Abstract
Background
As immune checkpoint inhibitors (ICI)-based combination therapies are increasingly explored for treating NK/T-cell lymphoma (NKTCL), there is a critical clinical need to identify patients who will benefit from ICI without relying on costly genomic testing.
Methods
A machine learning model was developed using routine blood tests and clinical characteristics from 364 ICI-treated NKTCL patients. The case records of 1259 NKTCL patients discharged from Sun Yat-Sen University Cancer Center, Guangzhou, between January 2018 and December 2023 were retrospectively analyzed. After screening, 364 ICI-treated patients were included in the study. These patients were randomly assigned to training (n = 254) and validation (n = 110) cohorts in a 2:1 ratio. Lasso regression and five machine learning algorithms, including random forest (RF), were applied for feature selection and clinical benefit prediction. The RF model demonstrated optimal predictive performance using five key features. To predict overall survival, we combined the RF model with two critical clinical factors-Ann Arbor stage and Eastern Cooperative Oncology Group (ECOG) performance status-to develop the stage-ECOG-RF (SER) model. This model generates a risk score to quantify the probability of poor survival following ICI treatment. In total, the SER model including seven features is significantly associated with clinical outcomes and long-term survival.
Results
Five feature variables-lymphocyte count, platelet count, bone marrow involvement, cholesterol (CHO), and EBV-DNA copy number-were selected from 23 laboratory tests and clinical characteristics with complete data (0% missing rate). In the training cohort, the RF algorithm showed an area under the receiver operating characteristic curve (AUC) of 0.878, outperforming extreme gradient boosting (XGBoost), support vector machine (SVM), decision trees (DT), logistic regression and SVM algorithms. The RF model demonstrated sensitivity of 83.3% and specificity of 78.9%. In the validation cohort, the AUC of the RF model was 0.752, with sensitivity of 68.8% and specificity of 69.1%. The SER model, which integrates the RF model with Ann Arbor stage and ECOG, attained time-dependent area under the receiver operating characteristic curve (AUC(t)) values of 0.736 and 0.650 for predicting 3- and 5-year overall survival. This surpasses the prognostic index of natural killer lymphoma (PINK-E) and the net reclassification index (NRI) models, which showed AUC(t) values of 0.722 and 0.532, and 0.707 and 0.541 at 3 and 5 years, respectively.
Conclusions
Based on routine blood tests and clinical data, the SER model for ICI therapy of NKTCL-optimized with the RF algorithm and incorporating Ann Arbor stage and ECOG-demonstrates superior predictive performance compared to PINK-E and NRI. It provides a valuable reference for early prediction of ICI therapy failure and long-term survival.
Keywords: Prognostic biomarker, PD-1, NKTCL, ICI, Machine learning
Introduction
NK/T-cell lymphoma is a rare and highly aggressive subtype of non-Hodgkin lymphoma (NHL) [1], which is closely related to Epstein-Barr virus (EBV) infection [2]. NKTCL shows marked a geographical variation in incidence, with a higher prevalence in Asia and Latin America [3]. NKTCL can be classified into two types, based on the site of primary tumor, as upper aerodigestive tract NKTCL (UAT-NKTCL) and non-upper aerodigestive tract NKTCL (NUAT-NKTCL) [4] clinical subtypes. UAT-NKTCL predominantly occurs in the nasal and paranasal regions (e.g., nasopharynx, nasal sinuses, Waldeyer’s ring, and oropharynx), accounting for over 80% of NKTCL cases, while NUAT-NKTCL primarily affects extranasal sites such as the skin, gastrointestinal tract, testes, liver, and lungs, representing 10–20% of cases but with a higher malignancy and poorer prognosis.
Immunohistochemical markers [5] for the diagnosis of NKTCL include cytoplasmic CD3ε, cytotoxic markers (TIA-1, Granzyme B, Perforin), and EBER in situ hybridization [6]. NKTCL primarily originates from the transformation of natural killer (NK) cells, with approximately 20% arising from cytotoxic T-cells [7]. The typical immunophenotype of NKTCL of NK-cell origin is characterized by cytoplasmic CD3ε positivity and CD56 positivity, while NKTCL of T-cell origin typically exhibit CD56 negativity. Other T-cell and NK-cell associated antigens, such as CD4, CD8, CD5, and CD7, show variable expressions.
Early-stage NKTCL patients treated with standard radiotherapy and chemotherapy achieve a 5-year overall survival (OS) rate of 70–80% [8], while patients with advanced or relapsed/refractory disease have poor outcomes, with a median OS of only 6 months after first-line non-anthracycline-based therapy [9]. Several small-sample studies have demonstrated that programmed death receptor 1 (PD1)/programmed death ligand 1 (PD-L1) inhibitors show potential antitumor activity in relapsed/refractory NKTCL, with single-agent overall response rate (ORR) ranging from 38.0% to 75.0% [10–12]. In non-treatment patients with early-stage NKTCL, the combination of ICI, chemotherapy with radiotherapy, demonstrated promising efficacy and a favorable safety profile. The overall response rate reached 87.8% [13]. The sintilimab plus P-GEMOX regimen (pegaspargase, gemcitabine, and oxaliplatin) showed promising clinical activity in treating advanced extranodal natural killer/T cell lymphoma (ENKTL), with an ORR of 100% in 34 efficacy-evaluable patients [14]. While the results of the aforementioned clinical trials are promising, the applicability and precise efficacy of ICI in combination with asparaginase-based chemotherapy for patients with varying stages and types of NKTCL remain to be confirmed through larger-scale studies. Ongoing clinical trials are actively addressing these questions. Our study aims to investigate potential biomarkers that could enhance the prediction of treatment efficacy and long-term survival outcomes in the context of immunotherapy combined with chemotherapy.
Machine learning (ML), an emerging statistical analysis method, can extract valuable insights from large datasets, thereby enhancing the precision of disease diagnosis and prognosis. Recently, researchers developed and evaluated SCORPIO [15], a machine learning system that predicts the prognosis and long-term survival of cancer patients treated with immune checkpoint inhibitors by using routine laboratory tests and clinical characteristics from 9745 ICI-treated patients across 21 cancer types. This system provides a valuable tool for predicting ICI outcomes without requiring complex genomic testing. Nevertheless, SCORPIO does not include EBV-associated cancers, such as nasopharyngeal carcinoma and NKTCL.
In this study, we trained and tested a predictive SER model, combined with RF machine learning model, Ann Arbor stage and ECOG, on 364 ICI-treated NKTCL patients to predict efficacy of immunotherapy and long-term survival.
Methods
Study participants inclusion
From March 2018 to December 2023, the retrospective study screened a total of 1259 NKTCL patients from Guangzhou Sun Yat-sen University Cancer Center. We retrospectively collected baseline clinical and laboratory data from 459 NKTCL patients prior to their first PD-1 inhibitor treatment. After excluding patients with incomplete medical records, those lost to follow-up, and those with missing baseline laboratory data, we ultimately included 364 patients for subsequent analysis and randomly grouped them into training (n = 254) and validation (n = 110) cohorts in a 2:1 ratio.
Baseline features collection
The baseline clinical features including age, gender, Ann Arbor stage, Eastern Cooperative Oncology Group (ECOG) Performance Status, multiple extranodal involvement, UAT/NUAT-NKTCL, accompanying with hemophagocytic lymphohistiocytosis (HLH), bone marrow involvement and primary tumor invasion (PTI) were collected. Bone marrow involvement is confirmed by bone marrow biopsy, smear and flow cytometric analysis. Laboratory features including pathological markers CD56, TIA-1, CD2, CD5, CD7, CD4, CD8, CD30, CD20, Granzyme B(GRB), Ki-67 and serum markers white blood count (WBC), lymphocyte count (LYM), neutrophil count (NE), platelet count (PLT), fibrinogen (Fbg), hemoglobin (Hb), lactate dehydrogenase (LDH), creatinine (CRE), alanine amino transferase (ALT), aspartate amino transferase (AST), triglycerides (TG), cholesterol (CHO), C-reaction protein (CRP), albumin, ferritin, EBV DNA copy number, β2-microglobulin, cytokines IL-2, IL-4, IL-6, IL-10, TNF-α, and IFN-γ.
Features selection
According to the RECIST 1.1 criteria, we evaluated therapeutic efficacy after the 2nd and 4th cycles of ICI treatment, classifying patients into progression disease (PD) and non-PD. Lasso regression was used to select key predictive parameters associated with the efficacy outcome from numerous variables. The Random Forest (RF) machine learning algorithm further assessed and confirmed the significance of these parameters. Finally, SHapley Additive exPlanations (SHAP) values were calculated to determine each parameter’s contribution to the model’s predictions, ensuring the effective identification of critical model parameters.
Development of the machine learning model
The filtered features were used to train and validate five machine learning models across training and validation cohorts including extreme gradient boosting (XGBoost), RF, support vector machine (SVM), decision trees (DT), and logistic regression. These performances were assessed using AUC, sensitivity, and specificity to determine the optimal algorithm for predicting treatment outcomes.
Nomogram development of SER prognosis model
To enhance the prognostic performance of the RF model, we performed univariate Cox regression analysis and multivariate Cox regression analysis to identify potential prognostic factors associated with ICI efficacy and survival outcomes. Variables such as age, Ann Arbor stage, ECOG, LDH levels, and PTI, which are components of the NLR and PINK-E prognostic models, were included in the analysis. Based on the results of the univariate and multivariate Cox regression analysis, a nomogram was developed to predict survival probabilities at 1, 3, and 5 years. The nomogram integrates the independent prognostic factors, assigning weights (points) to each variable according to its impact on survival. The total points for each patient are calculated by summing the points for each variable, and the corresponding survival probabilities are derived from the total points. The nomogram was validated using bootstrapping methods to assess its accuracy and reliability. Discrimination and calibration were evaluated to ensure the nomogram’s predictions closely matched actual survival outcomes.
Prognosis model comparison
To compare the prognostic performance of the SER, PINK-E, and NRI models, we utilized Concordance Index (C-index) analysis and Time-Dependent Receiver Operating Characteristic (ROC) analysis. Time-dependent ROC curves were plotted at specific time points (1, 3 and 5 years) to assess model performance at predicting survival at these intervals. The area under the curve quantifies the model’s accuracy, with higher AUC values, indicates better predictive ability. This analysis clearly shows the value of SER model in assessing the prognoses of NKTCL patients at different follow-up stages.
Results
Characteristics of patients
The study cohort comprised 364 patients, with 254 allocated to the training group and 110 to the validation group (Fig. 1). The median age was 45 years (range 3–93). The majority of patients were male (68.96% in the total cohort). Most patients had Ann Arbor Stage I-II disease (66.93% in the training cohort and 70.00% in the validation cohort), indicating an early-stage NKTCL. Multiple extranodal involvement was present in 29.13% and 24.55% of the training and validation cohorts, respectively. Upper aerodigestive tract (UAT) involvement was common, with 83.86% in the training group and 89.09% in the validation cohort. In the total cohort, bone marrow involvement was observed in 7.69% of patients. The incidence rate of hemophagocytic lymphohistiocytosis (HLH) was 6.87% in the total cohort. Most patients had an ECOG score of 0–1 (83.07% in the training cohort and 79.09% in the validation cohort). In terms of chemotherapy regimens, the distribution of p-GEMOX-based, pegaspargase-based, and chidamide-based regimens were consistent across the total cohort, training cohort, and validation cohort. After the 2nd or 4th cycle of ICI treatment, 14.29% of patients were classified as progression (PD), while 85.71% were classified as non-PD (Table 1).
Fig. 1.
Workflow of study and model development
Table 1.
Clinical characteristics of NKTCL patients
| Characteristic | Total cohort (n = 364) | Training cohort (n = 254) | Validation cohort (n = 110) | P value |
|---|---|---|---|---|
| No. (%) | No. (%) | No. (%) | ||
| Age, y | 0.0492 | |||
| Median (IQR) | 45 (36–57) | 48 (37–57) | 42 (34–56) | |
| Sex | 0.7115 | |||
| Male | 251 (68.96) | 177 (69.69) | 74 (67.27) | |
| Female | 113 (31.04) | 77 (30.31) | 36 (32.73) | |
| ANN ARBOR Stage | 0.6255 | |||
| 1–2 | 247 (67.86) | 170 (66.93) | 77 (70.00) | |
| 3–4 | 117 (32.14) | 84 (33.07) | 33 (30.00) | |
| Multiple extranodal involvement | 0.4445 | |||
| Yes | 101 (27.75) | 74 (29.13) | 27 (24.55) | |
| No | 263 (72.25) | 180 (70.87) | 83 (75.45) | |
| UAT | 0.2569 | |||
| Yes | 311 (85.44) | 213 (83.86) | 98 (89.09) | |
| No | 53 (14.56) | 41 (16.14) | 12 (10.91) | |
| Bone marrow or intracranial involvement | 0.0761 | |||
| Bone marrow | 20 (5.49) | 12 (4.72) | 8 (7.27) | |
| Bone marrow and intracrania | 8 (2.20) | 3 (1.18) | 5 (4.55) | |
| None | 336 (92.31) | 239 (94.09) | 97 (88.18) | |
| HLH | 0.2676 | |||
| Yes | 25 (6.87) | 15 (5.91) | 10 (9.09) | |
| No | 339 (93.13) | 239 (94.09) | 100 (90.91) | |
| ECOG | 0.3766 | |||
| 0–1 | 298 (81.87) | 211 (83.07) | 87 (79.09) | |
| 2–4 | 66 (18.13) | 43 (16.93) | 23 (20.91) | |
| Chemotherapy Regimens | 0.5058 | |||
| P-GEMOX-based regimens | 109 (29.95) | 77 (30.31) | 32 (29.09) | |
| Pegaspargase-based regimens | 114 (31.32) | 84 (33.07) | 30 (27.27) | |
| Chidamide-based regimens | 113 (31.04) | 73 (28.74) | 40 (36.36) | |
| Others | 28 (7.69) | 20 (7.87) | 8 (7.27) | |
| ICI outcomes | > 0.9999 | |||
| PD | 52 (14.29) | 36 (14.17) | 16 (14.55) | |
| non-PD | 312 (85.71) | 218 (85.83) | 94 (85.45) |
UAT Upper aerodigestive tract; HLH Hemophagocytic lymphohistiocytosis; ECOG Eastern Cooperative Oncology Group; ICI Immune checkpoint inhibitors; PD Progression disease
Feature selection and development of the machine learning model
To identify crucial features for ICI treatment outcomes, we conducted feature selection analyses. Table 2 presents the laboratory data summary. Considering of the missing rate, we excluded specific laboratory data from the machine learning and feature selection processes. The RF model’s feature importance plot revealed LYM and EBV DNA as the most influential, followed by PLT, bone marrow involvement, and CHO (Fig. 2a). SHAP analysis also highlighted these features, showing their significant impact on model predictions (Fig. 2b). LASSO regression selected features like CD56, ECOG, UAT/NUAT, and bone marrow involvement. After integrating RF, SHAP, and LASSO analyses, the top 5 features, LYM, EBV DNA, PLT, CHO, and bone marrow involvement, were determined and used for further analysis.
Table 2.
Laboratory features of NKTCL patients
| Laboratory features | Total cohort | Training cohort | Validation cohort | P value |
|---|---|---|---|---|
| (n = 364) | (n = 254) | (n = 110) | ||
| CBC, median, (IQR) | ||||
| WBC (10E9/L) | 5.84 (4.27–7.46) | 5.73 (4.06–7.53) | 6.04 (4.74–7.30) | 0.5089 |
| NE (10E9/L) | 3.67 (2.64–5.21) | 3.64 (2.58–5.24) | 3.71 (2.78–5.14) | 0.4460 |
| LYM (10E9/L) | 1.30 (0.89–1.76) | 1.29 (0.90–1.75) | 1.32 (0.85–1.85) | 0.9166 |
| Hb (g/L) | 127.00 (112.00–141.00) | 126.00 (110.00–141.00) | 130.00 (114.80–141.30) | 0.5882 |
| PLT (10E9/L) | 243.50 (186.00–303.80) | 244.00 (186.80–301.80) | 240.50 (178.50–304.30) | 0.9214 |
| Biochemical Parameters, median, (IQR) | ||||
| CHO (mmol/L) | 4.52 (3.93–5.38) | 4.56 (4.00–5.38) | 4.47 (3.82–5.35) | 0.4981 |
| TG (mmol/L) | 1.46 (1.08–2.12) | 1.49 (1.10–2.14) | 1.39 (1.02–2.09) | 0.4140 |
| CRP (mgl/L) | 6.17 (2.14–21.82) | 6.19 (2.18–21.90) | 6.15 (1.83–22.52) | 0.6581 |
| Albumin (g/L) | 41.70 (36.85–44.78) | 41.70 (37.58–45.00) | 41.60 (36.8–44.30) | 0.4974 |
| ALT (U/L) | 24.40 (15.70–41.80) | 24.40 (15.60–39.15) | 24.05 (15.78–49.48) | 0.6378 |
| AST (U/L) | 23.65 (17.90–38.03) | 23.75 (18.10–37.00) | 22.55 (17.40–43.28) | 0.9529 |
| CRE (μ mol/L) | 63.35 (52.40–74.05) | 62.60 (52.40–72.80) | 65.90 (52.15–75.13) | 0.3351 |
| LDH (U/L) | 211.25 (175.90–299.20) | 215.20 (178.10–303.90) | 200.90 (171.70–297.40) | 0.2922 |
| EBV-DNA, median, (IQR) | ||||
| EBV-DNA (copies) | 500 (0.00–4145.00) | 643 (7.00–4390.00) | 320 (0.00–3598.00) | 0.4085 |
| ICH, mean, (SD, n) | ||||
| CD56 | 0.80 (0.38, 364) | 0.81 (0.37, 254) | 0.79 (0.39, 110) | 0.7376 |
| TIA-1 | 0.97 (0.15, 340) | 0.98 (0.14, 237) | 0.97 (0.18, 103) | 0.7970 |
| CD2 | 0.90 (0.28, 227) | 0.90 (0.27, 161) | 0.89 (0.30, 66) | 0.8790 |
| CD5 | 0.21 (0.35, 331) | 0.19 (0.34, 230) | 0.25 (0.37, 101) | 0.1488 |
| CD7 | 0.68 (0.43, 207) | 0.69 (0.43, 142) | 0.66 (0.44, 65) | 0.6218 |
| CD4 | 0.26 (0.39, 231) | 0.26 (0.39, 164) | 0.26 (0.39, 67) | 0.8953 |
| CD8 | 0.29 (0.41, 228) | 0.27 (0.40, 159) | 0.33 (0.44, 69) | 0.5823 |
| CD30 | 0.36 (0.35, 229) | 0.36 (0.36, 152) | 0.35 (0.34, 77) | 0.8621 |
| CD20 | 0.03 (0.15, 334) | 0.03 (0.15, 230) | 0.03 (0.15, 104) | 0.4715 |
| GRB | 0.96 (0.18, 276) | 0.96 (0.17, 195) | 0.95 (0.20, 81) | 0.6861 |
| Ki67 | 0.70 (0.18, 364) | 0.71 (0.17, 254) | 0.67 (0.19, 110) | 0.0962 |
| Coagulation function, median, (IQR, n) | ||||
| Fbg (g/L) | 3.35 (2.61–4.38, 346) | 3.37 (2.62–4.43, 243) | 3.30 (2.52–4.30, 103) | 0.4469 |
| Cytokines, median, (IQR, n) | ||||
| IL-2 (pg/ml) | 2.50 (2.50–2.50, 231) | 2.50 (2.50–2.50, 156) | 2.50 (2.50–2.50, 75) | 0.1681 |
| IL-4 (pg/ml) | 2.50 (2.50–2.50, 231) | 2.50 (2.50–2.50, 156) | 2.50 (2.50–2.50, 75) | 0.7759 |
| TNF-α (pg/ml) | 2.50 (2.50–2.50, 232) | 2.50 (2.50–2.50, 157) | 2.50 (2.50–2.50, 75) | 0.1398 |
| IL-6 (pg/ml) | 8.24 (2.75–21.54, 232) | 9.46 (3.41–23.07, 157) | 6.07 (2.50–19.61, 75) | 0.1254 |
| IL-10 (pg/ml) | 2.50 (2.50–6.62, 231) | 2.50 (2.50–5.32, 156) | 2.50 (2.50–10.53, 75) | 0.8398 |
| IFN-γ (pg/ml) | 2.50 (2.50–8.38, 231) | 2.50 (2.50–8.29, 156) | 2.50 (2.50–11.09, 75) | 0.2798 |
CBC Complete blood count; WBC White blood count; NE Neutrophil count; LYM Lymphocyte count; Hb Hemoglobin; PLT Platelet count; CHO Cholesterol; TG triglycerides; CRP C-reaction protein; ALT Alanine amino transferase; AST Aspartate amino transferase; CRE Creatinine; LDH Lactate dehydrogenase
ICH Immunohistochemical; Fbg Fibrinogen; GRB Granzyme
Fig. 2.
Feature Selection. The feature importance plot was derived from the Random Forest (RF) model. Features are ranked by their importance scores (A); the SHAP summary plot illustrates the impact of each feature on the model’s predictions. Each dot represents a patient, with the color indicating the feature value (red for positive, blue for negative) (B); in the LASSO coefficient path, x-axis represents the log of the regularization parameter (λ), and the y-axis shows the coefficients of the features (C); the cross-validation results for the LASSO regression. The x-axis shows the log of λ, and the y-axis shows the misclassification error (D); the features selected after integrating results from RF, SHAP, and LASSO analyses (E); features are ranked by their importance scores, with the color gradient indicating the magnitude of their contribution
The filtered features were utilized to train five machine learning models across both training and validation cohorts. Within the training cohort, the RF model attained the highest AUC of 0.88, surpassing other models such as XGBoost, SVM, DT, and Logistic regression (Fig. 3a). The RF model demonstrated robust performance across sensitivity, specificity, accuracy, PPV, and NPV (Fig. 3b). It accurately classified 172 non-PD cases (68%) and 46 PD cases (18%), while producing 6 false positives (2%) and 30 false negatives (12%) (Fig. 3c). In the validation cohort, RF achieved an AUC of 0.75, followed by SVM, DT, and Logistic regression, although it was slightly lower than XGBoost (Fig. 3d). The RF model maintained a balanced performance in terms of sensitivity and specificity (Fig. 3e). It correctly classified 65 non-PD cases (59%) and 29 PD cases (26%), with 5 false positives (5%) and 11 false negatives (10%) (Fig. 3f). Collectively, the RF model exhibited superior AUC values and well-balanced performance metrics.
Fig. 3.
Machine Learning Model Performance Evaluation. The receiver operating characteristic (ROC) curves for five machine learning models in the training cohort (A) and the validation cohort (D); a bar chart comparison of key performance metrics for all models in the training cohort (B) and the validation cohort (E); the confusion matrix presents the classification results of the RF model in the training cohort (C) and the validation cohort (F)
Development of the SER nomogram model
Univariate Cox regression analysis explored the association between survival outcomes and variables like Ann Arbor stage, ECOG, LDH, and RF model (Fig. 4a). Multivariate Cox regression analysis further identified that Ann Arbor stage, ECOG, and RF model were independently associated with long-term survival (Fig. 4b). Consequently, a SER nomogram was constructed by integrating these independent prognostic factors-Ann Arbor stage, ECOG, and RF model-to predict 1-, 3-, and 5-year survival probabilities (Fig. 4c).
Fig. 4.
SER Prognostic Model Development. Univariate cox regression analysis (A), and Multivariate cox regression analysis (B) shows the hazard ratios (HR) and 95% confidence intervals (CI) for various clinical and laboratory features, PTI represents primary tumor invasion; the SER nomogram integrates independent prognostic factors Ann Arbor stage, ECOG, and RF (C). Each factor is assigned a point value based on its HR, and the total points are used to predict 1-, 3-, and 5-year survival probabilities
The survival curves consistently reveal significant differences between high-risk and low-risk groups across all cohorts, highlighting the SER model’s efficacy in survival prediction (Fig. 5a–c).
Fig. 5.
Survival analysis and prognostic model comparison. The Kaplan–Meier survival curves for the total cohort (A), the training cohort (B), and the validation cohort (C), comparing high-risk (SER high) and low-risk (SER low) groups; the concordance index (C-index) trends over time for the SER, PINK-E, and NRI models in the total cohort (D), the training cohort (E), and the validation cohort (F)
Prognosis model comparison
In our analysis, the C-index of the SER, PINK-E, and NRI models were compared over time. Results indicated that the SER model consistently achieved a higher C-index than PINK-E and NRI for predicting 3-, and 5-year survival, demonstrating superior long-term predictive accuracy although the SER model exhibited a slightly lower C-index than the PINK-E model for 1-year survival (Fig. 5d–f). Similarly, ROC analysis showed that the SER model outperformed PINK-E and NRI in terms of AUC values for 3- and 5-year survival predictions. Nevertheless, the SER model presented slightly lower AUC values than the PINK-E model for 1-year survival (Fig. 6). Overall, the SER model demonstrated extreme effectiveness in survival prediction across different time points, especially for the long-term survival prediction.
Fig. 6.
Comparison of time-dependent ROC curves and AUC values of prognostic model. The time-dependent ROC curves of the SER, PINK-E, and NRI models in the total cohort (A), the training cohort (C), and the validation cohort (E), evaluating their predictive performance for 1-, 3-, and 5-year survival; the AUC values of the SER, PINK-E, and NRI models in the total cohort (B), the training cohort (D), and the validation cohort (F)
Discussion
In the pre-asparaginase chemotherapy era for NKTCL, prognostic models primarily used the IPI [16], KPI, and nomogram, validated in patients on anthracycline-based chemotherapy [17]. However, with the shift to asparaginase-based regimens as the standard treatment, those models exhibited limitations and were successively replaced by the PINK-E [18] and NRI [19]. These newer models offer better risk-stratified prognosis differentiations and have been incorporated into the NCCN guidelines. While the efficacy of ICI in combination with asparaginase-based chemotherapy for NKTCL is still under clinical trial investigation, our study is the first to identify potential biomarkers that may enhance treatment efficacy and long-term survival prediction in this immunotherapy-chemotherapy combination setting.
As the largest cohort of ICI-treated NKTCL patients till now, our study incorporated 3 outcome features, 9 clinical features, and 34 laboratory features. Notably, although age is a significant factor in the PINK-E and NRI prognostic models, it was found to have no association with ICI outcomes or overall survival in our study (Table 3). HLH [20], a severe and poor-prognosis complication linked to NKTCL, was found to influence immunotherapy efficacy (Table 3). However, due to its low ranking in feature selection, HLH was not included in the survival analysis. As a part of NRI model, LDH did not emerge as a top feature in either machine learning or SHAP analysis though there was statistic difference between non-PD and PD patients (Table 4). Bone marrow involvement was an independent predictor of overall survival in NKTCL [21, 22] and was incorporated into the PIT prognostic model [23] for peripheral T-cell lymphoma unspecified (PTCL, NOS) [24]. While established models like PINK-E use stage IV, where bone marrow involvement is a key determinant, as a composite variable, our model’s novelty lies in the direct and explicit use of bone marrow involvement as an independent variable within a machine learning framework specifically tailored for NKTCL patients in the immunotherapy era, it may offer a different and potentially more granular perspective for predictive modeling in this specific therapeutic context.
Table 3.
Comparison of clinical characteristics among NKTCL patients with differential ICI efficacy
| Characteristic | Total (n = 364) | non-PD (n = 312) | PD (n = 52) | P value |
|---|---|---|---|---|
| No. (%) | No. (%) | No. (%) | ||
| Age, y | 0.6325 | |||
| Median (IQR) | 45 (36–57) | 45 (35–57) | 47 (37–55) | |
| Sex | 0.1444 | |||
| Male | 251 (68.96) | 220 (70.51) | 31 (59.62) | |
| Female | 113 (31.04) | 92 (29.49) | 21 (40.38) | |
| ANN ARBOR Stage | < 0.0001 | |||
| 1–2 | 247 (67.86) | 226 (72.67) | 20 (38.46) | |
| 3–4 | 117 (32.14) | 85 (27.33) | 32 (61.54) | |
| Multiple extranodal involvement | 0.0003 | |||
| Yes | 101 (27.75) | 237 (75.96) | 26 (50.00) | |
| No | 263 (72.25) | 75 (24.04) | 26 (50.00) | |
| UAT | < 0.0001 | |||
| Yes | 311 (85.44) | 277 (88.78) | 34 (65.38) | |
| No | 53 (14.56) | 35 (11.22) | 18 (34.62) | |
| Bone marrow or intracranial involvement | < 0.0001 | |||
| Bone marrow | 20 (5.49) | 10 (3.21) | 10 (19.23) | |
| Bone marrow and intracranial | 8 (2.20) | 3 (0.96) | 5 (9.62) | |
| None | 336 (92.31) | 299 (95.83) | 37 (71.15) | |
| HLH | 0.0009 | |||
| Yes | 25 (6.87) | 15 (4.81) | 10 (19.23) | |
| No | 339 (93.13) | 297 (95.19) | 42 (80.77) | |
| ECOG | 0.0007 | |||
| 0–1 | 298 (81.87) | 265 (84.94) | 33 (63.46) | |
| 2–4 | 66 (18.13) | 47 (15.06) | 19 (36.54) | |
| Chemotherapy Regimens | < 0.0001 | |||
| p-GEMOX-based regimens | 109 (29.95) | 102 (32.69) | 7 (13.46) | |
| Pegaspargase-based regimens | 114 (31.32) | 100 (32.05) | 14 (26.92) | |
| Chidamide-based regimens | 113 (31.04) | 82 (26.28) | 31 (59.62) | |
| Others | 28 (7.69) | 28 (8.97) | 0 (0.00) |
UAT Upper aerodigestive tract; HLH Hemophagocytic lymphohistiocytosis; ECOG Eastern Cooperative Oncology Group; PD Progression disease
Table 4.
Comparison of laboratory features among NKTCL patients with differential ICI efficacy
| Laboratory features | Total (n = 364) | non-PD (n = 312) | PD (n = 52) | P value |
|---|---|---|---|---|
| CBC, median, (IQR) | ||||
| WBC (10E9/L) | 5.84 (4.27–7.46) | 5.86 (4.57–7.51) | 5.65 (3.29–6.57) | 0.0489 |
| NE (10E9/L) | 3.67 (2.64–5.21) | 3.69 (2.67–5.21) | 3.55 (2.36–5.17) | 0.3494 |
| LYM (10E9/L) | 1.30 (0.89–1.76) | 1.35 (0.97–1.35) | 0.85 (0.50–1.57) | < 0.0001 |
| Hb (g/L) | 127.00 (112.00–141.00) | 128.00 (114.00–143.00) | 113.00 (89.25–131.50) | < 0.0001 |
| PLT (10E9/L) | 243.50 (186.00–303.80) | 247.00 (193.30–304.00) | 216.00 (111.30–296.30) | 0.0106 |
| Biochemical test, median, (IQR) | ||||
| CHO (mmol/L) | 4.52 (3.93–5.38) | 4.58 (4.01–5.40) | 4.11 (3.58–5.22) | 0.0189 |
| TG (mmol/L) | 1.46 (1.08–2.12) | 1.45 (1.06–2.15) | 1.50 (1.12–2.04) | 0.7931 |
| CRP (mgl/L) | 6.17 (2.14–21.82) | 5.69 (1.90–20.44) | 13.60 (3.60–41.65) | 0.0070 |
| Albumin (g/L) | 41.70 (36.85–44.78) | 42.20 (37.53–45.10) | 39.30 (34.30–41.85) | 0.0004 |
| ALT (U/L) | 24.40 (15.70–41.80) | 23.30 (15.60–39.00) | 29.65 (18.85–71.25) | 0.0133 |
| AST (U/L) | 23.65 (17.90–38.03) | 22.65 (17.73–36.18) | 30.20 (19.78–49.25) | 0.0093 |
| CRE (μ mol/L) | 63.35 (52.40–74.05) | 63.75 (53.55–75.20) | 58.10 (46.85–69.63) | 0.0263 |
| LDH (U/L) | 211.25 (175.90–299.20) | 207.70 (173.20–281.80) | 283.40 (194.60–424.50) | 0.0002 |
| EBV-DNA, median, (IQR) | ||||
| EBV-DNA (copies) | 500 (0.00–4145.00) | 313 (0.00–2560.00) | 9850 (645.00–97500.00) | < 0.0001 |
| ICH, mean, (SD, n) | ||||
| CD56 | 0.80 (0.38, 364) | 0.82 (0.36, 312) | 0.70 (0.45, 52) | 0.0411 |
| TIA-1 | 0.97 (0.15, 340) | 0.98 (0.14, 292) | 0.96 (0.20, 48) | 0.7516 |
| CD2 | 0.90 (0.28, 227) | 0.90 (0.28, 198) | 0.90 (0.29, 29) | 0.7599 |
| CD5 | 0.21 (0.35, 331) | 0.21 (0.35, 282) | 0.22 (0.36, 49) | 0.7247 |
| CD7 | 0.68 (0.43, 207) | 0.69 (0.43, 177) | 0.64 (0.44, 30) | 0.6233 |
| CD4 | 0.26 (0.39, 231) | 0.26 (0.38, 198) | 0.28 (0.43, 33) | 0.9417 |
| CD8 | 0.29 (0.41, 228) | 0.31 (0.42, 196) | 0.20 (0.37, 32) | 0.1591 |
| CD30 | 0.36 (0.35, 229) | 0.35 (0.34, 196) | 0.40 (0.41, 33) | 0.9733 |
| CD20 | 0.03 (0.15, 334) | 0.03 (0.16, 285) | 0.01 (0.05, 49) | 0.4262 |
| GRB | 0.96 (0.18, 276) | 0.96 (0.16, 235) | 0.90 (0.26, 41) | 0.0244 |
| Ki67 | 0.70 (0.18, 364) | 0.69 (0.18, 312) | 0.71 (0.18, 52) | 0.3923 |
| Coagulation function, median, (IQR, n) | ||||
| Fbg (g/L) | 3.35 (2.61–4.38, 346) | 3.39 (2.64–4.41, 297) | 3.11 (2.24–4.30, 49) | 0.0673 |
| Cytokines, median, (IQR, n) | ||||
| IL-2 (pg/ml) | 2.50 (2.50–2.50, 231) | 2.50 (2.50–2.50, 192) | 2.50 (2.50–2.50, 39) | 0.5234 |
| IL-4 (pg/ml) | 2.50 (2.50–2.50, 231) | 2.50 (2.50–2.50, 192) | 2.50 (2.50–2.50, 39) | 0.3642 |
| TNF-α (pg/ml) | 2.50 (2.50–2.50, 232) | 2.50 (2.50–2.50, 192) | 2.50 (2.50–2.50, 40) | 0.0980 |
| IL-6 (pg/ml) | 8.24 (2.75–21.54, 232) | 7.28 (2.60–17.90, 192) | 15.87 (5.32–58.54, 40) | 0.0041 |
| IL-10 (pg/ml) | 2.50 (2.50–6.62, 231) | 2.50 (2.50–3.88, 192) | 7.85 (2.50–56.43, 39) | < 0.0001 |
| IFN-γ (pg/ml) | 2.50 (2.50–8.38, 231) | 2.50 (2.50–6.93, 192) | 6.11 (2.5–23.15, 39) | 0.0011 |
CBC Complete blood count; WBC White blood count; NE Neutrophil count; LYM Lymphocyte count; Hb Hemoglobin; PLT Platelet count; CHO Cholesterol; TG triglycerides; CRP C-reaction protein; ALT Alanine amino transferase; AST Aspartate amino transferase; CRE Creatinine; LDH Lactate dehydrogenase
ICH Immunohistochemical; Fbg Fibrinogen; GRB Granzyme
Though prior studies have indicated that CD7 [25], CD30 [26] and Ki67 [27] expression may carry prognostic significance in NKTCL, our analysis did not uncover any meaningful prognostic biomarkers in this regard (Table 4). While CD56 and GRB expression levels did differ statistically between non-PD and PD patients, the overall positive rates were too close to signify a clear difference. Additionally, existing research has shown that there is no clinical or prognostic significance in differentiating the two subtypes of CD56 + and CD56-NKTCL based on their immunophenotypic profiles [28]. While acknowledging the limitation of missing cytokine data, our analysis highlighted potential prognostic significance of inflammatory cytokines such as IL-6, IL-10, and IFN-γ (Table 4), aligning with prior studies [29–31], nevertheless, further validation with larger cohorts is needed.
The final SER model identified Ann Arbor stage, ECOG, lymphocyte count, platelet count, bone marrow involvement, cholesterol (CHO), and EBV DNA copy number as the key features. The Ann Arbor stage is part of both the PINK and NRI models. EBV DNA is included in the PINK-E model, just like ECOG is part of the NRI model. But adding lymphocyte count, platelet count, CHO, and bone marrow involvement as prognostic biomarkers is a new contribution for NKTCL patients treated with ICIs.
Lymphopenia has been established as a prognostic indicator in various cancers, including advanced-stage carcinoma, sarcoma [32], Hodgkin lymphoma [33], diffuse large B-cell lymphoma [34], and adult T-cell leukemia/lymphoma [35]. Additionally, research has shown that low CD4 + T cell count is an independent predictor of worse survival in NKTCL [36]. Our study further confirms that the long-term prognostic value of lymphocyte counts in NKTCL patients with ICI treatment.
A low platelet count has been linked to poor outcomes in NKTCL [21], whereas a meta-analysis indicates that a high platelet count is a negative predictor for lung cancer survival [37]. Even in the recent SCORPIO model involving 9745 ICI-treated patients across 21 cancer types, the role of platelet count remains inconclusive [15]. Given the paucity of existing literature, the prognostic role of platelet count in NKTCL warrants further validation.
Another previous meta-analysis of 85,173 cancer patients showed that higher lipid metabolic biomarkers are significantly linked to better overall and disease-free survival in cancer patients [38]. Our study first highlights the prognostic importance of cholesterol in ICI-treated NKTCL patients.
We categorized chemotherapy regimens into p-GEMOX-based, pegaspargase-based, chidamide-based, and others. In the p-GEMOX group, 7 of 109 patients (6.42%) showed disease progression. In the pegaspargase group, 14 of 114 patients (12.28%) had PD. In the chidamide group, 31 of 113 patients (27.43%) experienced PD. Our findings suggest that the ICI-P-GEMOX combination is well-tolerated and effective for NKTCL treatment, showing promise for broader clinical use. However, due to the lack of rigorous prospective clinical trial data, we are unable to draw definitive conclusions regarding the superiority or inferiority of specific concomitant chemotherapy regimens. Although we did not include concomitant chemotherapy as a variable in the main predictive model, further ongoing clinical trials are actively addressing these questions. Further validation in larger-scale, prospective, multicenter cohorts with strictly controlled concomitant chemotherapy drugs is essential to confirm the generalizability and stability of our SER model.
In conclusion, our study developed an optimized SER model for ICI therapy in NKTCL, incorporating the RF algorithm along with Ann Arbor stage and ECOG. This model demonstrated superior long-term predictive performance compared to PINK-E and NRI, providing a valuable tool for early prediction of ICI therapy failure and long-term survival. Our study pioneers in proposing the first prognostic model for ICI immunotherapy in NKTCL and encompasses the largest patient cohort until now. However, its single-center design and absence of external validation cohorts might restrict the generalizability of the findings.
Our study has several limitations that must be acknowledged. First, the retrospective nature of this analysis, while enabling the assembly of a substantial cohort, inherently introduces risks of selection bias and unmeasured confounding. The patient population was drawn from a single tertiary center, which may attract more complex or severe cases, potentially limiting the generalizability of our findings to broader community settings. Furthermore, despite our efforts to include relevant clinical variables, there may be unrecorded or unknown confounding factors—such as specific genetic markers, detailed treatment compliance, or nuances in supportive care—that could influence both treatment response and prognosis, but which we were unable to account for in our model.
Second, the impact of missing data on our cohort selection is a significant consideration. From 459 ICIs-treated NKTCL patients, a substantial number were excluded due to incomplete clinical records. This process, while necessary to ensure data quality for model development, may have introduced a form of selection bias. It is plausible that patients with complete follow-up and well-documented records represent a subgroup with distinct characteristics, which could affect the external validity of the SER model. Although we employed statistical techniques to handle missing values where appropriate, the potential for bias in the final selected cohort remains a constraint.
Third, we must address the potential for model overfitting, as suggested by the observed performance drop between the training and internal validation cohorts. Although we implemented robust measures to mitigate this risk, including tenfold cross-validation and regularization within the machine learning algorithm, the finite size of our dataset, particularly the validation set, remains a constraint. The model’s performance metrics indicate a degree of optimism when applied to the training data. Therefore, the current version of the SER model should be viewed as a promising but preliminary predictive tool. Its ultimate clinical utility and generalizability are contingent upon future validation in larger, prospective, and multi-center cohorts. Such external validation is an essential next step to confirm the model’s stability, calibrate its predictive parameters, and establish its role in routine clinical practice for managing NKTCL in the immunotherapy era.
Despite these limitations, our study provides a foundational framework for prognostication in this challenging disease context and underscores the potential of integrating machine learning with clinical data to personalize treatment strategies.
Acknowledgements
We thank all the participants.
Author contributions
Ao Zhang wrote the manuscript, Runkun Han, Denghan Zhang, Bushu Xu organized the clinic data, Shenrui Bai provided clinical guidance, Hao Chen performed data analysis, Yifei Ma revised the manscript.
Funding
This research was supported by the National Natural Science Foundation of China youth science fund projects (82201922 to Ao Zhang and 82102687 to Yifei Ma).
Data availability
The primary datasets analyzed during this study have been deposited in the RDD system (Research Data Deposit) at http://www.researchdata.org.cn and have been assigned the approval RDD number of RDDA-2025226964.
Declarations
Ethics approval and consent to participate
The study was conducted according to the principles of the Declaration of Helsinki. The study protocol was approved by the Ethics Committee of Sun Yat-sen University Cancer Center.
Consent for publication
All patients provided written informed consent for the collection and publication of medical information related to their cases.
Competing interests
The authors declare no competing interests.
Footnotes
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Runkun Han, Denghan Zhang and Shenrui Bai have contributed equally to this work.
Contributor Information
Bushu Xu, Email: xubsh@sysucc.org.cn.
Hao Chen, Email: chenhao@sysucc.org.cn.
Ao Zhang, Email: zhangao@sysucc.org.cn.
References
- 1.Anna BA-N, Thankamma A, Sarah B, Daniel James H. Non-Hodgkin lymphoma. BMJ. 2018;362(0):1. [Google Scholar]
- 2.Miranda RN, Amador C, Chan JK, Guitart J, Rech KL, Medeiros LJ, et al. of the World Health Organization Classification of Tumors of the Hematopoietic and Lymphoid Tissues: mature T-Cell, NK-Cell, and Stroma-Derived Neoplasms of Lymphoid Tissues. Modern Pathol. 2024;37(8):100512. [DOI] [PubMed] [Google Scholar]
- 3.Malpica L, Idrobo H, Pavlovsky A, et al. Epidemiology, clinical features, and outcomes of peripheral T-cell lymphoma in Latin America: an international, retrospective, cohort study. Lancet Haematol. 2025;12(4):e258–68. [DOI] [PubMed] [Google Scholar]
- 4.Liu X, Cao D, Liu H, Ke X, Liu X, Xu X. Comparative analysis of upper aerodigestive tract and non-upper aerodigestive tract in NK/T-cell lymphoma. Clin Transl Oncol. 2023;26(1):214–24. [DOI] [PubMed] [Google Scholar]
- 5.Chan JK, Tsang WY, Ng CS. Clarification of CD3 immunoreactivity in nasal T/natural killer cell lymphomas: the neoplastic cells are often CD3 epsilon+. Blood. 1996;87(2):839–84. [PubMed] [Google Scholar]
- 6.Hue S, Oon M, Wang S, Tan S, Ng S. Epstein-Barr virus-associated T- and NK-cell lymphoproliferative diseases: an update and diagnostic approach. Pathology. 2020;52(1):111–27. [DOI] [PubMed] [Google Scholar]
- 7.Wang H, Fu BB, Gale RP, Liang Y. NK-/T-cell lymphomas. Leukemia. 2021;35(9):2460–8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Kim SJ, Yoon SE, Kim WS. Treatment of localized extranodal NK/T cell lymphoma, nasal type: a systematic review. J Hematol Oncol. 2018;11(1):140. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Lim SH, Hong JY, Lim ST, et al. Beyond first-line non-anthracycline-based chemotherapy for extranodal NK/T-cell lymphoma: clinical outcome and current perspectives on salvage therapy for patients after first relapse and progression of disease. Ann Oncol. 2017;28(9):2199–205. [DOI] [PubMed] [Google Scholar]
- 10.Tao R, Fan L, Song Y, et al. Sintilimab for relapsed/refractory extranodal NK/T cell lymphoma: a multicenter, single-arm, phase 2 trial (ORIENT-4). Signal Transduct Target Ther. 2021;6(1):365. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Huang H, Tao R, Hao S, et al. Sugemalimab monotherapy for patients with relapsed or refractory extranodal natural killer/T-cell lymphoma (GEMSTONE-201): results from a single-arm, multicenter, phase II study. J Clin Oncol. 2023;41(16):3032–41. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Kim S, Lim J, Laurensia Y, et al. Avelumab for the treatment of relapsed or refractory extranodal NK/T-cell lymphoma: an open-label phase 2 study. Blood. 2020;136(24):2754–63. [DOI] [PubMed] [Google Scholar]
- 13.Sun P, Li Y, Li C, et al. A phase II study of sintilimab, anlotinib, and pegaspargase sandwiched with radiotherapy as first-line therapy in patients with newly diagnosed, stage I-II extranodal natural-killer/T-cell lymphoma. Am J Hematol. 2023;98(7):1043–51. [DOI] [PubMed] [Google Scholar]
- 14.Tian X, Cai J, Xia Y, et al. First-line sintilimab with pegaspargase, gemcitabine, and oxaliplatin in advanced extranodal natural killer/T cell lymphoma (SPIRIT): a multicentre, single-arm, phase 2 trial. Lancet Haematol. 2024;11(5):e336–44. [DOI] [PubMed] [Google Scholar]
- 15.Yoo S, Fitzgerald C, Cho B, et al. Prediction of checkpoint inhibitor immunotherapy efficacy for cancer using routine blood tests and clinical data. Nat Med. 2025;31(3):869–80. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.LaCasce A. The international prognostic index: still relevant 30 years later. Haematologica. 2023;108(6):1453–4. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Au W, Weisenburger D, Intragumtornchai T, et al. Clinical differences between nasal and extranasal natural killer/T-cell lymphoma: a study of 136 cases from the International Peripheral T-Cell Lymphoma Project. Blood. 2009;113(17):3931–7. [DOI] [PubMed] [Google Scholar]
- 18.Kim S, Yoon D, Jaccard A, et al. A prognostic index for natural killer cell lymphoma after non-anthracycline-based treatment: a multicentre, retrospective analysis. Lancet Oncol. 2016;17(3):389–400. [DOI] [PubMed] [Google Scholar]
- 19.Chen S, Yang Y, Qi S, et al. Validation of nomogram-revised risk index and comparison with other models for extranodal nasal-type NK/T-cell lymphoma in the modern chemotherapy era: indication for prognostication and clinical decision-making. Leukemia. 2021;35(1):130–42. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Wen H, Ma H, Cai Q, et al. Recurrent ECSIT mutation encoding V140A triggers hyperinflammation and promotes hemophagocytic syndrome in extranodal NK/T cell lymphoma. Nat Med. 2018;24(2):154–64. [DOI] [PubMed] [Google Scholar]
- 21.Xu P, Wang Y, Shen Y, Wang L, Shen Z, Zhao W. Prognostic factors of Chinese patients with T/NK-cell lymphoma: a single institution study of 170 patients. Med Oncol. 2012;29(3):2176–82. [DOI] [PubMed] [Google Scholar]
- 22.Liu X, Cao D, Liu H, Ke X, Liu X, Xu X. Comparative analysis of upper aerodigestive tract and non-upper aerodigestive tract in NK/T-cell lymphoma. Clin Transl Oncol Off Publ Federat Span Oncol Soc Natl Cancer Inst Mexico. 2024;26(1):214–24. [DOI] [PubMed] [Google Scholar]
- 23.Lin H, Liu C, Pai J, et al. How to predict the outcome in mature T and NK cell lymphoma by currently used prognostic models? Blood Cancer J. 2012;2(10):e93. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Gallamini A, Stelitano C, Calvi R, et al. Peripheral T-cell lymphoma unspecified (PTCL-U): a new prognostic model from a retrospective multicentric clinical study. Blood. 2004;103(7):2474–9. [DOI] [PubMed] [Google Scholar]
- 25.Fu X, Wan W, Sun Z, et al. [Expression of CD7 and its correlation with prognosis in patients with NK/T-cell lymphoma]. Zhonghua xue ye xue za zhi = Zhonghua xueyexue zazhi. 2020;41(11):921–6. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Kim W, Nam S, Kim S, et al. Prognostic implications of CD30 expression in extranodal natural killer/T-cell lymphoma according to treatment modalities. Leuk Lymphoma. 2015;56(6):1778–86. [DOI] [PubMed] [Google Scholar]
- 27.He X, Chen Z, Fu T, et al. Ki-67 is a valuable prognostic predictor of lymphoma but its utility varies in lymphoma subtypes: evidence from a systematic meta-analysis. BMC Cancer. 2014;14:153. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Li Y, Wang H, Feng X, et al. Immunophenotypic characteristics and clinical relevance of CD56+ and CD56- extranodal nasal-type natural killer/T-cell lymphoma. Leuk Lymphoma. 2011;52(3):417–24. [DOI] [PubMed] [Google Scholar]
- 29.Hui Y, Gao Y, Li J, et al. Elevated serum IL-6 and total IgEAb are associated with poor survival in natural killer/T-cell lymphoma. Ann Hematol. 2024;103(4):1285–92. [DOI] [PubMed] [Google Scholar]
- 30.Wang H, Wang L, Wuxiao Z, et al. Increased serum levels of interleukin-10 predict poor prognosis in extranodal natural killer/T-cell lymphoma patients receiving asparaginase-based chemotherapy. OncoTargets Ther. 2015;8:2589–99. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31.Chan J, Lim J, Ong C. Towards next generation biomarkers in natural killer/T-cell lymphoma. Life (Basel, Switzerland). 2021. 10.3390/life11080838. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32.Ray-Coquard I, Cropet C, Van Glabbeke M, et al. Lymphopenia as a prognostic factor for overall survival in advanced carcinomas, sarcomas, and lymphomas. Cancer Res. 2009;69(13):5383–91. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33.Hasenclever D, Diehl V. A prognostic score for advanced Hodgkin’s disease. International Prognostic Factors Project on Advanced Hodgkin’s Disease. N Engl J Med. 1998;339(21):1506–14. [DOI] [PubMed] [Google Scholar]
- 34.Kim DH, Baek JH, Chae YS, et al. Absolute lymphocyte counts predicts response to chemotherapy and survival in diffuse large B-cell lymphoma. Leukemia. 2007;21(10):2227–30. [DOI] [PubMed] [Google Scholar]
- 35.Mitrovic Z, Perry A, Suzumiya J, et al. The prognostic significance of lymphopenia in peripheral T-cell and natural killer/T-cell lymphomas: a study of 826 cases from the International Peripheral T-cell Lymphoma Project. Am J Hematol. 2012;87(8):790–4. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 36.Zeng R, Zhu W, Jiang L, et al. Low absolute CD4(+) T cell counts in peripheral blood: an independent predictor of inferior survival in natural killer/T-cell lymphoma-a retrospective cohort study. Ann Transl Med. 2022;10(20):1119. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 37.Zhang X, Ran Y. Prognostic role of elevated platelet count in patients with lung cancer: a systematic review and meta-analysis. Int J Clin Exp Med. 2015;8(4):5379–87. [PMC free article] [PubMed] [Google Scholar]
- 38.Peng Q, Zhan C, Shen Y, et al. Blood lipid metabolic biomarkers are emerging as significant prognostic indicators for survival in cancer patients. BMC Cancer. 2024;24(1):1549. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Data Availability Statement
The primary datasets analyzed during this study have been deposited in the RDD system (Research Data Deposit) at http://www.researchdata.org.cn and have been assigned the approval RDD number of RDDA-2025226964.