A clinical-radiomic model based on best subset regression for prognostic prediction in natural killer/T-cell lymphoma Authors: Fangfang Tian, Longlan Chen, Xiaoliang Chen, Hua Pang Published: 2026 ABSTRACT AbstractBackgroundThe outcomes of patients with natural killer/T-cell lymphoma (NKTCL) are heterogeneous, thus, this study aimed to develop and validate prognostic models for overall survival (OS) and progression-free survival (PFS) in NKTCL by integrating clinicopathological and pre-treatment [18F]-fluorodeoxyglucose positron emission tomography/computed tomography ([18F]-FDG PET/CT)-derived variables, comparing the performance of univariate COX, least absolute shrinkage and selection operator (LASSO), and best subset regression (BSR) methods.MethodsThis retrospective study analyzed a cohort of 114 pre-treatment patients with NKTCL who have performed [18F]-FDG PET/CT scan. Predictors included clinical parameters (e.g., age, albumin (Alb)), pathological indices (e.g., Ki67), Epstein-Barr virus (EBV), and imaging metrics (e.g., radiomics score (Radscore), maximum standardized uptake value (SUVmax), lumbar-to-liver SUVmean ratio (LLR), and spleen-to-liver SUVmean ratio (SLR)). Variable selection was performed using univariate COX (P < 0.1), LASSO, and BSR techniques. The resulting models were compared using the area under the curve (AUC) and concordance index (C-index), with internal validation via bootstrapping.ResultsThe BSR-derived model demonstrated superior performance for predicting OS (3-year AUC: 0.883; C-index: 0.807). For OS, the optimal multivariate model identified Alb (hazard ratios (HR): 0.226, 95% confidence interval (CI): 0.073–0.697, P = 0.00961), EBV (HR: 3.826, 95% CI: 1.259–11.630, P = 0.017986), SUVmax (HR: 2.466, 95% CI: 1.123–5.416, P = 0.024517), LLR (HR: 0.195, 95% CI: 0.077–0.493, P = 0.000549), and Radscore (HR: 15.463, 95% CI: 5.313–45.001, P = 5.05e-07) as independent prognostic factors. For PFS, the optimal model comprised age (HR: 2.292, 95% CI: 1.076–4.884, P = 0.03166), EBV (HR: 4.168, 95% CI: 1.588–10.936, P = 0.00373), and Ki67 (HR: 2.608, 95% CI: 1.221–5.572, P = 0.01334).ConclusionComparative analysis revealed that BSR outperformed both univariate COX regression and LASSO in selecting variables for NKTCL prognosis modeling. Using BSR, we developed a robust model integrating essential clinical factors and PET/CT-derived radiomic features (e.g., Alb, EBV, SUVmax, LLR and Radscore for OS and age, EBV and Ki67 for PFS), offering a valuable tool for risk stratification and guiding individualized treatment strategies.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s40644-026-01074-0.Keywords: Natural killer/T-cell lymphoma, Overall survival, Progression-free survival, Model prediction Introduction Natural killer/T-cell lymphoma (NKTCL), particularly its extranodal form, is an aggressive malignancy with a heterogeneous clinical outcome. While some patients achieve long-term remission with current protocols, such as asparaginase-based regimens, a significant proportion experience refractory disease or early relapse, leading to poor survival [1]. This unpredictability in patient outcomes poses a major challenge for clinicians in selecting appropriate therapy intensity and counseling patients accurately. Current prognostic frameworks, including the Ann Arbor staging system and the Prognostic Index for NK/T-cell Lymphoma (Pinke), provide a foundational risk categorization but lack the granularity needed to guide individualized treatment decisions. Furthermore, although biomarkers like Epstein-Barr virus (EBV) infection are invaluable, they often focus on a single aspect of disease biology—such as tumor burden—and may not fully capture the complexity of both short-term and long-term outcomes. In recent years, several prognostic models for predicting progression-free survival (PFS) and overall survival (OS) have been developed using COX regression. Factors such as patient age, tumor node metastasis (TNM) staging, the Chinese Southwest Oncology Group and Asia Lymphoma Study Group ENKTL (CA) staging, Eastern Cooperative Oncology Group (ECOG) performance status, presence of B symptoms, and lactate dehydrogenase levels have been identified as independent risk factors in NKTCL. Additionally, [18F]-fluorodeoxyglucose positron emission tomography/computed tomography ([18F]-FDG PET/CT) has been recommended for prognostic stratification in patients with extranodal NKTCL [2, 3], and the radiomic features derived from these imaging studies have also been validated [4]. However, the methods used to develop prognostic models significantly affect their quality, and it remains uncertain whether conventional COX regression is the optimal approach. To address this gap, we compared multiple modeling strategies to identify the optimal approach for predicting OS and PFS in NKTCL patients. For variable selection, we employed three distinct methods: univariate COX regression [5, 6], the least absolute shrinkage and selection operator (LASSO) [7], and best-subset regression (BSR) [8]. Given that univariate COX remains the most widely used method in oncology due to its simplicity, we applied it as a benchmark. For comparison, we also utilized LASSO to handle multicollinearity and BSR to exhaustively evaluate all predictor combinations. This multi-method approach enabled a robust comparative assessment and a comprehensive evaluation of potential predictors. By systematically comparing these three modeling techniques—univariate COX, LASSO, and BSR—we aimed to identify the most parsimonious and powerful combination of pretreatment clinical and biological variables for predicting OS and PFS in NKTCL. We anticipate that the resulting models will enable more personalized risk assessment, assist clinicians in tailoring first-line therapy intensity based on PFS risk, and facilitate preemptive supportive care for patients at high risk of poor OS—ultimately aiming to improve both survival and quality of life for patients with this challenging disease. Methods Patients From February 2014 to June 2023, 114 consecutive patients with newly diagnosed NKTCL were enrolled in this retrospective study. The inclusion criteria were as follows: (a) Pathological confirmation of extranodal NKTCL, (b) completion of a pre-treatment ([18F]-FDG PET/CT scan, and (c) availability of follow-up data. Patients were excluded if they had (a) a fasting blood glucose level > 200 mg/dL before [18F]-FDG PET/CT scanning or (b) concurrent infections, multiple primary cancers, or a history of prior cancer (Fig. 1). Flowchart of patient selection The primary endpoints of our study were OS and PFS. PFS was calculated from the imaging date to the first occurrence of disease progression, including locoregional recurrence or metastasis, or to the date of death from any cause or the date of the last follow-up. OS was measured from the date of imaging to the date of death from any cause or the date of the last follow-up. The study protocol was approved by the local Ethics Committee of the First Affiliated Hospital of Chongqing Medical University (ZZ2025-230-01) and the Ethics Committee of Chongqing University Cancer Hospital (CZLS2025159-A). [18F]-FDG PET/CT scan All patients underwent whole-body [18F]-FDG PET/CT imaging, covering the area from the skull base to the upper third of the thigh, using a dedicated PET/CT system (Philips Gemini TF 64). Before imaging, patients adhered to a fasting period of at least 6 h to maintain blood glucose levels below 200 mg/dL. An intravenous dose of [18F]-FDG (3.7–4.4 MBq/kg) was then administered. Whole-body PET/CT scans were performed approximately 60 min after the injection. The images were subjected to attenuation correction using low-dose CT data, and the corrected PET images were reconstructed using an ordered-subset expectation-maximization iterative reconstruction algorithm. PET/CT image analysis The PET/CT images were independently evaluated by two experienced nuclear medicine physicians, who were provided with the clinical data but remained blinded to each other’s outcomes. Specifically, the pre-treatment PET/CT images were segmented, and features were extracted using LIFEx 7.2 (http://www.lifexsoft.org). The two nuclear medicine specialists (TFF and CLL), delineated the primary tumor, excluding lymph nodes, and established a computer-generated volume of interest, including voxels with intensities ≥ 40% of the maximum standardized uptake value (SUVmax). The images were resampled to achieve a 2 × 2 × 2 mm³ voxel size using three-dimensional Lagrangian polynomial interpolation. The Hounsfield units in the CT images were resampled into 400 discrete values, termed bins, through absolute discretization spanning − 1000 to 1000 Hounsfield units, yielding a fixed bin size of 5.0 Hounsfield units. We exclusively extracted first-order intensity and shape features, comprising 26 features. Absolute discretization was used to resample SUVmax units into 64 discrete levels in the PET imaging analysis, ranging from 0 to 30 SUV units, resulting in a fixed bin size of 0.476 SUV units, which aligned with typical SUV observed in NKTCL. Subsequently, data were extracted from 62 first-order intensity features, shape features, and texture indices. Texture features were examined using gray-level co-occurrence matrices, gray-level run-length matrices, neighborhood gray-tone difference matrices, and gray-level size zone matrices. The extracted radiomic feature values were normalized using the Z-transform method. The ComBat algorithm was then used to correct variations in the scanning data across different devices. In addition to radiomics, we also measured the spleen-to-liver standardized uptake value mean (SUVmean) ratio (SLR) and the lumbar-to-liver SUVmean ratio (LLR). SUVmean values for the spleen and liver were derived from a 5 cm diameter circular region of interest, avoiding large blood vessels, bile ducts, or calcified areas during delineation. For the lumbar spine, the SUVmean was measured by selecting the axial cross-section of the L2-L4 vertebral bodies, while excluding intervertebral discs and areas with bone destruction. Feature selection and radiomics signature construction The LASSO method was used to identify the most informative features for predictive analyses. We improved the model stability by determining the smallest penalty parameter λ using cross-validation. Subsequently, a linear combination of the selected features was used to calculate the radiomics score (Radscore) for each patient, with each feature weighted by its corresponding coefficient. Development and evaluation of individualized prediction models In this study, we conducted a comprehensive screening of variables to predict the prognosis of patients with NKTCL by employing univariate COX, LASSO, and BSR methodologies. The initial model incorporated the following indicators: age, Pinke, Ki67, treatment method, albumin (Alb) cancer site, tumor volume, Ann Arbor staging, CA staging, EBV, Radscore, LLR, SLR, and SUVmax. The EBV status was determined by Epstein-Barr virus-encoded RNA (EBER) in situ hybridization on tumor tissue sections. Cases with positive EBER staining were classified as EBV-positive, whereas those with negative EBER staining were classified as EBV-negative. Variables with P < 0.1 were initially filtered out using univariate COX analysis. Subsequently, we used BSR to identify the univariate combination that yielded the highest adjusted R-squared value. The univariate combination that yielded the λ value corresponding to the smallest mean square error (MSE) was identified using LASSO and cross-validation. Finally, the variables selected by these three methods were integrated into a multivariate COX regression model. We employed the bootstrap method to perform internal validation, thereby reinforcing the model’s robustness. The final model was determined using the stepwise backward regression method based on the minimum Akaike information criterion (AIC). The area under the curve (AUC) or Harrell’s concordance index (C-index) was computed to quantify the discrimination performance of the three models. The “bootstrapt” method was applied with 500 iterations. Clinical use of a predictive scoring model Nomograms were employed to visualize the outcomes of the model. Each variable is assigned a points score on the top scale; the total points from all variables are summed and projected to the bottom scales to obtain the predicted probability of 1-, 2-, and 3-year survival. Moreover, to define a clinically relevant dichotomy, the optimal cut-off value for the continuous risk score was determined using the maximally selected rank statistics (via the ‘X-tile’ software). Patients with risk scores greater than the cut-off were assigned to the high-risk group, and the remaining patients to the low-risk group. Finally, a calibration curve was utilized to further evaluate the model’s performance. Additionally, decision curve analysis (DCA) was employed to evaluate and compare the clinical utility of the prognostic models derived from different variable selection methods (Univariate COX regression, LASSO regression, and BSR). The analysis was performed for both OS and PFS. The net benefit of each model was calculated across a range of clinically plausible threshold probabilities, reflecting the varying levels of risk at which clinicians might intensify monitoring or treatment for NKTCL patients. The model with the highest net benefit across the relevant threshold range was considered the most clinically useful for decision support. Statistical analysis Continuous variables were transformed into categorical variables using the X-tile software to determine the optimal cut-off values. Univariate COX, LASSO, and BSR analyses were performed using R software (version 3.6.1), available at http://www.R-project.org, with results presented as hazard ratios (HRs) and 95% confidence intervals (CIs) to test whether they were predictors of OS and PFS. After identifying the optimal model, we evaluated its prognostic performance. The pre-existing clinical endpoints of OS and PFS were analyzed. Survival curves were generated using the Kaplan–Meier method, and the statistical significance of differences between the high- and low-risk groups was assessed using the log-rank test. All survival analyses were performed using the survival package in R. A P-value threshold of < 0.05 was used to determine statistical significance. Results Patient characteristics This study enrolled 114 patients who fulfilled the inclusion criteria (Fig. 1). The median age at diagnosis was 48.67 ± 14.48 years. The gender distribution indicated a male-to-female ratio of 1.71:1. A comprehensive overview of the clinical characteristics of these patients is presented in Table 1. [18F]-FDG PET/CT scans were performed for all patients. Following the scans, the treatment regimens varied: 39 patients received chemotherapy as the sole treatment, 54 received a combination of chemotherapy and radiotherapy, 5 were treated with radiotherapy alone, 5 received targeted therapy, and 11 did not underwent anti-caner treatment regimens. Consequently, patients were categorized into three groups based on their treatment regimen: chemotherapy, radiochemotherapy, and other therapeutic approaches. Table 1Bacis characteriscs of inlucded patientsVariablesPatients (n = 114)Ratio Sex Male7263.16%Female4236.84% Age (years) < 582723.68%≥ 588776.32% Tumor volume (cm 3 ) < 13.88271.93%≥ 13.83228.07% Cancer site Nasal cavity6557.02%Nasopharynx4942.98% EBER positivity Yes7263.16%No4236.84% Pinke ≤ 18171.05%> 13328.95% Ki67 ≤ 0.54640.35%> 0.56859.65% Ann-Arbor staging Stage I-II10188.60%Stage III-IV1311.40% CA staging Stage I-II6355.26%Stage III-IV5144.74% Lymph nodes affected Yes4539.47%No6960.53% Alb (g/L) ≤ 417364.04%> 414135.96% Radscore ≤-0.72824.56%>-0.78675.44% Treatment Radiochemotherapy5447.37%Chemotherapy3934.21%Radiotherapy54.39%Target54.39%No anti-tumor treatment119.65%CA staging: Chinese Southwest Oncology Group and Asia Lymphoma Study Group ENKTL staging, Alb: Albumin, EBV: Epstein-Barr virus, Radscore: Radiomics score Bacis characteriscs of inlucded patients CA staging: Chinese Southwest Oncology Group and Asia Lymphoma Study Group ENKTL staging, Alb: Albumin, EBV: Epstein-Barr virus, Radscore: Radiomics score Patient follow-up The median follow-up period was 19.35 months (3.1–77.1 months). Disease progression was observed in 32 of the 114 patients, and mortality was recorded in 27. PFS rates at 1 and 3 years were 56% and 11%, respectively. Correspondingly, OS rates for the same time intervals were 83% and 28%, respectively. We used three distinct models to predict OS and PFS. Development of individualized prediction models For the OS analysis, variables selected through univariate COX regression models, based on a criterion of P < 0.1, included age, Alb, Pinke, EBV, treatment method, Radscore, and SUVmax (Table 2). The BSR identified the following variables: age, Alb, SUVmax, EBV, Ann Arbor staging, LLR, and Radscore. In contrast, the LASSO regression selected age, Pinke, EBV, treatment method, Radscore, and SUVmax. Subsequently, we performed survival analysis using multivariate regression on these three sets of variables. The final variables, determined by the minimum AIC value, were identified as age, Alb, EBV, and Radscore (univariate COX/LASSO, AIC = 196.0) and Alb, SUVmax, LLR, EBV, and Radscore (BSR, AIC = 186.9) (Table 3). Table 2Univariate analysis of factors affecting OS and PFSOSPFS P HR (95%CI) P HR (95%CI)Sex0.6731.18 (0.55–2.55)0.6861.16 (0.57–2.35)Age0.0030.31 (0.14–0.66)0.0130.4 (0.19–0.83)Cancer site0.6191.21 (0.57–2.6)0.7111.14 (0.57–2.3)Lymph node affected0.9881.01 (0.47–2.17)0.2680.67 (0.34–1.35)CA staging0.3541.43 (0.67–3.05)0.2171.56 (0.77–3.14)Ann arbor staging0.3391.61 (0.61–4.25)0.5030.66 (0.2–2.2)Pinke0.0931.97 (0.89–4.33)0.1611.67 (0.82–3.42)Alb0.0190.31 (0.12–0.82)0.1220.53 (0.23–1.19)Ki670.6791.18 (0.54–2.54)0.0072.71 (1.31–5.58)EBV0.0163.31 (1.25–8.78)0.0053.9 (1.5-10.17)Treament method0.2620.65 (0.3–1.38)0.2470.66 (0.33–1.33)Radscore0.0005.19 (2.29–11.79)0.3570.66 (0.27–1.6)SUVmax0.0352.26 (1.06–4.84)0.1091.79 (0.88–3.63)Tumor volume0.1251.83 (0.84–3.97)0.8900.94 (0.42–2.12)SLR0.1750.59 (0.28–1.26)0.7710.9 (0.45–1.81)LLR0.1200.54 (0.25–1.17)0.8350.93 (0.46–1.88)CA staging: Chinese Southwest Oncology Group and Asia Lymphoma Study Group ENKTL staging, Alb: Albumin, EBV: Epstein-Barr virus, Radscore: Radiomics score, SLR: spleen-to-liver SUV ratio, LLR: lumbar-to-liver SUVmean ratio Univariate analysis of factors affecting OS and PFS CA staging: Chinese Southwest Oncology Group and Asia Lymphoma Study Group ENKTL staging, Alb: Albumin, EBV: Epstein-Barr virus, Radscore: Radiomics score, SLR: spleen-to-liver SUV ratio, LLR: lumbar-to-liver SUVmean ratio Table 3Prognostic performance of the three models for OS and PFS AUC(97.5%CI)C-index (95%CI)AICUnivariate COXOS (age + Alb + EBV + Radscore)82.7[68.8 − 92.3]0.787[0.720–0.898]196.0PFS (age + EBV + Ki67)79.7[67.7–90.3]0.724[0.649–0.824]237.3LASSOOS (age + Alb + EBV + Radscore)82.7[68.8 − 92.3]0.787[0.720–0.898]196.0PFS (age + EBV + Ki67)79.7[67.7–90.3]0.724[0.649–0.824]237.3BSROS (Alb +SUVmax +EBV+LLR+ Radscore)88.3[78.6 − 96.3]0.807[0.762–0.927]186.9PFS (age + EBV + Ki67)79.7[67.7–90.3]0.724[0.649–0.824]237.3OS: Overall survival, PFS: progression-free survival, LASSO: least absolute shrinkage and selection operator, BSR: Best Subset Regression, Alb: Albumin, EBV: Epstein-Barr virus, Radscore: Radiomics score, LLR: lumbar-to-liver SUVmean ratio, AUC: Area Under the ROC Curve, AIC: Akaike’s Information Criterion Prognostic performance of the three models for OS and PFS OS: Overall survival, PFS: progression-free survival, LASSO: least absolute shrinkage and selection operator, BSR: Best Subset Regression, Alb: Albumin, EBV: Epstein-Barr virus, Radscore: Radiomics score, LLR: lumbar-to-liver SUVmean ratio, AUC: Area Under the ROC Curve, AIC: Akaike’s Information Criterion For the PFS analysis, variables selected through univariate COX regression models, based on a criterion of P < 0.1, included age, Ki67 and EBV (Table 2). The BSR identified the following variables: age, Ki67, EBV, Ann arbor staging, CA staging, Radscore, tumor volume, and SUVmax. In contrast, the LASSO regression selected age, Alb, Ki67, Ann arbor staging, CA staging, EBV. Subsequently, we performed survival analysis using multivariate regression on these three sets of variables. The final variables, determined by the minimum AIC value, were identified as age, Ki67 and EBV (univariate COX/LASSO/BSR, AIC = 237.3) (Table 3). Comparison of models Through a comparative analysis of three models—univariate COX, LASSO, and BSR, it was observed that the univariate COX and LASSO yielded identical models in predicting OS. In contrast, the BSR-based model exhibited superior performance, with an AUC of 88.3 (97.5% CI: 78.6–96.3) and a C-index of 0.807 (95% CI: 0.762–0.927) as shown in Table 3; Fig. 2. Despite its higher accuracy in predictions of OS, the BSR model demonstrated performance comparable to that of the COX and LASSO models in predictions of PFS, as illustrated in Supplementary Fig. 1. Performance comparison of three prognostic models for overall survival (OS). Receiver operating characteristic (ROC) and decision curve analysis (DCA) curves comparing univariate COX regression, least absolute shrinkage and selection operator (LASSO), and best subset regression (BSR) models for predicting 1-, 2-, and 3-year OS. (A–C) ROC curves for 1-, 2-, and 3-year OS, respectively (Univariate COX / LASSO (overlap)). (D–F) DCA curves for 1-, 2-, and 3-year OS, respectively (Univariate COX / LASSO (overlap)) Performance and validation of prognostic models Nomograms were employed to visualize the outcomes of the model, resulting in the generation of OS and PFS nomograms (Fig. 3). Kaplan-Meier survival curves for OS and PFS are presented in Fig. 4, stratified by the BSR model-derived risk score and EBV status. Analyses incorporating additional variables from the BSR model are shown separately for OS and PFS in Supplementary Figs. 2 and 3, respectively. COX regression showed that variables from model of BSR were significant independent predictors of both OS and PFS (Table 4). Calibration curves calculated for 1, 2, and 3 years demonstrated strong concordance with the OS and PFS nomograms (Fig. 3). Furthermore, the DCA for OS prediction demonstrated that the prognostic model derived from BSR yielded the highest net benefit across the broadest clinically relevant range of threshold probabilities (approximately 20% to 60%) for both 1-year and 3-year predictions, outperforming models built using univariate COX or LASSO selection. While the univariate COX and LASSO models showed a marginal advantage in narrower, extreme ranges (60–70% for 1-year and 10–20% for 3-year predictions). It suggested that the three models showed nearly overlapping net benefits for OS only at 2 years, they exhibited complete overlap in net benefit for PFS at all evaluated time points (1, 2, and 3 years). Collectively, these findings indicate that clinical decisions informed by the BSR-derived model would offer the most favorable and robust trade-off in identifying high-risk patients. Prognostic nomograms and their calibration curves for overall survival (OS) and progression-free survival (PFS) prediction. (A) A nomogram developed based on the model derived from univariate COX and LASSO regression (shared model) for predicting OS. (B) Calibration curves for the nomogram in (A) at 1-, 2-, and 3-year time points, comparing predicted versus observed OS probabilities. (C) A nomogram developed based on the best subset regression (BSR) model for predicting OS. (D) Calibration curves for the nomogram in (C) at 1-, 2-, and 3-year time points. (E) A nomogram for predicting PFS, developed based on the consensus model (shared by univariate COX, LASSO, and BSR). (F) Calibration curves for the nomogram in (E) at 1-, 2-, and 3-year time points Kaplan-Meier survival analysis stratified by prognostic risk score and tumor EBER status. (A) Overall survival (OS) curves for patients stratified into high- and low-risk groups by the risk score derived from the OS-specific best subset regression (BSR) model. (B) Progression-free survival (PFS) curves for patients stratified into high- and low-risk groups by the risk score derived from the PFS-specific BSR model. (C) OS curves comparing patients with (EBVpos) and without (EBVneg) evidence of tumor EBER positivity. (D) PFS curves comparing patients with (EBVpos) and without (EBVneg) evidence of tumor EBER positivity Table 4Multivariable COX regression analyses for OS and PFS across the three models OS Models Variables P HR (95%CI) Univariate COX/LASSOage0.0089473.008 (1.3173–6.868)EBV0.0267143.207 (1.144–8.992)Radscore0.0002435.532 (2.219–13.794)Alb0.0514890.355 (0.126–1.007)BSRAlb0.009610.226 (0.073–0.697)SUVmax0.0245172.466 (1.123–5.416)EBV0.0179863.826 (1.259–11.630)LLR0.0005490.195 (0.077–0.493)Radscore5.05e-0715.463 (5.313–45.001) PFS Model Variables P HR (95%CI) Univariate COX/LASSO/BSRage0.031662.292 (1.076–4.884)Ki670.013342.608 (1.221–5.572)EBV0.003734.168 (1.588–10.936)OS: Overall survival, PFS: progression-free survival, LASSO: least absolute shrinkage and selection operator, BSR: Best Subset Regression, Alb: Albumin, EBV: Epstein-Barr virus, Radscore: Radiomics score, LLR: lumbar-to-liver SUVmean ratio Multivariable COX regression analyses for OS and PFS across the three models OS: Overall survival, PFS: progression-free survival, LASSO: least absolute shrinkage and selection operator, BSR: Best Subset Regression, Alb: Albumin, EBV: Epstein-Barr virus, Radscore: Radiomics score, LLR: lumbar-to-liver SUVmean ratio Discussion Our study leveraged a comparative modelling approach to elucidate the distinct biological drivers of short-term progression and long-term survival in NKTCL. The central finding is that the BRS is the optimal predict model for OS and PFS. The fact that selected variables based on the BSR model are fundamentally different for OS and PFS provides a nuanced understanding of prognosis of NKTCL and underscores the necessity of endpoint-specific risk stratification. Interpretation of the OS model The optimal OS model identified by the BSR algorithm integrates five key variables: Alb, EBV, SUVmax, LLR and Radscore. This composite model effectively captures a holistic prognostic panorama spanning from systemic host factors to metabolic activity and tumor-intrinsic features. Alb serves as a robust marker of systemic inflammation and nutritional status. Low Alb levels, indicative of a poor-performance status and a pro-tumorigenic systemic environment, are consistently associated with worse long-term survival across malignancies [9–11]. This underscores the critical impact of host factors on a patient’s ability to withstand both the disease and the rigours of long-term therapy. EBV is more than a mere marker of tumor burden, it is a direct driver of pathogenesis. Its integration into our model highlights its non-redundant role in predicting long-term outcomes. The potential mechanism underpinning its predictive superiority extend beyond viral infection, potentially involving the modulation of the tumor immune microenvironment (TIME). EBV can promote an immunosuppressive milieu by upregulating programmed death-ligand 1 (PD-L1) and recruiting regulatory T cells, thereby facilitating immune evasion and ultimately leading to poorer survival [12, 13]. This positions circulating EBV not only as a prognostic marker but also as a promising dynamic biomarker for guiding future immunotherapy strategies. The metabolic imaging biomarkers, SUVmax and LLR, provide complementary information. SUVmax, representing the peak metabolic activity within the most aggressive tumor subregion, is a well-established surrogate for high tumor proliferative activity and cellular density [14]. Our findings confirm its enduring prognostic value in NKTCL. More intriguing is the protective role of a pre-treatment LLR > 1. This suggests that elevated bone marrow metabolic activity prior to therapy may be a surrogate for a more robust and effective systemic immune response against the tumor, potentially reflecting better functional reserve and immune competence [15]. Conversely, the Radscore quantifies intratumoral heterogeneity from anatomical imaging, reflecting underlying malignant biological behavior beyond conventional size metrics and metabolic activity [16]. Intriguingly, our analysis revealed that these pre-treatment radiomic features held significant prognostic value for OS but not for PFS. This critical distinction suggests that pre-treatment imaging may more accurately reflect the fundamental, innate biological properties of the tumor which ultimately dictate long-term survival outcomes [17]. In contrast, PFS is more susceptible to the influence of treatment modalities, early therapeutic responses, and transient phenomena like pseudoprogression [18, 19]. This is consistent with an emerging paradigm wherein pre-treatment radiomics predict OS, while post-treatment functional imaging (e.g., interim PET/CT) is better suited for assessing treatment response and predicting PFS [20]. Therefore, the OS model synthesizes information across multiple levels, which provides a more robust and comprehensive definition of long-term survival risk. Interpreting the PFS model and the divergence from OS The optimal PFS model identified by BRS, comprised of Ki67, age and EBV, reflects a paradigm centered on tumor proliferative capacity and therapeutic tolerance. Ki67, a direct marker of tumor cell proliferation, is an intuitive driver of short-term disease progression. A high Ki67 index signifies rapid tumor growth and aggressiveness, directly increasing the likelihood of early relapse or resistance to first-line therapy [21–23]. Age may influence PFS indirectly, primarily through its correlation with tolerance to intensive therapy. Younger patients are often more capable of tolerating aggressive multi-agent regimens without dose reductions or delays, thereby sustaining longer remission periods [24]. Furthermore, age may serve as a surrogate for differences in tumor biology and resilience. The consistent presence of EBV in both OS and PFS models confirms its role as a core oncogenic driver critical to both short-term and long-term disease dynamics. The divergence between the PFS and OS models can be explained by a conceptual framework: PFS is more directly driven by treatment-sensitive, proliferative tumor biology (Ki67) and the patient’s capacity to tolerate therapy (age) [25]. In contrast, OS is additionally modulated by host factors (Alb) and the tumor’s long-term metastatic and immune-evasion potential (captured by Radscore and EBV’s role in TIME), alongside systemic metabolic and immune responses (LLR) [26]. Methodological and clinical implications The superior performance of our BSR model stems from its synergy with our study’s goal: to identify a sparse, robust prognostic signature from moderate-dimensional clinical-radiological data. Unlike univariate COX or LASSO methods, BSR’s key advantage is its exhaustive search, which reliably identifies the globally optimal variable combination, avoiding local optima [27]. This is crucial for capturing complex interactions in NKTCL biology. This approach minimizes bias and yields a more interpretable model for clinical translation. Thus, BSR is not just a statistical tool but a principled framework for our context, effectively distilling a complex feature space into a concise and high-performing prognostic model. In addition, our findings have immediate translational implications. For clinical trial design, PFS-based models could identify candidates for novel or intensified first-line therapies, while OS-based models might better select patients for pre-emptive supportive care or maintenance strategies. This facilitates a more personalized treatment approach. Study limitations and future directions Limitations of this study should be acknowledged. First, although the sample size was sufficient for model development, external validation in larger, multi‑center cohorts is necessary to confirm the generalizability of our findings. Second, the heterogeneity of treatment regimens across patients may have influenced the assessment of treatment‑related effects on progression‑free survival (PFS), and we were unable to fully account for these variations in the current analysis. Third, due to the limited sample size, post‑treatment PET imaging data were not included; however, this represents a promising avenue for future investigation. Lastly, the Radscore derived from radiomic features is inherently complex, which poses challenges for clinical interpretability. Future studies incorporating explainable artificial intelligence approaches may help address this limitation. Despite these limitations, our study provides valuable insights into the prognostic stratification of NKT‑cell lymphoma. Conclusion In conclusion, we have developed and validated a distinct prognostic model based on BSR for OS and PFS in NKTCL, revealing divergent biological pathways governing short-term progression and long-term survival. By moving beyond a one-size-fits-all approach, our models lay the groundwork for more personalized and endpoint-specific risk stratification, ultimately aiming to improve clinical outcomes. Their translation into routine practice necessitates further prospective validation. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Author contributions P.H. and C.X.L. contributed to the study conception and design and supervised the study conduction. T.F.F. and C.L.L. conducted the systematic retrieval. T.F.F. and C.L.L. collected the data and performed data analysis and were responsible for data interpretation and accuracy of the data analysis. T.F.F. and C.L.L. also wrote the original draft and P.H. and C.X.L. contributed to the main writing review and advised on study design. Finally, the manuscript was reviewed by all authors and approved for publication. Funding This research has received funding from new diagnosis and treatment protocol of radioiodine-refractory differentiated thyroid cancer (National Engineering Research Center for Isotopes and Drugs) (Grant No. 03010103KY_2023_11_22) and Postdoctoral Research Funding (Grant No. 0303020201P0613). Data availability Data used for analyses can be obtained from the corresponding author upon making a reasonable request. Declarations Ethical approval and consent to participate The study protocol was approved by the local Ethics Committee of the First Affiliated Hospital of Chongqing Medical University (ZZ2025-230-01) and the Ethics Committee of Chongqing University Cancer Hospital (CZLS2025159-A). This study was conducted in accordance with the Declaration of Helsinki. Clinical trial number: not applicable. Consent for publication Not applicable. 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. Fangfang Tian and Longlan Chen contributed equally to this work. Xiaoliang Chen and Hua Pang are co-corresponding authors. Contributor Information Xiaoliang Chen, Email: chengxiaoliang26@163.com. Hua Pang, Email: phua1973@163.com. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement Data used for analyses can be obtained from the corresponding author upon making a reasonable request.