Biology-informed risk stratification of glioblastoma by integrating MRI-based intratumoral heterogeneity with clinical features: a multicenter validation study Authors: Junfeng Zhang, Hong Guo, Xiaoxiao Ma, Yujue Zhong, Xiaojun Yu, Xiangbing Bian, Jianxing Hu, Caohui Duan, Yilong Huang, Jialing Wu, Mingliang Yang, Jian Hu, Xiaobo Zhang, Luhua Zhang, Rui Jiang, Xin Lou Published: 2026 ABSTRACT AbstractBackgroundRefined risk stratification before randomization is clinically important for reducing prognostic imbalance across study arms when evaluating novel therapies for glioblastoma. However, current artificial intelligence-assisted prognostic models are often limited by complex computational pipelines, limited bedside applicability, and insufficient biological interpretability. This study aimed to develop a prognostic model, supported by an online-accessible platform, for individualized risk stratification of glioblastoma using MRI-based intratumoral heterogeneity and routinely available clinical features, and to investigate the biological meaning of model-driven risk stratification.MethodsThis retrospective multicenter study included 836 patients with isocitrate dehydrogenase-wildtype glioblastoma from six centers between October 1996 and May 2025. The habitat risk score (HRS) for each patient was derived from a proposed intratumoral heterogeneity index and quantitative metrics extracted from three-dimensional preoperative MRI-based vascular habitat mappings. Independent predictors of overall survival (OS) were identified using Cox proportional hazards regression analysis, and three prognostic models (HRS model, clinical model, and radio-clinical model) were developed and validated in spatial and temporal external sets. Model interpretability was assessed using time-stratified Shapley additive explanations (SHAP) analysis. A web-based interactive platform was implemented for rapid individualized risk assessment. The biological meaning of model-driven risk stratification was explored using transcriptomic and histologic profiling.ResultsHRS, Karnofsky performance status (KPS), O6-methylguanine-DNA methyltransferase promoter methylation status, and extent of resection were identified as independent predictors of OS, with KPS and HRS contributing most strongly to survival prediction in SHAP analysis. The radio-clinical model demonstrated good predictive performance and outperformed the clinical model and the HRS model, with C-indexes of 0.74 and 0.77 in the spatial and temporal validation sets, respectively. It also effectively stratified patients into low- and high-risk groups regardless of first-line therapeutic regimen (log-rank, P < 0.05). Mechanistically, high-risk tumors showed increased tumor stemness and expanded HIF1α-positive regions, whereas low-risk tumors exhibited an immune-stimulatory phenotype. The deployed web-based platform enabled rapid patient-specific risk estimation to support bedside application.ConclusionsThis study establishes an interpretable and clinically deployable framework for glioblastoma risk stratification by integrating imaging-derived intratumoral heterogeneity with routine clinical features, without requiring complex computational infrastructure. The model also provides biologically grounded insights into model-driven risk stratification.Graphical Abstract Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s40164-026-00788-y.Keywords: Glioblastoma, Magnetic resonance imaging, Intratumoral heterogeneity, Risk stratification, Biological interpretability Background Glioblastoma is the most common and aggressive primary malignant brain tumor, and patient prognosis remains dismal despite maximal safe resection followed by chemoradiotherapy under the Stupp protocol [1, 2]. As newly diagnosed glioblastoma patients are increasingly enrolled in numerous prospective clinical trials evaluating novel therapeutic efficacy and combination strategies, refined risk stratification before randomization is essential to reduce prognostic imbalances between study arms, thereby minimizing confounding bias and enabling objective, reproducible, and comparable assessment of treatment efficacy [3, 4]. Recent advances in generalized foundation models have accelerated the development of artificial intelligence-assisted diagnosis, yet critical limitations persist [5]. Many models remain “black boxes” with limited interpretability, offering vague information on variable contributions or decision pathways [6–8]. Furthermore, model deployment requires dedicated computational infrastructure, high computational costs, and sustained technical maintenance, hindering integration with routine clinical workflows, particularly in non-tertiary healthcare institutions with insufficient informatization in developing countries [9]. These challenges underscore the need for easy-to-use prognostic tools that enhance bedside feasibility, data interpretability, and individualized risk stratification for glioblastoma. MRI (magnetic resonance imaging) features, including semantic features, quantitative physiological parameters, radiomics, and deep learning-associated features, have been widely explored to offer rapid risk estimation, transparent decision-making, and smooth integration into routine clinical practice [10–13]. Nevertheless, previous studies have primarily focused on imaging features extracted from the entire tumor region, overlooking intratumoral heterogeneity at the subregional level, which conveys additional pivotal prognostic information [14, 15]. Meanwhile, clinical prediction models, typically represented as nomograms, rely on manual scoring with vertical-line alignment in paper-based report documents, introducing potential subjective variability and low interactivity for parameter adjustment [16]. Additionally, with the shifting diagnostic criteria for glioblastoma confined to isocitrate dehydrogenase (IDH) wildtype tumors according to the 2021 World Health Organization (WHO) classification, prognostic models developed in earlier patient populations are prone to calibration drift and diminished precision for risk stratification [17–19]. More importantly, the biological meaning of model-driven risk stratification has not been fully investigated, providing insufficient mechanistic insight for clinicians and trial designers to guide individualized management and to verify the rationale underlying model output [20, 21]. These gaps highlight the urgent need to develop a biology-interpretable, clinic-feasible approach that integrates imaging features reflecting intratumoral heterogeneity for refined risk stratification of glioblastoma and to validate it rigorously in large-scale samples adhering to updated classification criteria. To the best of our knowledge, no studies addressing these issues have been reported. The purpose of our study is to develop an online-accessible prognostic model for overall survival (OS) prediction and individualized risk stratification of glioblastoma using MRI-based intratumoral heterogeneity metrics and routinely available clinical features, to validate the model spatiotemporally across multicenter external cohorts per the 2021 WHO classification, and to elucidate the biological underpinnings of model-driven risk stratification. Materials and methods Participants and study design This multicenter, retrospective study was conducted in three medical centers (centers A-C) and three public datasets (TCGA-GBM [center D], IvyGAP [center E], UPENN-GBM [center F]). Preoperative MRI data from patients with suspected supratentorial brain tumors were retrospectively collected and converted to the Neuroimaging Informatics Technology Initiative format from Digital Imaging and Communications in Medicine format. A total of 1778 glioblastoma patients were retrospectively collected. The inclusion criteria were: integrated diagnosis of newly diagnosed IDH-wildtype glioblastoma according to the 2021 WHO classification criteria; aged 18 years or older. We excluded 942 patients for the following exclusion criteria: a history of treatment before surgery or recurrent disease (n = 201); poor imaging quality or incomplete preoperative MRI sequences (n = 150); unknown O6-methylguanine-DNA methyltransferase (MGMT) promoter methylation status (n = 353); or incomplete clinical demographic data or follow-up records (n = 238). The remaining 836 patients were enrolled in this study. Consecutive eligible patients from center A (n = 349; January 2013-January 2022) constituted the training set. A spatial validation set (n = 315; October 1996-January 2022) was constructed from consecutive eligible patients at two external centers (centers B-C) and three public datasets (centers D-F) from The Cancer Imaging Archive. Consecutive eligible patients from center A (n = 172; February 2022-May 2025) formed the temporal validation set. For biological interpretability analysis, 63 consecutive eligible patients were included, comprising 24 patients from center E with paired MRI and transcriptomics data and 39 patients from center C with available tumor tissue blocks. The following clinical features were obtained for each patient: age, sex, Karnofsky performance status (KPS) at diagnosis, MGMT promoter methylation status, first-line therapy, and extent of resection (EOR). Overall survival (OS) was defined as the interval from the date of diagnosis to death from any cause, with censored observations defined at last follow-up. This study was approved by the institutional review boards of all participating institutions and was conducted in compliance with the Declaration of Helsinki, with informed consent waived because of the retrospective nature of the study, and only pre-existing medical data were used. The workflow of this study is shown in Fig. S1. MRI scanning protocol, image processing, and metrics measurement MRI data were acquired across all six participating centers, including T1-weighted, T2-weighted, fluid-attenuated inversion recovery (FLAIR), T1-weighted contrast-enhanced (T1-C), and dynamic susceptibility contrast (DSC) perfusion imaging sequences. The details of imaging protocols are provided in the Supplementary Materials. The MRI scanners, acquisition sequences, and imaging parameters for the preoperative multiparametric MRI examinations are summarized in Table S1. MRI vascular habitat mapping was generated using the vascular habitats service of ONCOhabitats (an online platform developed by Polytechnic University of Valencia (Universitat Politècnica de València); released on February 25, 2025; https://www.oncohabitats.upv.es/) to extract habitat-derived metrics as follows: (i) Image preprocessing: all MRI sequences underwent standardized image preprocessing, including correction of magnetic field inhomogeneities, removal of noise and spike artifacts, automated image registration, skull stripping, and voxel-intensity normalization, resulting in harmonized high-quality multiparametric MRI datasets. (ii) Tumor segmentation: glioblastoma tissues were then automatically segmented into enhancing tumor, necrosis, and nonenhancing edema using a three-dimensional U-net convolutional neural network trained for biomedical image segmentation based on the four structural MRI sequences (T1-weighted, T2-weighted, FLAIR, and T1-C) [22]. (iii) DSC perfusion quantification: DSC T2*-weighted images were registered to the T2-weighted anatomical images, and a gamma-variate function was applied to correct for Gd-based contrast agent leakage effect, allowing calculation of relative cerebral blood volume (rCBV) and mean transit time (MTT) mappings using established algorithms as previously described [23]. (iv) Generation of tumor vascular habitats: based on the structurally segmented images and registered DSC T2*-weighted images, distinct vascular habitats within the entire tumor volume were subsequently identified using an unsupervised clustering algorithm based on the directional class-adaptive spatially varying finite mixture model described by Juan-Albarracín et al. [24], which defined four vascular habitats: high angiogenic tumor (HAT), low angiogenic tumor (LAT), infiltrative peripheral edema (IPE), and vasogenic peripheral edema (VPE). (v) Measurement of habitat-derived metrics: 16 habitat-derived metrics were automatically computed, including the absolute volume and proportional volume fraction of each habitat as well as the median rCBV and MTT values for HAT, LAT, IPE, and VPE. The details of vascular habitat analysis are described in the Supplementary Materials. Due to differences in MRI scanner types, field strength, and acquisition protocols among centers, the ComBat harmonization method was applied using the sva R package to correct batch effects while preserving the intrinsic clinical heterogeneity of the real-world cohorts [25]. Definition and calculation of the intratumoral heterogeneity index An intratumoral heterogeneity (ITH) index was proposed to characterize tumor spatial heterogeneity by quantifying the distribution of the four vascular habitats generated above from a three-dimensional whole-tumor perspective by using connected component analysis. The procedures were as follows: (i) The three-dimensional tumor habitat images (processed via standardized Z-score normalization and resampled to 1 × 1 × 1 mm3) were obtained. Each habitat was treated as an independent label, resulting in four distinct labels corresponding to the four tumor vascular habitats analyzed in this study. (ii) The total voxel count within the entire tumor volume was calculated, and all vascular habitat labels were iteratively traversed. For each non-zero habitat label, the corresponding habitat region was binarized and extracted. (iii) A 26-connected neighborhood algorithm was applied to identify connected components within each non-zero habitat label. Noise components with a single-voxel volume were removed, and the number of connected components and the volume of the largest connected component for each habitat label were recorded. (iv) The ITH index was determined by two variables: the number of connected regions (denoted as ni) and the maximum volume of connected regions (denoted as Vi,max). More connected regions and a smaller maximum volume per tumor habitat indicate a more heterogeneous pattern. Thus, the ITH index was calculated as follows: Where represents the total tumor volume, and V denotes the number of habitats per tumor (set to four in this study). The ITH index ranges from 0 to 1. A higher ITH index indicates a more scattered distribution of tumor habitats within the tumor volume, which implies a more heterogeneous and chaotic cellular composition. For clarity, a schematic illustration of the ITH calculation is provided in Fig. S2. The code in Python for the ITH index calculation is available at https://github.com/301hospitalzhangjunfeng/Intratumoral-Heterogeneity-Index-Calculation. Definition and assessment of tumor resection The extent of resection (EOR) for glioblastoma patients from centers A-C was determined according to the Response Assessment in Neuro-Oncology (RANO) resect group classification [26]. In this study, RANO class 1 (supramaximal contrast-enhancing resection) and RANO class 2 (maximal contrast-enhancing resection) were categorized as gross total resection; RANO class 3 was categorized as subtotal resection; RANO class 4 was categorized as biopsy. If possible, tumor volumetric analysis was conducted using pre- and postoperative MRI within 72 h after resection to measure EOR in patients from centers A to C, as previously described [26, 27]. The contrast-enhancing tumor was volumetrically quantified on gadolinium-enhanced T1-weighted sequences, whereas nonenhancing tumor was assessed on FLAIR or T2-weighted imaging. To differentiate nonenhancing tumor from vasogenic edema or postoperative reactive changes, the anatomical tissue integrity of adjacent parenchyma, together with the signal intensity, was evaluated as previously described [26–29]. Other conventional sequences, including non-enhanced T1-weighted imaging, were also reviewed to distinguish postoperative hematoma from residual tumor. Postoperative T2/FLAIR abnormalities suspicious for ischemia were excluded by corresponding diffusion-weighted imaging, when required [29]. For patients with multifocal disease, volumes from all discrete tumor foci were summed to obtain a total tumor burden [30]. Absolute tumor volumes were measured, and patients were subsequently classified according to the RANO resect criteria for extent of resection. For patients from centers D to F, available EOR information was collected from The Cancer Imaging Archive (https://www.cancerimagingarchive.net/access-data/) for our review. Category of first-line therapeutic regimens The first-line therapeutic regimens following surgery in this study were summarized and prespecified as Stupp and non-Stupp regimens for therapy subgroup analysis. Specifically, non-Stupp regimens include chemoradiotherapy with non-Stupp regimens (lomustine + temozolomide/radiotherapy or lomustine + radiotherapy), chemotherapy only (temozolomide or lomustine), radiotherapy only, none or best supportive care, and not available (unavailable treatment history in medical records or documentation). Model development and validation An established rule of thumb for ensuring at least 10 events for each variable was used to calculate the sample size [31]. Hence, the training set of 349 patients with 334 events provided adequate power for developing a survival model. Univariable and multivariable Cox proportional hazards regression analyses were conducted to estimate hazard ratios and 95% confidence intervals of imaging and clinical variables for overall survival prediction. Specifically, the proportional hazards assumption of each variable was verified using Scaled Schoenfeld residuals. In accordance with recommendations from the Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD) [32], continuous variables that satisfied both the linearity assumption and the proportional hazards assumption were retained in their original form to preserve the full informational content inherent in continuous data. Because age, as a continuous variable, violated both the proportional hazards and linearity assumptions, it was transformed into three predefined categories as previously described: 18–39 years, 40–65 years, and > 65 years [33]. After collinearity diagnostics to ensure parameter estimation, variables of prognostic significance with P values less than 0.05 in univariable analysis were entered into a multivariable analysis based on Akaike information criterion and Bayesian information criterion. Three Cox analysis-based prognostic models were developed in the training set: a habitat risk score (HRS) model using HRS only; a clinical model using clinical features; and a radio-clinical model incorporating HRS and clinical features. Model performance was evaluated using the Harrell C-index, calibration curves, time-dependent receiver operating characteristic (ROC) analysis, and decision curve analysis. The areas under the ROC curves (AUCs) for 6-, 12-, and 24-month overall survival (OS) were calculated. C-index values of < 0.6, 0.6–0.7, and > 0.7 were considered indicative of poor, moderate, and good predictive performance, respectively. Internal validation of the final model was performed in the training cohort using bootstrap resampling with 1,000 repetitions to estimate optimism-corrected performance. To further examine model generalization and robustness, external validation was carried out in both spatial and temporal validation sets. Comparisons of the C-index and time-dependent AUCs between the reference model (clinical model) and the HRS model or radio-clinical model were conducted using the compareC and timeROC R packages, respectively. Furthermore, a web-based interactive graphical user interface derived from the radio-clinical model was deployed to enable rapid individualized risk assessment (available at: https://tmmubisuzjf.shinyapps.io/dynnomapp/). Risk stratification and survival analysis For risk stratification, the model-derived risk score of each patient was calculated according to the equation based on multivariable Cox regression analysis as follows: Where denotes the prognostic variable, is the corresponding regression coefficient, and the number of predictive variables is represented as p. Then, risk scores were ranked in ascending order, and the optimal cutoff value was determined on the training set using the surv_cutpoint function in the R package survminer as previously described [34]. The cutoff values derived from the training set were subsequently applied to the external validation sets and the biological set. Based on the risk score cutoffs derived from prognostic models, patients were stratified into low- and high-risk groups. Kaplan-Meier analysis and log-rank tests were used to compare OS between groups. The R scripts for Cox regression analysis and risk stratification are available at https://github.com/301hospitalzhangjunfeng/Statistical-Analysis. Gene expression, pathway enrichment, and immune infiltration analyses A total of 24 glioblastoma patients with paired MRI and transcriptomics data were obtained from the center E dataset. RNA-sequencing data in fragments per kilobase of transcript per million mapped reads format were downloaded as a gene-by-sample matrix and converted to transcripts per million (TPM). Batch effects were corrected using the ComBat algorithm. Differential gene expression between low- and high-risk groups was assessed on log2-transformed TPM values [log2(TPM + 1)] using the R package limma. Genes with |log2(fold change)|≥ 0.5 and a false discovery rate (FDR) < 0.05 were considered differentially expressed [35, 36]. Gene ontology enrichment analyses across the biological process, molecular function, and cellular component categories were performed using the R package clusterProfiler, with significance defined as FDR < 0.05. Gene set enrichment analysis was additionally conducted on the hallmark gene sets (https://www.gsea-msigdb.org/gsea/msigdb) for pathway enrichment, and gene set variation analysis scores were calculated to quantify patient-specific pathway activity (FDR < 0.05). Immune infiltration was evaluated using both the ESTIMATE (stromal score, immune score, ESTIMATE score, and tumor purity) and xCell (immune cell component) algorithms. Cytolytic activity was quantified using the cytolytic score described by Rooney et al. [37]. The code for transcriptomics analysis in R is available at https://github.com/301hospitalzhangjunfeng/Biological-Interpretability-Analysis. Spatial expression of stemness-associated genes by in situ hybridization analysis In situ hybridization data from the 24 patients in center E were used to assess the difference in tumor stemness between the low- and high-risk groups. As shown in Fig. S3, several anatomic feature regions in each tumor tissue section were predefined according to histologic structure in glioblastoma from hematoxylin & eosin-stained sections as previously described (https://glioblastoma.alleninstitute.org/) [38]: (i) leading-edge region (LE region); (ii) infiltrative tumor region (IT region); (iii) cellular tumor region (CT region); (iv) hypervascular region of cellular tumor (CThypervascular region); (v) perinecrotic region of cellular tumor (CTperinecrotic region); (vi) necrosis of cellular tumor (CTnecrosis region). Then, the data on stemness-associated genes in these distinct regions were downloaded from the IvyGAP dataset. Finally, the expression energy of each stemness-associated gene in the same anatomic structure region between low- and high-risk groups was compared. Immunofluorescence staining of tumor microenvironment markers Available formalin-fixed, paraffin-embedded tumor tissue blocks from 39 patients with glioblastoma in center C were collected and cut into consecutive 5 μm sections. A multiplex immunofluorescence staining protocol based on tyramide signal amplification (TSA) was performed on formalin-fixed, paraffin-embedded tumor sections to concurrently visualize CD31, hypoxia-inducible factor 1α (HIF1α), CD8α, and α-smooth muscle actin (αSMA), with 4′-6′-diamidino-2-phenylindole (DAPI) used for nuclear counterstaining. Tissue sections were deparaffinized, rehydrated, and subjected to antigen retrieval using AR9 buffer (AR900250ML; Akoya Biosciences, USA), followed by cooling to room temperature. After blocking endogenous peroxidase activity, primary antibodies against CD31, HIF1α, CD8α, and αSMA were applied sequentially, with each antibody incubated for 30 min at room temperature. Detailed product information and staining dilution for the primary antibodies used are summarized in Table S2. Corresponding HRP-conjugated secondary antibodies and TSA fluorophores were then applied iteratively. Signal development for each marker was achieved using a five-color TSA-PPD dye kit (TSA-N-8275; Panovue, Beijing, China). Between cycles, antibody stripping was performed to prevent cross-reactivity. After the final staining cycle, nuclei were counterstained with DAPI (D8714; Sigma-Aldrich, USA), and sections were mounted with antifade medium for imaging. Whole-slide fluorescent imaging was performed using an Olympus VS200 scanner (Olympus, Germany) equipped with a UPLXAPO 20 × objective. The acquired whole-slide fluorescence images were subsequently processed and analyzed using OLYMPUS OlyVIA software (version 2.9; Olympus Soft Imaging Solutions GmbH) and ImageJ (version 1.52a; National Institutes of Health, USA) as follows: multichannel fluorescence images were separated into individual channels and converted to grayscale images; positive-staining areas were identified using the same thresholding strategy across samples; and the percentage of positive-staining area per field area was calculated for each marker. For each sample, three consecutive sections representing the maximal tumor cross-sectional area were selected for analysis, and five random fields were chosen in each section under 20 × magnification. The percentages of positive-staining area from these fields were averaged for each sample. The box-whisker plots with actual data points were depicted by using GraphPad Prism (version 9.4.0; GraphPad Software Inc, USA). Statistical analysis Continuous variables with normal distribution are presented as means ± standard deviations (SDs); otherwise, medians with interquartile ranges (IQRs) are presented. Categorical variables are presented as n (%). Between-group comparisons were conducted using the Pearson χ2 test or the Fisher exact test for categorical variables. For continuous variables, the Student’s t test was used when normality and equal variances were met; otherwise, the Mann-Whitney U test was applied. The Kruskal–Wallis test was used to compare more than two groups. Time-stratified Shapley additive explanations (SHAP) analysis was performed using Python (version 3.10.16; Python Software Foundation) to evaluate how the relative contribution of each variable to the model-predicted risk varied across the observed survival intervals. Details of the SHAP methodology are described in the Supplementary Materials. For the calculation of median follow-up, the reverse Kaplan-Meier method was applied. Two-sided P < 0.05 was considered statistically significant. Statistical analyses were conducted using R (version 4.4.3; R Foundation for Statistical Computing) and SPSS (version 27.0.1.0; IBM SPSS Statistics). The R scripts used for statistical analysis are available at https://github.com/301hospitalzhangjunfeng/Statistical-Analysis. Results Data from 1778 patients with glioblastoma were initially collected from three centers and three publicly available datasets. After excluding 942 patients, 836 patients were enrolled (Fig. 1). Compared with the training set, the spatial and temporal validation sets had younger median ages (63 years vs. 59 years vs. 61 years, respectively; P < 0.001), a higher frequency of O6-methylguanine-DNA methyltransferase (MGMT) promoter methylation (36.4% vs. 39.0% vs. 48.8%, respectively; P = 0.02), a higher proportion of patients receiving the Stupp regimen (60.5% vs. 70.2% vs. 74.4%, respectively; P = 0.02), and longer median overall survival (12.3 months vs. 18.7 months vs. 17.6 months, respectively; P < 0.001) (Table 1), reflecting the clinical, molecular, and therapeutic heterogeneity across real-world patient cohorts over a prolonged time period. The pipeline of patient enrollment. GBM, glioblastoma; MGMT, O6-methylguanine-DNA methyltransferase Table 1Baseline of clinical characteristicsClinical FeaturesTraining Set (n = 349)Spatial Validation Set (n = 315)Temporal Validation Set (n = 172)P Value§Age at diagnosis (y)*63 (56–71)59 (49–66)61 (51–70) < 0.001 Number of patients aged 18–39 years9 (2.6)26 (8.3)24 (14.0) Number of patients aged 40–65 years188 (53.9)186 (59.0)85 (49.4)Number of patients aged > 65 years152 (43.5)103 (32.7)63 (36.6)Number of male patients206 (59.0)192 (61.0)93 (54.1)0.33Extent of resection < 0.001 GTR212 (60.7)233 (74.0)83 (48.3) Others†137 (39.3)82 (26.0)89 (51.7)MGMT promoter status0.02 Methylated127 (36.4)123 (39.0)84 (48.8) Unmethylated222 (63.6)192 (61.0)88 (51.2)KPS at diagnosis*90 (70–90)90 (80–90)90 (80–90)0.01First-line therapy following surgery0.02 Stupp regimen211 (60.5)221 (70.2)128 (74.4) Chemoradiotherapy with non-Stupp regimens71 (20.3)49 (15.5)25 (14.5) Chemotherapy only17 (4.9)13 (4.1)4 (2.3) Radiotherapy only11 (3.2)5 (1.6)7 (4.1) None or best supportive care34 (9.7)21 (6.7)8 (4.7) NA5 (1.4)6 (1.9)0 (0)Overall survival (mo)*12.3 (6.0–19.3)18.7 (10.4–30.2)17.6 (7.0–24.4) < 0.001Median follow-up (mo)*40.8 (21.7–46.8)37.7 (26.0–75.5)30.3 (25.8–35.1) < 0.001Unless otherwise indicated, data are numbers of patients, with percentages in parentheses. GTR, gross total resection; KPS, Karnofsky performance status; MGMT, O6-methylguanine-DNA methyltransferase; NA, not available*Data are medians, with interquartile ranges in parentheses†Subtotal resection and biopsy§P values are for comparison of variables across all sets Baseline of clinical characteristics Unless otherwise indicated, data are numbers of patients, with percentages in parentheses. GTR, gross total resection; KPS, Karnofsky performance status; MGMT, O6-methylguanine-DNA methyltransferase; NA, not available *Data are medians, with interquartile ranges in parentheses †Subtotal resection and biopsy §P values are for comparison of variables across all sets Univariable and multivariable Cox analyses to identify predictors of overall survival Univariable Cox analysis identified 12 quantitative imaging-derived metrics associated with OS in the training set (all P < 0.05; Table 2). After three metrics were excluded because of collinearity (Table S3), multivariable Cox analysis identified the intratumoral heterogeneity (ITH) index (hazard ratio [HR], 14.49; 95% confidence interval [CI]: 3.64, 57.65; P < 0.001), relative cerebral blood volume (rCBV) in high angiogenic tumor (HAT) (HR, 1.18; 95% CI: 1.12, 1.25; P < 0.001), infiltrative peripheral edema (IPE) volume (HR, 1.04; 95% CI: 1.01, 1.06; P = 0.002), and low angiogenic tumor (LAT) fraction (HR, 13.28; 95% CI: 4.14, 42.63; P < 0.001) as independent predictors (Table 2). Based on these parameters and corresponding coefficients, the habitat risk score (HRS) of each patient was calculated as follows: Table 2Univariable and multivariable Cox regression analyses to determine imaging predictors of overall survival in the training setVariablesUnivariable AnalysisMultivariable AnalysisRegression CoefficientHazard RatioP valueRegression CoefficientHazard RatioP valueITH index2.5312.49 (3.77, 41.32) < 0.0012.6714.49 (3.64, 57.65) < 0.001rCBV in HAT0.161.18 (1.12, 1.23) < 0.0010.171.18 (1.12, 1.25) < 0.001rCBV in LAT0.201.22 (1.14, 1.31) < 0.0010.051.05 (0.80, 1.37)0.71rCBV in IPE0.131.14 (1.02, 1.28)0.03-0.030.97 (0.79, 1.19)0.79rCBV in VPE0.031.03 (0.84, 1.26)0.81MTT in HAT0.011.01 (0.42, 2.46)0.98MTT in LAT1.133.08 (1.15, 8.28)0.030.832.29 (0.53, 9.92)0.27MTT in IPE1.785.93 (1.81, 19.36)0.003-0.590.56 (0.09, 3.35)0.52MTT in VPE0.481.62 (0.52, 5.11)0.41HAT volume*0.071.07 (1.05, 1.10) < 0.001LAT volume*0.041.04 (1.03, 1.05) < 0.001IPE volume0.071.07 (1.05, 1.09) < 0.0010.041.04 (1.01, 1.06)0.002VPE volume0.001.00 (0.99, 1.01)0.65HAT fraction2.7816.16 (2.57, 101.61)0.0030.972.65 (0.18, 40.06)0.48LAT fraction2.087.97 (3.02, 21.05) < 0.0012.5913.28 (4.14, 42.63) < 0.001IPE fraction1.243.44 (0.79, 14.97)0.10VPE fraction*-1.240.29 (0.16, 0.54) < 0.001Data in parentheses are 95% confidence intervals. Variables with P < 0.05 at univariable analysis were included in the multivariable analysis. HAT, high angiogenic tumor; ITH, intratumoral heterogeneity index; IPE, infiltrative peripheral edema; LAT, low angiogenic tumor; MTT, mean transit time; rCBV, relative cerebral blood volume; VPE, vasogenic peripheral edema*HAT volume, LAT volume, and VPE fraction were excluded from the multivariable analysis because of collinearity Univariable and multivariable Cox regression analyses to determine imaging predictors of overall survival in the training set Data in parentheses are 95% confidence intervals. Variables with P < 0.05 at univariable analysis were included in the multivariable analysis. HAT, high angiogenic tumor; ITH, intratumoral heterogeneity index; IPE, infiltrative peripheral edema; LAT, low angiogenic tumor; MTT, mean transit time; rCBV, relative cerebral blood volume; VPE, vasogenic peripheral edema *HAT volume, LAT volume, and VPE fraction were excluded from the multivariable analysis because of collinearity Upon inclusion of clinical variables, HRS, MGMT promoter methylation status, gross total resection (GTR) status, and Karnofsky performance status (KPS) at diagnosis remained independent predictors of OS (Table 3). As such, an optimal radio-clinical model consisting of the four variables was developed, yielding the highest C-index of 0.78 and the lowest Akaike information criterion and Bayesian information criterion values (3043.70 and 3058.95, respectively) in the training set (Table S4). Table 3Univariable and multivariable Cox regression analyses for predictors of overall survival after incorporating the habitat risk score and clinical features in the training setVariablesUnivariable AnalysisMultivariable AnalysisRegression CoefficientHazard RatioP ValueRegression CoefficientHazard RatioP ValueHRS1.02.72 (2.25, 3.28) < 0.0010.892.42 (1.98, 2.95) < 0.001Age (y) 18–39ReferenceReferenceReferenceReference 40–650.561.75 (0.86, 3.55)0.130.041.04 (0.51, 2.15)0.91 > 650.862.36 (1.16, 4.82)0.020.271.31 (0.63, 2.73)0.48Sex MaleReferenceReference Female-0.050.95 (0.76, 1.19)0.66MGMT promoter status UnmethylatedReferenceReferenceReferenceReference Methylated-0.450.64 (0.51, 0.80) < 0.001-0.390.67 (0.54, 0.85) < 0.001Extent of resection Others†ReferenceReferenceReferenceReference GTR-0.560.58 (0.46, 0.72) < 0.001-0.290.75 (0.59, 0.95)0.02KPS-0.050.95 (0.94, 0.96) < 0.001-0.050.96 (0.95, 0.96) < 0.001Data in parentheses are 95% confidence intervals. Variables with P < 0.05 at univariable analysis were included in the multivariable analysis. GTR, gross total resection; HRS, habitat risk score; KPS, Karnofsky performance status; MGMT, O6-methylguanine-DNA methyltransferase†Subtotal resection, partial resection, and biopsy Univariable and multivariable Cox regression analyses for predictors of overall survival after incorporating the habitat risk score and clinical features in the training set Data in parentheses are 95% confidence intervals. Variables with P < 0.05 at univariable analysis were included in the multivariable analysis. GTR, gross total resection; HRS, habitat risk score; KPS, Karnofsky performance status; MGMT, O6-methylguanine-DNA methyltransferase †Subtotal resection, partial resection, and biopsy Model performance In bootstrap-based internal validation, the radio-clinical model showed good robustness, with a bootstrap-corrected C-index of 0.78 (95% CI: 0.76, 0.80) (Table S5). In external validation, the model maintained good predictive performance, achieving a C-index of 0.74 (95% CI: 0.71, 0.78) in the spatial validation set and 0.77 (95% CI: 0.73, 0.81) in the temporal validation set. It outperformed the clinical model and the HRS model, with favorable time-dependent discrimination across all sets (Table 4; Table S6). The time-dependent areas under the ROC curves (AUCs) of the radio-clinical model for 6-, 12-, and 24-month OS prediction were 0.96 (95% CI: 0.93, 0.98), 0.86 (95% CI: 0.77, 0.90), and 0.73 (95% CI: 0.67, 0.79) in the spatial validation set, respectively, and 0.95 (95% CI: 0.92, 0.98), 0.86 (95% CI: 0.80, 0.92), and 0.78 (95% CI: 0.69, 0.87), in the temporal validation set, respectively. AUCs for 6- and 12-month OS prediction of the radio-clinical model were higher than those of the clinical model and the HRS model (Fig. 2; Table 4). Table 4Performance of three models in predicting overall survival up to 24 months in the training and external validation setsModels and SetsC-IndexP Value*AUC6mOSP Value*Clinical model Training set0.73 (0.71, 0.76)Reference0.82 (0.77, 0.88)Reference Spatial validation set0.68 (0.65, 0.72)Reference0.81 (0.74, 0.88)Reference Temporal validation set0.71 (0.66, 0.76)Reference0.85 (0.79, 0.93)ReferenceHRS model Training set0.69 (0.67, 0.72)0.020.78 (0.73, 0.83)0.27 Spatial validation set0.68 (0.65, 0.73)0.860.92 (0.88, 0.97)0.01 Temporal validation set0.73 (0.68, 0.78)0.550.87 (0.81, 0.93)0.81Radio-clinical model Training set0.78 (0.76, 0.81) < 0.0010.88 (0.84, 0.92) < 0.001 Spatial validation set0.74 (0.71, 0.78)0.0010.96 (0.93, 0.98) < 0.001 Temporal validation set0.77 (0.73, 0.81)0.020.95 (0.92, 0.98)0.02Models and SetsAUC12mOSP Value*AUC24mOSP Value*Clinical model Training set0.83 (0.79, 0.87)Reference0.79 (0.74, 0.84)Reference Spatial validation set0.76 (0.69, 0.82)Reference0.74 (0.68, 0.80)Reference Temporal validation set0.73 (0.65, 0.81)Reference0.83 (0.77, 0.90)ReferenceHRS model Training set0.79 (0.74, 0.84)0.240.79 (0.74, 0.85)0.91 Spatial validation set0.80 (0.74, 0.86)0.350.65 (0.58, 0.71)0.03 Temporal validation set0.83 (0.81, 0.93)0.030.70 (0.58, 0.81)0.03Radio-clinical model Training set0.89 (0.86, 0.93) < 0.0010.87 (0.82, 0.91) < 0.001 Spatial validation set0.86 (0.77, 0.90) < 0.0010.73 (0.67, 0.79)0.72 Temporal validation set0.86 (0.80, 0.92) < 0.0010.78 (0.69, 0.87)0.29Data in parentheses are 95% confidence intervals*P value for comparison of the C-index or area under the receiver operating characteristic curve with that of the clinical model (reference) in the corresponding set. AUC, area under the receiver operating characteristic curve; HRS, habitat risk score; OS, overall survival; 6 m, 6-month; 12 m, 12-month; 24 m, 24-month Performance of three models in predicting overall survival up to 24 months in the training and external validation sets Data in parentheses are 95% confidence intervals *P value for comparison of the C-index or area under the receiver operating characteristic curve with that of the clinical model (reference) in the corresponding set. AUC, area under the receiver operating characteristic curve; HRS, habitat risk score; OS, overall survival; 6 m, 6-month; 12 m, 12-month; 24 m, 24-month Discrimination performance of models in predicting 6-, 12-, and 24-month overall survival. Graphs show time-dependent areas under the receiver operating characteristic curves for the habitat risk score model (A), clinical model (B), and radio-clinical model (C) in the training set (left column), spatial validation set (middle column), and temporal validation set (right column). HRS, habitat risk score; AUC, area under the receiver operating characteristic curve Additionally, the radio-clinical model demonstrated good concordance between predicted and observed probabilities in bootstrap-based internal validation and external validation sets, with the favorable Brier scores of 0.07 (95% CI: 0.05, 0.09), 0.13 (95% CI: 0.11, 0.15), and 0.22 (95% CI: 0.19, 0.26) for 6-month, 12-month, and 24-month OS prediction in the spatial validation set, respectively, and 0.10 (95% CI: 0.07, 0.13), 0.18 (95% CI: 0.14, 0.22), and 0.15 (95% CI: 0.10, 0.19) in the temporal validation set, respectively (Fig. S4; Table S5; Table S7). Decision curve analysis indicated that, across most threshold probabilities, the radio-clinical model provided greater net benefit than the clinical model and the HRS model for predicting 6- and 12-month OS, whereas its net-benefit advantage over the clinical model and the HRS model was less apparent for 24-month OS prediction (Fig. S5). Risk stratification by models Optimal cutoff values of HRS, clinical risk score, and radio-clinical risk score were − 0.17, − 0.84, and − 0.70, respectively. All three model-derived risk scores effectively divided patients into low- and high-risk groups, with high-risk patients exhibiting shorter OS than low-risk patients (log-rank test, all P < 0.05; Fig. 3). Considering that first-line therapeutic regimens are important factors affecting patient prognosis, we analyzed the interaction between the radio-clinical risk score and therapeutic regimens. The results demonstrated that the radio-clinical risk score was an independent predictor of OS (HR, 2.34 [95% CI: 2.03, 2.69]; P < 0.001), with no statistically significant interaction with therapeutic regimen (P = 0.54; Table S8), suggesting robust predictive performance of the radio-clinical risk score. Similarly, survival analysis in therapeutic subgroups demonstrated that the optimal cutoff for radio-clinical risk score effectively stratified patients into low- and high-risk groups with distinct OS in both the Stupp and the non-Stupp therapy subgroups (log-rank test, all P < 0.05; Figs. S6A-S6F). Overall survival analysis of low- and high-risk groups stratified by model-derived risk scores. Plots show Kaplan–Meier survival curves of low- and high-risk groups stratified by the habitat risk score (A), clinical risk score (B), and radio-clinical risk score (C) in the training set (left column), spatial validation set (middle column), and temporal validation set (right column). Plots below the Kaplan–Meier survival curves show the distribution of risk scores and predictive variables for individual patients. HRS, habitat risk score; MGMT, O6-methylguanine-DNA methyltransferase; GTR, gross total resection; KPS, Karnofsky performance status SHAP analysis and clinical feasibility of the radio-clinical model In the training set, SHAP analysis demonstrated a positive association between HRS and the predicted risk, whereas KPS, MGMT promoter methylation status, and GTR status showed inverse associations. HRS and KPS were the major contributors to OS prediction. Notably, the contribution of individual features varied over time (Fig. 4A-D). To improve bedside applicability and compatibility with contemporary electronic medical systems, an online-accessible platform with an interactive graphical user interface was developed based on the radio-clinical model, enabling rapid calculation of individualized survival probabilities after entry of patient-specific parameters. The platform was demonstrated using two representative patients: the high-risk patient exhibited 6-, 12-, and 24-month OS probabilities of 0.28, 0.01, and 0, while the low-risk patient showed 0.97, 0.90, and 0.62, respectively (Fig. 5). Feature importance analysis of the radio-clinical model for overall survival prediction. A The summary plot shows the overall feature importance for the radio-clinical model in the training set. Each dot represents a single patient, and color denotes the feature value (red for high, blue for low) ranked by mean absolute SHAP value. B The bar plot shows mean absolute SHAP values ranking the relative importance of predictive features. C The stacked plot shows the changes in mean absolute and relative SHAP values of each feature across observed survival intervals. D The force plot demonstrates the individualized contribution of each feature for a typical patient in the spatial validation set with a habitat risk score of 1.36, Karnofsky performance status of 80, unmethylated status of O6-methylguanine-DNA methyltransferase promoter, and non-gross total resection. HRS, habitat risk score; KPS, Karnofsky performance status; MGMT, O6-methylguanine-DNA methyltransferase; GTR, gross total resection Representative use of the online platform based on the radio-clinical model for overall survival prediction. The MRI scans shown are axial sections. Patient A and patient B had similar ages and the same sex, but they differed in O6-methylguanine-DNA methyltransferase promoter methylation status (negative vs. positive), gross total resection status (negative vs. positive), and Karnofsky performance status (60 vs. 80). The vascular habitat mappings show that patient A’s tumor exhibited greater intratumoral heterogeneity than patient B’s tumor did, with a higher habitat risk score of 5.08 than that of 2.67 in patient B. After inputting patient-specific parameters, patient A was stratified into the high-risk group and had poor predicted survival probabilities of 0.28, 0.01, and 0 at 6, 12, and 24 months, respectively, whereas patient B was stratified into the low-risk group and had corresponding survival probabilities of 0.97, 0.90, and 0.62, respectively. KPS, Karnofsky performance status; MGMT, O6-methylguanine-DNA methyltransferase; GTR, gross total resection; OS, overall survival Biological interpretability of the radio-clinical model A total of 968 differentially expressed genes were identified between the low- and high-risk groups (Fig. 6A), and the top 20 genes are summarized in Table S9. Pathway enrichment analysis showed that metabolism and cell proliferation-associated pathways, such as oxidative phosphorylation, E2F transcription factor targets, and G2/M cell cycle (G2M) checkpoint, were significantly upregulated in the high-risk group, whereas interferon-γ–related response, inflammatory response, and tumor necrosis factor signaling were predominantly activated in the low-risk group, indicating distinct tumor proliferation/stemness and immune microenvironment profiles between the two groups (Fig. 6B, C; Table S10). These findings were validated by using patient-specific pathway enrichment based on gene set variation analysis (Fig. 6D; Table S11). To further investigate the difference in immune microenvironment between the two groups, several immune infiltration analyses were conducted. Heterogeneity in immune cell infiltration and weak correlation among immune cell subpopulations were observed in all tumor samples (Fig. 7A, B). Multiple immune cell infiltration analyses confirmed that immune scores, M1-macrophage infiltration, and activated myeloid dendritic cells were prominently higher in the low-risk group (Fig. 7C, D; Table S12), whereas tumor purity exhibited the opposite result (Fig. 7E). However, no evidence of a statistically significant difference in cytolytic activity was observed between the two risk groups (Fig. 7F). Consistent with the transcriptomics results, histologic staining verified that CD8α- and α-smooth muscle actin (αSMA)-positive areas were markedly increased in the low-risk tumors, whereas hypoxia-inducible factor 1α (HIF1α)-positive regions were expanded in the high-risk group, indicative of an immune-stimulatory microenvironment in low-risk patients (Fig. 7G-K). Given that the oxidative phosphorylation, E2F targets, and G2M checkpoint pathways were upregulated in the high-risk group, which implies tumor stemness activation, we performed in situ hybridization analysis to characterize the spatial expression of stemness-associated genes within the predefined tumor subregions for validation (Fig. 8A; Table S13). The findings demonstrated that expression of most stemness-associated genes was elevated in the high-risk group compared with the low-risk group, especially in the leading-edge region (Fig. 8B; Figs. S7-S11; Tables S14-S19). Differentially expressed genes and pathway enrichment analyses of low- and high-risk groups stratified by the radio-clinical model. A Volcano plot of differentially expressed genes between low- and high-risk groups. The top 10 genes with the highest fold change are marked. B, C Bar charts of gene and pathway enrichments by gene ontology analysis (left) and gene set enrichment analysis (right). D Comparisons of gene set variation analysis scores of the top 10 enriched pathways between low- and high-risk groups Characterization of differences in the tumor immune microenvironment between low- and high-risk groups. A The heatmap shows the Spearman correlation among immune cell types. B The stacked bar chart shows the relative proportion of immune cells in all samples. C Comparisons of ESTIMATE, immune, and stromal scores between low- and high-risk groups. D Comparisons of xCell-inferred immune cell infiltration between the two risk groups. E Comparison of tumor purity. F Comparison of cytolytic activity (CYT). G–K Comparisons of tumor microenvironment markers between low- and high-risk groups. CYT, cytolytic activity; HIF1α, hypoxia-inducible factor 1α; αSMA, α-smooth muscle actin Comparisons of stemness-associated gene expression between low- and high-risk groups stratified by the radio-clinical model. A Representative images of stemness-associated gene expression in low- and high-risk patients by in situ hybridization analysis. B The box-whisker plots with actual data points show comparisons of stemness-associated gene expression in the leading-edge region between low- and high-risk groups. LE, leading-edge Discussion Novel therapeutic strategies and clinical trials are actively being pursued to improve outcomes in glioblastoma. However, the lack of prognostic tools that combine clinical feasibility with biological interpretability continues to limit efforts to reduce prognostic imbalance across study arms before randomization. Existing models often have limited interactivity, lack validation in contemporary real-world populations, and provide insufficient biologic rationale for their predictions. In this study, we developed a user-friendly prognostic model embedded in an online platform that integrates MRI-derived intratumoral heterogeneity metrics with clinical features to predict overall survival in glioblastoma. The model demonstrated good discrimination in both spatial and temporal validation cohorts, with C-indexes ranging from 0.74 to 0.77. Although direct cross-study comparisons should be made cautiously, this performance appears comparable to, and in some instances better than, that reported for several previously published prognostic models [21]. Moreover, the radio-clinical risk score effectively stratified patients into low- and high-risk groups regardless of therapeutic regimen. Importantly, we also provided biologically interpretable evidence supporting this data-driven risk stratification by demonstrating distinct immune microenvironment and stemness-related features between the two risk groups. Advanced MRI enables noninvasive assessment of tumor physiology [39]. Although prior studies have shown that physiologic parameters such as rCBV are associated with survival in glioblastoma and that radiomics may further enhance prediction [40, 41], most studies have focused on features extracted from the entire tumor and have not fully captured prognostically relevant spatial heterogeneity [42]. To address this limitation, we performed MRI-based tumor habitat analysis and extracted a series of quantitative physiologic metrics from vascular habitat mappings to quantify heterogeneity at the subregional level. We also proposed an intratumoral heterogeneity index that characterizes spatial heterogeneity from a three-dimensional whole-tumor perspective by quantifying the distribution of habitat components. The intratumoral heterogeneity index, median rCBV in the high angiogenic tumor habitat, infiltrative peripheral edema volume, and the fraction of low angiogenic tumor habitat were identified as independent predictors of OS. These variables were then integrated into the habitat risk score, which captures the prognostic information encoded in tumor spatial heterogeneity. Given that clinical features also substantially influence prognosis, we evaluated HRS together with clinical features and ultimately developed a radio-clinical model incorporating HRS, KPS, MGMT promoter methylation status, and gross total resection status. This model demonstrated robust performance in validation sets and outperformed the model based solely on clinical features, underscoring the incremental value added by HRS. Interestingly, decision curve analysis showed that the net-benefit advantage of the radio-clinical model over the clinical model was less apparent for predicting 24-month OS. This partially reflects an inherent limitation of preoperative MRI data, which captures tumor biology only at baseline. Preoperative MRI features are informative for early and intermediate prognosis in glioblastoma, but their predictive value may attenuate with treatment response and posttreatment tumor evolution. Recurrent glioblastoma often represents substantial molecular and phenotypic divergence from the initial tumor, indicating that preoperative imaging features may not fully recapitulate the determinants of long-term survival [43–45]. Thus, the decreasing relative contribution of HRS and the increasing contribution of KPS in longer-term survival prediction suggest that baseline habitat heterogeneity may be particularly relevant to early and mid-term risk stratification, whereas long-term prognosis is more strongly influenced by interval clinical course and evolving tumor biology. The development of longitudinal prognostic models integrating dynamic MRI data for improved long-term survival prediction is warranted. Although many predictive models have shown promising performance, their clinical implementation remains difficult because they often depend on high-performance computational infrastructure, specialized analytic platforms, and complex deployment workflows [8, 46]. Thus, we implemented the radio-clinical model through a web-based graphical user interface to improve clinical usability and facilitate bedside application. By replacing a technically demanding analytical workflow with an interactive interface, the platform improves translational potential and represents a practical step toward real-world implementation. However, the current platform does not yet fully integrate automated image preprocessing, habitat quantification, and ITH calculation, and therefore still depends on postprocessed inputs. A fully automated end-to-end pipeline would further simplify use, improve robustness and reproducibility, and strengthen the model’s applicability across diverse clinical settings. An important finding of our study is that the radio-clinical risk score served as an independent predictor of OS, effectively stratifying patients into low- and high-risk groups. Subgroup analyses across Stupp and non-Stupp regimens demonstrated the stability of this stratification across different therapy interventions. Biological analysis further validated the mechanistic underpinnings of this risk stratification. High-risk tumors mainly displayed upregulated oxidative phosphorylation, reflecting elevated proliferative activity mediated by stemness-associated metabolic reprogramming, whereas low-risk tumors demonstrated enrichment of interferon-mediated antitumor immune pathways, indicative of robust effector T cell infiltration and antitumor immune response [47, 48]. Histology and in situ hybridization findings further confirmed that high-risk tumors displayed reduced CD8 T cell infiltration and elevated expression of stemness-associated genes, particularly at the invasive margins. This finding is consistent with previous studies demonstrating that the prominent stemness and immunosuppressive phenotype of residual tumor cells at the tumor leading edge promote therapy resistance and tumor progression, and may partially explain why high-risk patients have a poor prognosis [25, 49–51]. Cytolytic score is a marker of cytolytic immune activity across multiple cancer types, and higher levels of immune cytotoxicity are commonly observed in tumors with pronounced immune infiltration [52]. Similarly, higher cytolytic activity is associated with higher histological grade and more aggressive phenotypes in gliomas [53]. However, our study found no statistically significant difference in cytolytic activity between low-risk and high-risk groups. This discrepancy may be attributed to several factors. The small sample size of the biological set in our study could lead, at least in part, to insufficient statistical power. Furthermore, bulk transcriptomic deconvolution may be affected by technical noise and biological confounding arising from mixed cell populations, reference profile selection, and tumor-specific expression, which may compromise the accuracy of immune cell quantification [54, 55]. Sampling bias may also contribute to cytolytic score evaluation. Immune infiltration can vary substantially within a tumor, and local sampling may not capture the full spatial heterogeneity of the immune microenvironment [56]. Collectively, although larger sample sizes and multidimensional omics data are required for further validation, our findings provide preliminary biology-informed evidence supporting the risk stratification achieved by the radio-clinical risk score. Our study has several limitations. First, the retrospective design may have introduced inherent selection bias, although independent external validation was performed. Second, the transcriptomic analysis was based on bulk RNA sequencing only. Given the marked cellular heterogeneity of glioblastoma, future prospective studies incorporating single-cell and spatial transcriptomics are warranted to identify key molecular drivers of prognosis. Finally, the HRS was derived only from conventional and perfusion MRI sequences and may not fully capture metabolic or proliferative heterogeneity. Future work incorporating metabolic imaging and advanced multi-omics profiling may further improve both interpretability and generalizability. Conclusion This study proposes a habitat risk score that quantifies intratumoral heterogeneity from a three-dimensional whole-tumor perspective using multiparametric MRI-based vascular habitat mappings and integrates it with clinical features to construct a radio-clinical model for overall survival prediction in glioblastoma. Embedded in an online-accessible interactive platform, this model demonstrates favorable performance for survival prediction and risk stratification and shows potential for integration into routine clinical workflows to support bedside application. Furthermore, distinct immune and stemness-related phenotypes were identified across risk groups, providing biologically grounded evidence for model-based risk stratification. Overall, this biology-informed radio-clinical model offers a promising tool for optimizing treatment planning, patient management, and clinical trial enrollment. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1. Supplementary Material 2. Supplementary Material 3. Acknowledgements We thank the Cancer Imaging Archive, the Cancer Genome Atlas, the University of Pennsylvania Health System, and the Ivy Glioblastoma Atlas Project for providing open-source datasets. Abbreviations Extent of resection Gross total resection Habitat risk score Karnofsky performance status O6-methylguanine-DNA methyltransferase Overall survival Isocitrate dehydrogenase Relative cerebral blood volume Mean transit time High angiogenic tumor Low angiogenic tumor Infiltrative peripheral edema Vasogenic peripheral edema Intratumoral heterogeneity Hypoxia-inducible factor 1α α-smooth muscle actin Author contributions J.Z., H.G., and X.M. contributed to the conceptualization, data analysis and curation, methodology, statistical analysis, software development, and writing the original and revised drafts; Y.Z., and X.Y. contributed to the data collection, data preprocessing, and resources; X.B., J.H., C.D., Y.H., and J.W. contributed to the data acquisition and imaging methodology; M.Y. and J.H. contributed to data visualization; X.Z., L.Z., and R.J. contributed to data interpretation; X.L. contributed to the conceptualization, study supervision, funding support, and project administration. Funding This study was supported by the National Natural Science Foundation of China (grants 82502283 and 82327803) and the Youth Science and Technology Talent Program of General Hospital of Western Theater Command (grant 414378). Data availability To protect patients’ privacy in the Chinese PLA General Hospital, General Hospital of Western Theater Command of PLA, and Army Medical University, the raw MRI data are private and not publicly available. Deidentified patient-level clinical and survival data reported in this paper will be shared by the corresponding author (louxin@301hospital.com.cn) upon reasonable request. Declarations Ethics approval and consent to participate This retrospective study was conducted in compliance with the Declaration of Helsinki, with informed consent waived because of the retrospective nature of the study, and only pre-existing medical data were used. Ethics approval was obtained by the institutional review boards of Army Medical Center (Chongqing, China; Approval No. 2024[75]), Chinese PLA General Hospital (Beijing, China; Approval No. S2021-542-01), and General Hospital of Western Theater Command of PLA (Chengdu, China; Approval No. 2024EC7-ky037). Consent for publication This retrospective study was approved by the institutional review boards of all participating institutions, with informed consent waived. The consent for publication was approved by the institutional review boards of all participating institutions. 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. Junfeng Zhang, Hong Guo and Xiaoxiao Ma contributed equally to this work. References Associated Data Supplementary Materials Supplementary Material 1. Supplementary Material 2. Supplementary Material 3. Data Availability Statement To protect patients’ privacy in the Chinese PLA General Hospital, General Hospital of Western Theater Command of PLA, and Army Medical University, the raw MRI data are private and not publicly available. Deidentified patient-level clinical and survival data reported in this paper will be shared by the corresponding author (louxin@301hospital.com.cn) upon reasonable request.