Machine learning-based prediction and external validation of treatment-related myelosuppression in patients with non-small cell lung cancer receiving PD-1 inhibitors plus platinum-doublet chemotherapy Authors: Jingyao Hui, Rui Hu, Zhenni Yang, Jiale Ding, Yaru Guan, Xiaoya Li, Long Li Published: 2026 ABSTRACT AbstractBackgroundThe application of programmed cell death protein 1 (PD-1) inhibitors combined with platinum-based double-drug chemotherapy in patients with non-small cell lung cancer (NSCLC) is becoming increasingly widespread. However, the impact of this treatment regimen on the bone marrow hematopoietic system is well-defined. Therefore, we have to consider the risk of bone marrow suppression in NSCLC patients after receiving this treatment regimen. Our objective was to identify risk factors for the risk of myelosuppression, one of the serious complications of PD-1 inhibitor plus platinum-based dual-agent chemotherapy, in patients with NSCLC and to develop an effective machine learning (ML) model to predict this risk.MethodsWe retrospectively enrolled patients with NSCLC who received PD-1 inhibitor plus platinum-doublet chemotherapy at the Department of Respiratory Medicine, The First Hospital of Lanzhou University between July 2018 and March 2026. A subset of these patients was randomly divided into a training set (70%) and a test set (30%). In the training set, feature selection was performed using recursive feature elimination (RFE), least absolute shrinkage and selection operator (LASSO), and random forest (RF). Multiple ML models were constructed and evaluated, with the area under the curve (AUC) as the primary performance metric. Model interpretability was assessed using Shapley Additive Explanations (SHAP). External validation was performed using a temporally distinct subsequent cohort.ResultsFeature selection using RFE, LASSO, and RF identified age, body mass index (BMI), tumor size, platelet count, red cell distribution width (RDW), total protein, white blood cell count (WBC), and red blood cell count (RBC) as significant risk factors for myelosuppression in patients with NSCLC receiving PD-1 inhibitor plus platinum-doublet chemotherapy. Among the developed ML models, light gradient boosting machine (LightGBM) demonstrated the best performance, achieving AUCs of 0.898 in the training set, 0.841 in the test set, and 0.793 in external validation.ConclusionsThe LightGBM model effectively predicts the risk of myelosuppression in patients with NSCLC receiving PD-1 inhibitor plus platinum-doublet chemotherapy and may provide useful support for clinical decision-making.Keywords: Non-small cell lung cancer (NSCLC), chemotherapy, machine learning (ML), prediction, myelosuppression Highlight box Key findings • We developed a machine learning (ML) model to predict the risk of myelosuppression in patients with non-small cell lung cancer (NSCLC) receiving programmed cell death protein 1 (PD-1) inhibitors combined with platinum-based doublet chemotherapy, and conducted rigorous external validation. What is known and what is new? • Myelosuppression is a common adverse effect of chemotherapy, but previous studies have primarily focused on traditional chemotherapy regimens. There remains a gap in predictive models for myelosuppression risk specifically associated with the combination of PD-1 inhibitors and platinum-based doublet chemotherapy. • This study is the first to identify eight independent risk factors (age, body mass index, tumor size, platelet count, red blood cell distribution width, total protein, white blood cell count, and red blood cell count) in this specific patient population, and has developed a ML prediction model based on these factors. What is the implication, and what should change now? • The identified independent risk factors should be incorporated into clinical practice to guide early monitoring and timely intervention for these indicators, thereby reducing the risk of myelosuppression in such patients. The light gradient boosting machine model, which performed best in the study, demonstrated strong predictive performance and can provide valuable support for personalized clinical decision-making. Introduction Lung cancer (LC) is the most commonly diagnosed cancer and the leading cause of cancer-related death worldwide (1). In 2022, approximately 2.3 million new cases and 1.9 million deaths were reported globally (1). Non-small cell lung cancer (NSCLC), the predominant histological subtype of LC, accounts for more than 85% of all cases. Notably, around 40% of patients are ineligible for surgery resection and therefore require systemic chemotherapy (2,3). Multimodal therapy is currently the recommended treatment strategy for NSCLC (4-6). In particular, the combination of programmed cell death protein 1 (PD-1) inhibitors with platinum-based doublet chemotherapy has become a standard regimen for both adjuvant and neoadjuvant treatment (7-10). This approach not only improves prognosis (10). but may also increase resectability in selected patients by influencing surgical evaluation (11). However, not all patients benefit from this regimen (12,13). Nearly half of patients with NSCLC experience adverse events following combined immunotherapy and chemotherapy, and approximately 20% require hospital readmission due to treatment-related toxicity (14). Myelosuppression is one of the most common adverse events associated with antitumor therapy. Its pathophysiological basis involves the cytotoxic effects of anticancer agents on the bone marrow hematopoietic system (15). These effects can lead to reductions in platelet count, red blood cell count (RBC), and white blood cell count (WBC), thereby increasing the risks of bleeding, infection, and other complications. Myelosuppression may also disrupt treatment continuity and, in severe cases, become life-threatening (16,17). Cancer-related clinical characteristics are highly complex and often strongly intercorrelated. Traditional statistical methods, such as logistic regression, are therefore increasingly insufficient for risk factor identification and risk prediction in oncology (18,19). By contrast, machine learning (ML) algorithms often perform better when applied to high-dimensional and complex datasets (20,21). However, the “black-box” nature of ML models limits interpretability. Shapley Additive Explanations (SHAP), a game theory-based method, has emerged as one of the most widely used approaches for interpreting ML models and addressing this limitation. Although several previous studies have investigated adverse events following multimodal therapy in patients with cancer (22-24), few have specifically examined the risk of myelosuppression in patients with NSCLC receiving PD-1 inhibitors combined with platinum-based doublet chemotherapy. Therefore, we collected data from patients treated at the Department of Respiratory Medicine, The First Hospital of Lanzhou University between July 2018 and March 2026. After feature selection, we constructed multiple ML models, including support vector machine (SVM), gradient boosting machine (GBM), neural network (NNET), eXtreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), and categorical boosting (CatBoost), to predict the risk of myelosuppression in this population. GBM and XGBoost capture complex nonlinear interactions through gradient boosting. LightGBM and CatBoost optimize training efficiency and missing value handling. NNET can fit highly unstructured patterns. Multi-model comparison can avoid the contingency of a single algorithm, and enhance the reliability and generalization ability of the prediction results. We present this article in accordance with the TRIPOD reporting checklist (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-1555/rc). Methods Data source A total of 400 patients with NSCLC who received PD-1 inhibitor plus platinum-doublet chemotherapy at the Department of Respiratory Medicine, The First Hospital of Lanzhou University, were included according to predefined inclusion and exclusion criteria. PD-1 inhibitors include pembrolizumab, sintilizumab, and camrelizumab. Platinum agents include cisplatin or carboplatin. The combination regimen of two-drug chemotherapy included gemcitabine, etoposide, pemetrexed, and paclitaxel. The study period spanned July 2018 to March 2026. Patients treated before January 2025 were assigned to the model development cohort (training and testing set, n=335), while those treated between January 2025 and March 2026 were assigned to a temporally distinct external validation cohort (n=65). This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. All patient data were de-identified, and the requirement of informed consent for this retrospective study was waived. The study was approved by the Ethics Committee of The First Hospital of Lanzhou University (No. LDYYLL2026-615). Data collection All characteristics, except for the variables used to define the outcome, were defined as the first available measurement after hospital admission but before the receipt of a PD-1 inhibitor plus platinum-based dual-agent chemotherapy to ensure that the independent variables were captured strictly before the onset of the myelosuppression outcome measure. Data collection was performed strictly according to the prespecified inclusion and exclusion criteria. The inclusion criteria were: (I) pathologically confirmed NSCLC; (II) age 18–80 years; and (III) receipt of at least one cycle of PD-1 inhibitor plus platinum-doublet chemotherapy. The exclusion criteria were: (I) missing key data; (II) severe organ failure; (III) pregnancy or lactation; and (IV) a history of severe allergy to combined immunotherapy and chemotherapy. Collected features included: (I) general characteristics, such as sex, age, and body mass index (BMI); (II) medical history, such as diabetes, hypertension, and smoking history; (III) treatment-related variables, such as treatment regimen, bone metastasis, and tumor size; and (IV) hematological and biochemical indices, such as RBC, WBC, red cell distribution width (RDW), and total protein (TP). Outcome measure According to guideline recommendations (25), myelosuppression was defined as a positive outcome if the minimum values of the relevant indicators within 21 days after PD-1 inhibitor plus platinum-doublet chemotherapy met the criteria for grade ≥2 myelosuppression. Specifically, the thresholds were as follows: neutrophil count <1.5×109/L, WBC <3.0×109/L, platelet count <75×109/L, and hemoglobin <100 g/L. Model development and evaluation Patients treated before January 2025 were randomly divided into a training set (70%) and a test set (30%). In the training set, features selected by at least two of the following three algorithms, including RFE, LASSO, and RF, were retained as the final predictors. Multiple ML models were constructed, and feature importance was ranked within each model. Model performance was evaluated using the receiver operating characteristic (ROC) curve, recall, sensitivity, and other metrics, with AUC as the primary index. Calibration curves and decision curve analysis (DCA) were also used to identify the optimal model. SHAP was applied to quantify the marginal contributions of individual features to the overall prediction and model output. Statistical analysis The training set was first preprocessed. Samples with more than 20% missing values were excluded to reduce bias; otherwise, missing values were imputed using predictive mean matching (PMM). The specific parameters were set as follows: the imputation method was PMM; the number of imputation was 5 times (m=5). The number of iterations was 5. All continuous variables with outliers were Winsorization, with the upper and lower limits truncated by 1% to minimize the effect of extreme values. Data standardization: Before model training, all continuous variables were Z-score standardized, and the standardized parameters were calculated based on the training set only and applied synchronously to the test set and external validation set to avoid data leakage. Categorical variables were directly included in the model after factorization without standardization. Duplicate records were removed. The preprocessing strategy derived from the training set was then applied to the test set and external validation set. The sample imbalance issue is addressed by using the Synthetic Minority Over-sampling Technique (SMOTE) algorithm to achieve sample balance. Feature selection was performed using RFE, RF, and LASSO, and 5-fold cross-validation during model development. Continuous variables are presented as mean ± standard deviation, and categorical variables as frequency and percentage. Categorical variables were compared using the chi-square test. Continuous variables were compared using the t-test when the assumptions of normality and homogeneity of variance were satisfied; otherwise, the Mann-Whitney U test was used. A two-sided P<0.05 was considered statistically significant. All figures and tables were generated using R version 4.5.2 (R Foundation for Statistical Computing, Vienna, Austria). Results Baseline characteristics Univariate analysis was first performed on the 235 patients in the training set. Significant differences between the case and control groups were observed in tumor size, age, BMI, platelet count, basophil count, TP, serum potassium, and serum calcium. No statistically significant differences were observed for the remaining features. Detailed results of the univariate analysis are presented in Table 1, and the study flowchart is shown in Figure 1. These results suggest that the patient’s age, BMI, tumor size and abnormal inflammatory/immune-related indicators in peripheral blood may jointly constitute a susceptible background for bone marrow suppression in NSCLC patients receiving PD-1 inhibitor combined with platinum-based doubuse chemotherapy, and provide a certain reference for the subsequent construction of a multi-dimensional risk model. Table 1Univariate analysis results for the training setFeaturesNon-myelosuppression (n=210)Myelosuppression (n=25)P valueGender (female) 54 (25.7)5 (20.0)0.71Smoking history (yes)106 (50.5)12 (48.0)0.98Drinking history (yes) 161 (76.7)20 (80.0)0.90Diabetes (yes) 24 (11.4)3 (12.0)>0.99Hypertension (yes) 162 (77.1)19 (76.0)>0.99Hyperlipidemia (yes)4 (1.9)1 (4.0)>0.99Arrhythmia (yes) 8 (3.8)1 (4.0)>0.99COPD (yes)26 (12.4)3 (12.0)>0.99Bone metastasis (yes) 52 (24.8)8 (32.0)0.59Bone protection (yes) 40 (19.0)5 (20.0)>0.99PD1_platinum_gemcitabine (yes) 0 (0.0)1 (4.0)0.20PD1_platinum_etoposide (yes) 73 (34.8)8 (32.0)0.96PD1_platinum_pemetrexed (yes) 32 (15.2)4 (16.0)>0.99PD1_platinum_paclitaxel (yes) 101 (48.1)11 (44.0)0.86Radiotherapy (yes) 32 (15.2)4 (16.0)>0.99Surgery (yes) 49 (23.3)6 (24.0)>0.99Stage (III-IV) (yes) 177 (84.3)23 (92.0)0.33Squamous cell carcinoma82 (39.0)9 (36.0)0.82Adenocarcinoma126 (60.0)16 (64.0)0.71Large cell2 (1.0)0 (0.0)0.89Tumor size (mm)41.56±21.1764.01±46.76<0.001Age (years)60.24±8.1865.88±11.240.002BMI (kg/m2)22.79±3.0324.99±3.790.001Procalcitonin (μg/L)0.18±0.080.17±0.100.47Interleukin-6 (pg/mL)233.82±314.24213.40±291.890.76RBC (×109/L)4.62±0.504.45±0.700.13Hemoglobin (g/dL)141.53±16.09137.88±21.450.30WBC (×109/L)6.29±2.146.76±2.070.30Platelets (×109/L)128.70±94.57182.88±108.110.008MCV (fL)92.91±4.7793.56±4.890.52Lymphocyte (×109/L)0.94±0.660.80±0.530.30Monocyte (×109/L)0.94±0.631.03±0.540.53Neutrophil (×109/L)2.36±2.381.62±1.990.14Eosinophil (×109/L)2.48±2.732.83±2.320.54Basophil (×109/L)0.07±0.090.14±0.170.001RDW (%)6.61±6.754.74±7.070.20INR1.07±0.101.08±0.110.48Prothrombin time (s)11.67±0.9811.84±1.180.43Prothrombin activity (%)92.07±11.4691.16±12.480.71D dimer (μg/L)1.11±1.240.67±0.600.08CRP (mg/L)27.80±27.0629.44±30.280.78AST (U/L)24.20±11.0226.84±9.140.25ALT (U/L)25.88±17.9629.81±19.940.31TBIL (μmol/L)15.73±11.4814.41±7.460.57Total protein (g/L)66.90±9.0772.30±11.080.007Albumin (g/L)41.91±4.1743.28±4.210.12Globulin (g/L)29.06±4.8329.36±4.460.77BUN (mg/dL)5.44±1.775.41±1.380.94Creatinine (μmol/L)67.64±13.4370.53±10.990.30Potassium (mmol/L)4.01±0.394.22±0.340.01Sodium (mmol/L)139.14±2.91138.69±3.720.48Chloride (mmol/L)104.32±3.47103.97±4.170.64Calcium (mmol/L)2.20±0.132.28±0.120.005Data are presented as mean ± standard deviation or frequency (percentage). ALT, alanine aminotransferase; AST, aspartate aminotransferase; BUN, blood urea nitrogen; COPD, chronic obstructive pulmonary disease; CRP, C-reactive protein; INR, international normalized ratio; MCV, mean corpuscular volume; RBC, red blood cell count; RDW, red cell distribution width; TBIL, total bilirubin; WBC, white blood cell count. Univariate analysis results for the training set Data are presented as mean ± standard deviation or frequency (percentage). ALT, alanine aminotransferase; AST, aspartate aminotransferase; BUN, blood urea nitrogen; COPD, chronic obstructive pulmonary disease; CRP, C-reactive protein; INR, international normalized ratio; MCV, mean corpuscular volume; RBC, red blood cell count; RDW, red cell distribution width; TBIL, total bilirubin; WBC, white blood cell count. Research mechanism diagram. AUC, area under the curve; BMI, body mass index; GBM, gradient boosting machine; LightGBM, light gradient boosting machine; NNET, neural network; RBC, red blood cell; RDW, red cell distribution width; SVM, support vector machine; WBC, white blood cell. Feature selection Using a combined strategy based on LASSO, RFE, and RF in the training set, we identified age, BMI, tumor size, platelet count, RDW, TP, WBC, and RBC as risk factors for myelosuppression in patients with NSCLC receiving PD-1 inhibitor plus platinum-doublet chemotherapy (Figure 2A-2E). This means that patients with advanced age, poor nutritional status, larger tumor diameter, and abnormal hematological indicators such as platelet, WBC, RBC, RDW and TP may be more prone to bone marrow suppression. Flow chart of feature selection. (A) Binomial deviation curve of LASSO regression; (B) coefficient shrinkage path plot of LASSO regression; (C) importance ranking plot of RF features; (D) RFE cross-validation accuracy curve; (E) feature selection Venn diagram. BMI, body mass index; LASSO, least absolute shrinkage and selection operator; RBC, red blood cell; RDW, red cell distribution width; RFE, recursive feature elimination; TP, total protein; WBC, white blood cell. Model performance in the training set Based on the selected features, multiple ML models were developed. Their performance in the training set was evaluated using the AUC as the primary metric, along with calibration curves and DCA as supplementary measures. The AUC values were as follows: SVM, 0.889 [95% confidence interval (CI): 0.844–0.934]; GBM, 0.890 (95% CI: 0.849–0.932); NNET, 0.828 (95% CI: 0.777–0.879); XGBoost, 0.880 (95% CI: 0.837–0.924); LightGBM, 0.898 (95% CI: 0.857–0.938); and CatBoost, 0.891 (95% CI: 0.850–0.933). Among these models, LightGBM showed the best performance, whereas NNET performed the worst, After grid optimization combined with 5-fold cross-validation, the hyperparameters of the LightGBM model were determined as follows: learning rate =0.02, max depth =3, num leaves =5, min data in leaf =30, subsample =0.5, colsample bytree =0.5, nrounds =172L The ROC curves for the training set are shown in Figure 3A, the DCA curves in Figure 3B, the predicted probability distributions in Figure 3C, the forest plot of 95% CIs in Figure 3D, the visualization of evaluation metrics in Figure 3E, and The calibration curve of the training set is shown in 3F. The feature importance rankings in Figure 4A-4F. The performance indicators of each ML model are shown in Table 2. Evaluation results of model performance in the training set. (A) ROC curve of training set, the performance of the model is mainly evaluated by the area under the curve AUC, among which the AUC of LightGBM is optimal up to 0.898. (B) DCA curve of training set, curve showed that the model was higher than the full intervention and no intervention thresholds in most risk threshold ranges, indicating good clinical net benefits. (C) Forecast probability distribution diagram of training set. Each model has a strong ability to distinguish between negative and positive samples, among which the LightGBM model has the most obvious separation degree of prediction probability of negative and positive samples. (D) Forest diagram of AUC 95% CI of training set, which further shows that the AUC of LightGBM model is in the best state. (E) Histogram of model performance index of training set. In addition to AUC, other performance indicators such as accuracy and sensitivity were visually compared, and the LightGBM model still had the best performance. (F) The calibration curve of the training set model, reflecting the calibration ability of each model, among which the LightGBM model was closer to the diagonal line. AUC, area under the curve; CatBoost, categorical boosting; CI, confidence interval; GBM, gradient boosting machine; LightGBM, light gradient boosting machine; NNET, neural network; ROC, receiver operating characteristic; SVM, support vector machine; XGBoost, eXtreme gradient boosting. Graph showing the ranking of the importance of internal features of the model. Internal feature importance ranking for (A) SVM, (B) GBM, (C) NNET, (D) XGBoost, (E) LightGBM, (F) CatBoost. BMI, body mass index; CatBoost, categorical boosting; GBM, gradient boosting machine; LightGBM, light gradient boosting machine; NNET, neural network; RBC, red blood cell; RDW, red cell distribution width; SVM, support vector machine; TP, total protein; WBC, white blood cell; XGBoost, eXtreme gradient boosting. Table 2Results of predictive model performance evaluationModelDataAUCAccuracySensitivitySpecificityPrecisionF1 scoreSVMTraining set0.8890.8840.7840.9840.9800.871Test set0.7360.6100.9000.5780.1910.316GBMTraining set0.8900.9200.9360.9040.9070.921Test set0.7650.6500.9000.6220.2090.340NNETTraining set0.8280.2120.2480.1760.2310.239Test set0.7260.3900.9000.3330.1300.228XGBoostTraining set0.8800.8560.7760.9360.9240.843Test set0.7170.6100.8000.5890.1780.291LightGBMTraining set0.8980.8800.8240.9360.9280.873Test set0.8410.7700.8000.7670.2760.410Validation set0.7930.8000.8460.7880.5000.629CatBoostTraining set0.8910.8600.7920.9280.9170.850Test set0.8290.5700.9030.5220.1890.317AUC, area under the curve; CatBoost, categorical boosting; GBM, gradient boosting machine; LightGBM, light gradient boosting machine; NNET, neural network; SVM, support vector machine; XGBoost, eXtreme gradient boosting. Results of predictive model performance evaluation AUC, area under the curve; CatBoost, categorical boosting; GBM, gradient boosting machine; LightGBM, light gradient boosting machine; NNET, neural network; SVM, support vector machine; XGBoost, eXtreme gradient boosting. Model performance in the test set and external validation In the test set, the AUC values were as follows: SVM, 0.736 (95% CI: 0.544–0.927); GBM, 0.765 (95% CI: 0.617–0.913); NNET, 0.726 (95% CI: 0.536–0.916); XGBoost, 0.717 (95% CI: 0.559–0.875); LightGBM, 0.841 (95% CI: 0.693–0.988); and CatBoost, 0.829 (95% CI: 0.707–0.950). LightGBM again showed the best performance, The performance degradation of the model (only about 0.057 from the training set) was the smallest among all the models, showing superior robustness, while XGBoost performed the worst. The ROC curves for the test set are shown in Figure 5A, the DCA curves in Figure 5B, the predicted probability distributions in Figure 5C, the forest plot of 95% CIs in Figure 5D, the visualization of evaluation metrics in Figure 5E, and calibration curves for the test set are shown in Figure 5F. In the external validation cohort, the LightGBM model achieved an AUC of 0.793 (95% CI: 0.672–0.914), Although the external validation performance decreased slightly, this result was within the clinically acceptable range of discrimination, and the stable performance of the model across time periods initially confirmed the reliability of the tool in the actual clinical field, reducing the concern of overfitting. The external validation results are shown in Figure 6A-6E. Evaluation results of model performance in the test set. (A) ROC curve of test set, in which the AUC of LightGBM model reached 0.841, which was the best model. (B) DCA curve of test set. LightGBM maintained a high level of the curve within all risk thresholds, corresponding to high clinical net benefits. (C) Forecast probability distribution diagram of test set, LightGBM model also shows a higher forecast probability separation state. (D) Forest diagram of AUC 95% CI of test set, which also showed the best AUC value and 95% CI in the test set. (E) Histogram of model performance index of test set. LightGBM’s model performance advantage is also reflected in F1 score and accuracy. (F) The calibration curve of the test set, LightGBM also showed good calibration in the test set. AUC, area under the curve; CatBoost, categorical boosting; CI, confidence interval; GBM, gradient boosting machine; LightGBM, light gradient boosting machine; NNET, neural network; ROC, receiver operating characteristic; SVM, support vector machine; XGBoost, eXtreme gradient boosting. Evaluation results of model performance in the validation set. (A) ROC curve of validation set, LightGBM model also showed the best performance in the external validation with an AUC of 0.793. (B) DCA curve of validation set, the curve was mostly above the baseline of full intervention and no intervention. (C) Forecast probability distribution diagram of validation set, which indicates that LightGBM has strong ability to separate negative and positive samples. (D) Forest diagram of AUC 95% CI of validation set, which again reflects that LightGBM model has the best performance. (E) Histogram of model performance index of validation set. The model also performed well in other performance indicators in the external validation. AUC, area under the curve; CI, confidence interval; LightGBM, light gradient boosting machine; ROC, receiver operating characteristic. SHAP-based model interpretation To address the limited interpretability of ML models, SHAP was applied to provide an in-depth interpretation of the optimal model. In the LightGBM model, SHAP feature importance ranking indicated that age, BMI, tumor size, platelet count, RDW, TP, WBC, and RBC contributed to the model in descending order (Figure 7A). Additional SHAP visualizations included the summary plot (Figure 7B), force plots (Figure 7C,7D), the SHAP feature dependence plot visualizes the relationship between each variable and the model’s predictive contribution. For instance, age, BMI, and TP all exhibit a clear threshold effect, with their positive predictive contribution significantly increasing after exceeding a specific threshold (Figure 7E). The cumulative contribution curve indicates that the top four most important features (age, BMI, tumor size, and platelets) contribute the majority of the model’s predictive information, and the marginal contribution of subsequent features gradually decreases (Figure 7F). The sample-feature heatmap intuitively displays the feature contribution patterns of different patient groups (Figure 7G). The average absolute bar chart quantifies and validates the stability of each feature’s contribution, where age, BMI, tumor size, and platelets not only have the highest contribution but also show the most consistent effect across different samples (Figure 7H). SHAP Visualization. (A) SHAP importance bar plot. (B) SHAP summary plot, mainly show that age, BMI and tumor size are the three most influential variables in LightGBM model prediction, and the contribution of other blood indicators decreases in turn. (C,D) SHAP single sample effort plot, which specifically illustrates the contribution of single sample internal characteristics to the final prediction results. (E) SHAP scatter plot, there is an obvious critical inflection point for each index, and the SHAP value changes greatly after exceeding the threshold, such as the higher the age and the larger the tumor, the risk of disease continues to increase. (F) SHAP cumulative contribution plot, which further verified that the predicted contribution of the model was highly concentrated in the three core variables of age, BMI and tumor size, and the cumulative contribution of the remaining indicators was very low. (G) SHAP heat map, the risk-driven characteristics of different samples are significantly different, the high SHAP red line corresponds to high-risk individuals, mainly driven by high-contribution features such as age and tumor size. (H) SHAP mean bar plot, age, BMI and tumor size not only had the highest average influence, but also had greater fluctuations in the effect strength between different samples, which were the core factors to distinguish the risk of the population. LightGBM, light gradient boosting machine; BMI, body mass index; RDW, red cell distribution width; RBC, red blood cell; TP, total protein; SHAP, Shapley Additive Explanations; WBC, white blood cell. Discussion In this study, we included 335 patients with NSCLC from the Department of Respiratory Medicine, The First Hospital of Lanzhou University, all of whom received PD-1 inhibitor plus platinum-doublet chemotherapy. Among them, 35 patients (10.45%) developed myelosuppression, which is lower than the previously reported incidence of 36.47% (26). One possible explanation is that grade 1 myelosuppression typically does not require clinical intervention; therefore, our outcome definition included only grade ≥2 myelosuppression. In addition, we performed external validation using a temporally distinct cohort of 65 patients. Using RF, LASSO, and RFE for feature selection, we identified age, BMI, tumor size, platelet count, RDW, TP, WBC, and RBC as key risk factors for myelosuppression in patients with NSCLC receiving PD-1 inhibitor plus platinum-doublet chemotherapy. Among all constructed ML models, the LightGBM showed the best performance, with AUCs of 0.898, 0.841, and 0.793 in the training, test, and external validation sets, respectively. In addition, DCA showed that the net clinical benefit of the model was better than that of “all intervention” or “none intervention” strategies in most coverage ranges of threshold probability, indicating that the model could effectively guide preventive intervention decision-making without increasing additional clinical burden, and had practical clinical application value. LightGBM employs a histogram-based algorithm and a leaf-wise growth strategy, enabling efficient capture of complex interactions among high-risk features. It also incorporates gradient-based one-side sampling to focus on difficult-to-classify samples, thereby improving the identification of minority-class myelosuppression events. In addition, its built-in regularization and efficient handling of complex features may contribute to its superior performance in the present study. In order to further promote the practical application of our research results, we built an interactive web risk prediction tool based on the LightGBM model developed in this study, which has been deployed on the shinyapps.io cloud platform. Clinicians can directly by the following link to access and use the tool: . Li et al. (27). predicted the risk of myelosuppression in patients with LC receiving platinum-doublet chemotherapy in 2025; however, their study neither considered the use of PD-1 inhibitors nor included external validation. Although Dong et al. (26). applied ML methods to predict chemotherapy related adverse events, only a single RF model was used. In 2026, Chen et al. (28). developed a model to predict chemotherapy-related adverse events in patients with LC; however, their approach relied only on a nomogram rather than ML methods. To address these limitations, we employed RF, LASSO, and RFE for feature selection and developed multiple ML models, including SVM, GBM, NNET, XGBoost, LightGBM, and CatBoost, to predict the risk of myelosuppression in patients with NSCLC receiving PD-1 inhibitor plus platinum-doublet chemotherapy. Older patients were more likely to develop myelosuppression. One possible explanation is that impaired renal and hepatic function in older adults reduces the clearance of chemotherapeutic agents, thereby increasing systemic toxicity (29). In addition, older patients are often in a state of chronic inflammation, which may further suppress bone marrow hematopoietic function (1). These findings are consistent with previous reports (27). Patients with low body weight are more likely to be malnourished, which may reduce bone marrow reserve capacity and decrease the number of hematopoietic stem cells. Moreover, insufficient protein and nutrient intake may impair bone marrow repair, thereby increasing the risk of myelosuppression (30,31). This observation is also supported by previous studies (32). Larger tumors are generally associated with higher chemotherapy doses and frequencies and may also release increased levels of inflammatory mediators that suppress bone marrow hematopoiesis. Lower platelet counts and WBC levels in patients with NSCLC often indicate diminished bone marrow reserve, while concomitant bleeding and infection may further reduce tolerance to chemotherapy (33). Low RBC levels combined with elevated RDW may indicate decompensated bone marrow reserve with ongoing ineffective hematopoiesis, thereby increasing the risk of myelosuppression. In addition, low TP levels may increase the free fraction of chemotherapeutic agents, thereby enhancing toxicity; inadequate protein availability is also associated with impaired bone marrow repair (34). Limitations Several limitations of this study should be acknowledged. First, as a single-center study, it may be subject to selection bias. Second, the relatively small sample size may have increased the risk of type II error. Third, due to limitations in data availability, we were unable to assess dynamic changes in clinical features over time; therefore, trajectory-based analyses were not performed. Conclusions We developed and validated multiple ML models to predict the risk of myelosuppression in patients with NSCLC receiving PD-1 inhibitor plus platinum-doublet chemotherapy. Among these models, LightGBM demonstrated the best predictive performance and may serve as a valuable tool for risk stratification in this population. Supplementary The article’s supplementary files as