM.HumeH.MagdalinskiA. J.McculloughJ. J.et al (2018). Platelet transfusion for patients with cancer: American society of clinical oncology clinical practice guideline update. J. Clin. Oncol.36, 283ā299. 10.1200/JCO.2017.76.1734
24
ShoukierM.BorthakurG.JabbourE.RavandiF.Garcia-ManeroG.KadiaT.et al (2021). The effect of eltrombopag in managing thrombocytopenia associated with tyrosine kinase therapy in patients with chronic myeloid leukemia and myelofibrosis. Haematologica106, 2853ā2858. 10.3324/haematol.2020.260125
25
ShuL.YanH.WuY.YanT.YangL.ZhangS.et al (2024). Explainable machine learning in outcome prediction of high-grade aneurysmal subarachnoid hemorrhage. Aging (Albany NY)16, 4654ā4669. 10.18632/aging.205621
26
SongA. B.Al-SamkariH. (2025). Chemotherapy-induced thrombocytopenia: modern diagnosis and treatment. Br. J. Haematol.206, 1062ā1066. 10.1111/bjh.20037
27
SongY.ZhouF.SongN. X.LiuX. M.YuZ.XieL. N.et al (2018). Impact on platelet recovery of recombinant human thrombopoietin in severe aplastic anemia patients with allogeneic hematopoietic stem cell transplantation. Zhonghua Xue Ye Xue Za Zhi39, 207ā211. 10.3760/cma.j.issn.0253-2727.2018.03.007
28
SwansonK.WuE.ZhangA.AlizadehA. A.ZouJ. (2023). From patterns to patients: advances in clinical machine learning for cancer diagnosis, prognosis, and treatment. Cell.186, 1772ā1791. 10.1016/j.cell.2023.01.035
29
TangS. Q.TangL. L.MaoY. P.LiW. F.ChenL.ZhangY.et al (2021). The pattern of time to onset and resolution of immune-related adverse events caused by immune checkpoint inhibitors in cancer: a pooled analysis of 23 clinical trials and 8,436 patients. Cancer Res. Treat.53, 339ā354. 10.4143/crt.2020.790
30
YangH.XuX.TanM.GaoJ.FangR.LiuX.et al (2025). Efficacy and safety of different treatments in chemotherapy-induced thrombocytopenia: a systematic review and network meta-analysis. Front. Pharmacol.16, 1549214. 10.3389/fphar.2025.1549214
31
YuK. M.LauJ. Y.FokM.YeungY. K.FokS. P.ZhangS.et al (2018). Preclinical evaluation of the mono-PEGylated recombinant human interleukin-11 in cynomolgus monkeys. Toxicol. Appl. Pharmacol.342, 39ā49. 10.1016/j.taap.2018.01.016
32
ZeienJ.QiuW.TriayM.DhaibarH. A.Cruz-TopeteD.CornettE. M.et al (2022). Clinical implications of chemotherapeutic agent organ toxicity on perioperative care. Biomed. Pharmacother.146, 112503. 10.1016/j.biopha.2021.112503
33
ZeylabiF.Nameh Goshay FardN.ParsiA.PezeshkiS. M. S. (2023). Bone marrow alterations in COVID-19 infection: the root of hematological problems. Curr. Res. Transl. Med.71, 103407. 10.1016/j.retram.2023.103407
34
ZhangJ. J.ZhaoR.XiaF.LiY.WangR. W.GuanX.et al (2022). Cost-effectiveness analysis of rhTPO and rhIL-11 in the treatment of chemotherapy-induced thrombocytopenia in hematological tumors based on real-world data. Ann. Palliat. Med.11, 2709ā2719. 10.21037/apm-22-880
35
ZhaoX.ShanX.SuiS.SongQ.ChengM.ZhaoY. (2025). Cancer treatment-induced thrombocytopenia: diagnosis, mechanisms and management. Front. Immunol.16, 1634688. 10.3389/fimmu.2025.1634688
Summary
Keywords
cancer therapy-induced thrombocytopenia (CTIT), machine learning, predictive model, random forest (RF), recombinant human interleukin-11 (rhIL-11)
Citation
Wen Y, Xia J, Dong Y, Liu Z, Yu S, Wang Y, Qing S, Li X, Miao J, Tang D and Yao Y (2026) Machine learning-based models for predicting the efficacy and safety of recombinant human interleukin-11 in the treatment of cancer therapy-induced thrombocytopenia: exploration and preliminary validation from a multicenter retrospective study. Front. Cell Dev. Biol. 14:1882398. doi: 10.3389/fcell.2026.1882398
Received
15 May 2026
Revised
13 June 2026
Accepted
22 June 2026
Published
09 July 2026
Volume
14 - 2026
Edited by
Fuying Dao, Nanyang Technological University, Singapore
Reviewed by
Junqi Liu, The First Affiliated Hospital of Zhengzhou University, China
Yefeng Yin, Chinese Academy of Medical Sciences and Peking Union Medical College, China
Updates
Copyright
Ā© 2026 Wen, Xia, Dong, Liu, Yu, Wang, Qing, Li, Miao, Tang and Yao.
This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
*Correspondence: Yi Yao, yaoyi2018@whu.edu.cn
ā These authors have contributed equally to this work
Disclaimer
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.