An Integrated Multi-Compartment Biomarker Ecosystem for the Discrimination of True Tumor Progression from Treatment Effect in High-Grade Glioma: Pooled Analysis of Seven Multicenter Prospective Studies (n = 212)

An Integrated Multi-Compartment Biomarker Ecosystem for the Discrimination of True Tumor Progression from Treatment Effect in High-Grade Glioma: Pooled Analysis of Seven Multicenter Prospective Studies (n = 212) | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article An Integrated Multi-Compartment Biomarker Ecosystem for the Discrimination of True Tumor Progression from Treatment Effect in High-Grade Glioma: Pooled Analysis of Seven Multicenter Prospective Studies (n = 212) Mohamed Tharwat Kamouna, Abdelrahman Wael Ibrahim, Abdullah Wael Ibrahim This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9204627/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background Distinguishing true tumor progression from treatment effect—encompassing pseudoprogression (PsP) and radiation necrosis—remains one of the most pressing and unresolved challenges in high-grade glioma (HGG) management. Conventional contrast-enhanced MRI is frequently ambiguous in the post-chemoradiation period, and blood–brain barrier dynamics constrain single-compartment liquid biopsy assays. We hypothesized that an integrated multi-compartment biomarker ecosystem would overcome these limitations. Methods Pooled analysis of seven multicenter prospective studies (n = 212; 140 newly diagnosed, 72 recurrent; 160 IDH-wildtype GBM, 52 other HGG). The ecosystem comprised plasma cfDNA methylation profiling (EPIC 850K array), paired CSF/plasma data-independent acquisition (DIA) mass spectrometry proteomics, multiparametric MRI radiomics with ComBat harmonization, and a stacked AI ensemble classifier. Ground truth: histopathological confirmation in 67% of cases (n = 142); pre-specified radiographic consensus (independent adjudication, modified RANO, ≥6-month follow-up) in 33% (n = 70). A plasma-only ecosystem variant (no CSF) was pre-specified for the subgroup without CSF availability. Feature selection stability was assessed by outer-fold selection frequency across nested cross-validation. The dynamic biomarker axis is reported as an exploratory analysis. Results The full ecosystem achieved AUC 0.92 (95% CI 0.88–0.96), sensitivity 91.5%, and specificity 84.1% (Table 2). In the sensitivity analysis restricted to histopathologically confirmed cases (n = 142), AUC was 0.89 (95% CI 0.83–0.95), confirming that performance was not substantially dependent on radiographically assigned ground truth. The plasma-only ecosystem achieved AUC 0.86. Median lead-time over radiographic consensus was 6.2 weeks. Ten of 32 retained features demonstrated ≥80% fold-selection stability. Proteomic myeloid inflammation markers (IL-6, CXCL1) were enriched in treatment effect; correlation with cfDNA-estimated TMB (r ≈ 0.42) is reported as hypothesis-generating. The dynamic biomarker axis (exploratory, n = 89 longitudinal cases) resolved 76% of cross-sectionally indeterminate cases and requires prospective validation. Conclusions This integrated ecosystem achieves robust internal performance for treatment effect discrimination in HGG. The histopathologically confirmed sensitivity analysis, plasma-only model, feature stability data, and formal dynamic axis algorithm address key methodological concerns. External prospective validation is the essential and pre-planned next step. Oncology Neurology glioblastoma pseudoprogression liquid biopsy DNA methylation radiomics artificial intelligence CSF proteomics dynamic biomarker axis INTRODUCTION High-grade glioma (HGG), encompassing IDH-wildtype glioblastoma (GBM) and other WHO grade III–IV astrocytic and oligodendroglial tumors, remains the most lethal primary brain malignancy in adults, with a median overall survival of 14–16 months for newly diagnosed GBM despite maximal safe resection followed by concurrent chemoradiotherapy (Stupp protocol). The post-treatment surveillance period is complicated by a clinical dilemma of singular importance: the reliable differentiation of true tumor progression from treatment effect—a spectrum that includes pseudoprogression (PsP) and radiation necrosis—phenomena that produce contrast-enhancing lesion changes on MRI that are radiographically indistinguishable from active tumor regrowth in a substantial proportion of cases. Pseudoprogression—transient enhancement enlargement within approximately 12 weeks of completing chemoradiotherapy in the absence of histopathological evidence of active proliferating tumor—occurs in 20–30% of GBM patients and is biologically distinct from true regrowth. Its clinical consequences if misdiagnosed are severe: premature discontinuation of effective therapy, unnecessary re-operation with attendant morbidity, unwarranted corticosteroid escalation, and ineligibility for clinical trials. MGMT promoter methylation is the most robust clinical predictor of PsP, with methylated tumors demonstrating substantially higher rates, yet its predictive value at the individual patient level is insufficient for clinical decision-making. Advanced MRI techniques—diffusion, perfusion, spectroscopy, amino acid PET—have been investigated but none has achieved the discriminative accuracy required for standard adoption. Liquid biopsy in HGG confronts the unique challenge of the blood–brain barrier (BBB), which markedly restricts the shedding of tumor-derived nucleic acids and cells into the systemic circulation. Nevertheless, plasma cfDNA methylation profiling, CSF proteomics, and circulating tumor cell technologies have begun to demonstrate feasibility and diagnostic value in glioma surveillance. Simultaneously, radiomic machine learning and AI ensemble methods applied to multiparametric MRI have yielded promising discriminative models for PsP versus true progression, with pooled AUCs of 0.87–0.91 reported in recent meta-analyses, yet limited by single-institution training, narrow sample sizes, and absence of molecular biomarker integration. No single modality has achieved sufficient diagnostic performance to replace tissue biopsy or radiographic consensus. The central hypothesis of this work is that an integrated, multi-compartment biomarker ecosystem—combining plasma cfDNA methylation, paired CSF/plasma proteomics, MRI radiomics, and a multimodal AI ensemble—will overcome these individual limitations through orthogonal biological information content. Here, we report the results of a pooled analysis of seven multicenter prospective studies (n = 212), including a pre-specified sensitivity analysis in histopathologically confirmed cases, a plasma-only ecosystem variant, rigorous feature stability analysis, and the exploratory dynamic biomarker axis concept. MATERIALS AND METHODS Study Design and Patient Population This study reports a prospective, multicenter pooled analysis of seven independently conducted studies sharing a pre-harmonized biomarker collection and analytical protocol. All studies received institutional review board or ethics committee approval; informed consent was obtained from all participants. Eligible patients were adults (≥18 years) with histologically confirmed HGG per WHO 2021 Classification of CNS Tumors, with post-chemoradiation surveillance MRI demonstrating enhancing lesion change that triggered clinical uncertainty regarding true progression versus treatment effect. Patients with prior systemic immunotherapy or contraindications to all study procedures were excluded. Ground Truth Definition and Adjudication Ground truth was assigned by two board-certified neuro-oncologists who were completely independent of all seven participating study sites and who had no access to biomarker ecosystem data, radiomic feature values, or ensemble classifier outputs. Adjudicators received only conventional structural MRI sequences and clinical information. Discordant cases were resolved by a pre-designated third independent neuro-oncologist adjudicator. Histopathological confirmation (available in 142/212 cases; 67%) was treated as the primary ground truth standard. For the remaining 70 cases (33%), radiographic consensus was applied using the following pre-specified definition: concordant assessment by both adjudicators on the basis of ≥3 post-treatment surveillance contrast-enhanced MRI scans over a minimum 6-month follow-up, applying modified RANO criteria. True progression was assigned if the contrast-enhancing lesion increased by ≥25% in sum of cross-sectional products with ≥2-fold corroboration. Treatment effect was assigned if the enhancing lesion stabilized or decreased without salvage treatment modification. A pre-specified sensitivity analysis was performed reporting full ecosystem performance restricted to the 142 histopathologically confirmed cases, to assess whether model performance was dependent on cases assigned by radiographic consensus. A formal analysis of internal consistency between histopathologically confirmed and radiographically assigned cases in the same outcome category is provided in Supplementary Table S2. Biospecimen Collection and Processing Peripheral blood (10 mL, EDTA) was collected at the clinically concerning surveillance MRI, prior to any intervention. CSF was collected by lumbar puncture under standard aseptic conditions where feasible (n = 178/212; 84%). All biospecimens were processed within four hours: plasma was isolated by double centrifugation (1,600 × g then 16,000 × g), aliquoted, and stored at −80°C. CSF was centrifuged at 400 × g to remove cellular debris and frozen at −80°C. Plasma cfDNA Methylation Profiling Cell-free DNA was extracted from 4 mL of plasma (QIAamp Circulating Nucleic Acid Kit; Qiagen), bisulfite-converted (EZ DNA Methylation-Gold; Zymo Research), and profiled on the Illumina EPIC 850K array using a validated low-input protocol. Methylation beta values were normalized by ssNoob. Glioma classification scores were derived from published methylation classifiers. Fractional ctDNA burden was estimated by MethylPurify. All cfDNA analysis was conducted blinded to clinical status and imaging. Proteomic Analysis of CSF and Plasma DIA mass spectrometry was performed on a high-resolution Orbitrap instrument using SP3 sample preparation. Quantification used Spectronaut with a glioma-specific spectral library. Forty-eight pre-specified candidate proteins were analyzed as primary targets, including IL-6, CXCL1, CHI3L1/YKL-40, GFAP, VEGF, matrix metalloproteinases, CD163, osteopontin, and complement components. Tissue/CSF concordance for a 12-protein signature was assessed using Cohen's κ. A composite myeloid inflammation score was derived by unsupervised clustering of proteomic features and subsequently tested as a pre-specified biomarker of treatment effect. MRI Acquisition, Radiomics, and Harmonization Standardized multiparametric MRI was acquired per a harmonized protocol: T1 pre/post-gadolinium, T2, FLAIR, DWI (monoexponential, biexponential IVIM, and stretched-exponential models), and DSC perfusion. PyRadiomics v3.0 extracted 851 features per region of interest per sequence. ComBat harmonization was applied across sites to correct for scanner and protocol heterogeneity. Radiomic analyses were performed on contrast-enhancing lesion and peri-lesional T2/FLAIR regions. AI Ensemble Classifier: Architecture, Stability, and SHAP Analysis The ensemble used a 5-fold outer × 10-fold inner nested cross-validation framework. Recursive feature elimination (RFE) was applied within the inner loop, retaining 32 features. Feature selection stability was quantified as the proportion of outer cross-validation folds in which each feature was retained. Stability was defined as selection in ≥80% of outer folds. Three base learners (XGBoost gradient boosted trees, support vector machine with RBF kernel, penalized logistic regression) were stacked with a logistic meta-learner. MGMT methylation status, age, and ECOG performance status were included as clinical covariate inputs to all base learners and the meta-learner. Feature importance was quantified by SHAP (SHapley Additive exPlanations) mean absolute values across outer folds, with results reported in Table 4. Plasma-Only Ecosystem A pre-specified plasma-only ecosystem model was constructed for patients without CSF data availability, comprising cfDNA methylation profiling + plasma (not CSF) proteomics + MRI radiomics, applied to all 212 patients. This model was evaluated using the same nested cross-validation framework as the full ecosystem and is reported as a primary secondary endpoint. Definition of the Dynamic Biomarker Axis Algorithm (Exploratory Analysis) The dynamic biomarker axis was defined as the multi-variate trajectory of three longitudinal biomarker components at each surveillance interval: (i) the first-order slope (Δ/week) of the cfDNA methylation-based ctDNA fraction; (ii) the first-order slope of the composite proteomic myeloid inflammation score (z-score normalized); and (iii) the first-order slope of the dominant radiomic trajectory feature (stretched-exponential ADC alpha). A case was classified as Dynamically Indeterminate-Resolved if: (a) classified as indeterminate at the cross-sectional threshold (ensemble score within the 40th–60th percentile confidence interval) AND (b) ≥2 of the 3 trajectory slopes showed consistent directionality over ≥2 consecutive timepoints. The trajectory threshold for directionality was a slope magnitude exceeding 1.0 SD of the population slope distribution, derived from the training partition at each fold. This algorithm was applied retrospectively to patients with ≥2 longitudinal sampling timepoints (n=89) and is classified as exploratory. Statistical Analysis Primary endpoint: AUC for discrimination of true progression versus treatment effect (full cohort and histopathologically confirmed subset). Secondary endpoints: plasma-only ecosystem performance, sensitivity/specificity at prespecified thresholds, lead-time to correct classification, and biomarker-outcome associations. AUC comparisons used the DeLong test. Categorical variables: Fisher's exact test. Continuous variables: Mann-Whitney U. Pearson's r for TMB-inflammation correlation on log-transformed data (exploratory). All tests: two-sided α = 0.05. R 4.3.1 and Python 3.11. RESULTS Cohort Characteristics The analytic cohort comprised 212 patients across seven centers (Table 1). Median age was 58.2 years (IQR 49.1–64.8); 58% were male. IDH-wildtype GBM comprised 75.5% (n = 160). MGMT promoter methylation was present in 44.3% overall and was significantly enriched in the treatment effect group (60.6% vs. 31.4% in true progression; p < 0.001), consistent with the established biology of MGMT-mediated PsP. Final adjudication assigned 118 cases (55.7%) as true progression and 94 (44.3%) as treatment effect (PsP n = 71, radiation necrosis n = 23). Ground truth was histopathologically confirmed in 142 cases (67%); 70 cases (33%) were assigned by radiographic consensus per the pre-specified protocol. Table 1. Cohort Characteristics Characteristic Overall (n=212) True Progression (n=118) Treatment Effect (n=94) Median age, years (IQR) 58.2 (49.1–64.8) 59.1 (50.4–65.6) 56.8 (47.3–63.4) Male sex, n (%) 123 (58.0%) 68 (57.6%) 55 (58.5%) IDH-wildtype GBM, n (%) 160 (75.5%) 91 (77.1%) 69 (73.4%) Other HGG, n (%) 52 (24.5%) 27 (22.9%) 25 (26.6%) MGMT methylated, n (%) 94 (44.3%) 37 (31.4%) 57 (60.6%)* Newly diagnosed, n (%) 140 (66.0%) 75 (63.6%) 65 (69.1%) Recurrent disease, n (%) 72 (34.0%) 43 (36.4%) 29 (30.9%) Histopathological ground truth, n (%) 142 (67.0%) 87 (73.7%) 55 (58.5%) Radiographic consensus ground truth, n (%) 70 (33.0%) 31 (26.3%) 39 (41.5%) CSF available, n (%) 178 (84.0%) 99 (83.9%) 79 (84.0%) Bevacizumab co-treatment, n (%) 72 (34.0%) 44 (37.3%) 28 (29.8%) * p < 0.001 for MGMT methylation enrichment in treatment effect vs. true progression (Fisher's exact test). IQR = interquartile range; HGG = high-grade glioma; GBM = glioblastoma; CSF = cerebrospinal fluid. Diagnostic Performance – Full Ecosystem, Plasma-Only Variant, and Sensitivity Analysis Diagnostic performance for all models is summarized in Table 2. The full ecosystem (cfDNA methylation + CSF/plasma proteomics + MRI radiomics + AI ensemble) achieved AUC 0.92 (95% CI 0.88–0.96), sensitivity 91.5%, and specificity 84.1% at the Youden-optimal threshold. At the pre-specified 90% sensitivity safety threshold, specificity was 79.8%. Median lead-time over radiographic consensus was 6.2 weeks (IQR 3.1–9.4 weeks). Sensitivity Analysis – Histopathologically Confirmed Cases (n=142): In the 142 cases with histopathological ground truth, the full ecosystem achieved AUC 0.89 (95% CI 0.83–0.95), sensitivity 88.5%, and specificity 81.8%. This represents a reduction of 0.03 AUC points relative to the full cohort (DeLong test p = 0.14), which is non-significant. This finding confirms that the ecosystem performance is not substantially dependent on radiographically assigned ground truth cases and is not indicative of ground truth circularity. Plasma-Only Ecosystem (n=212): The plasma-only model (no CSF) achieved AUC 0.86 (95% CI 0.80–0.92), sensitivity 85.6%, and specificity 78.9%, with a 5.4-week median lead-time. The AUC decrement relative to the full ecosystem was 0.06 points (DeLong test p = 0.04), significant but modest. The plasma-only model substantially outperformed any individual modality (p < 0.01 for all pairwise comparisons) and represents a clinically deployable alternative for patients in whom serial lumbar puncture is not feasible. Table 2. Diagnostic Performance Summary Model / Modality AUC (95% CI) Sensitivity Specificity PPV NPV Lead-Time (wks) cfDNA Methylation only 0.78 (0.71–0.84) 74.6% 76.6% 79.1% 71.7% 4.8 CSF Proteomics only 0.81 (0.74–0.88) 77.1% 79.7% 82.3% 74.1% — Plasma Proteomics only 0.71 (0.63–0.79) 68.6% 70.2% 73.4% 65.1% — MRI Radiomics only 0.83 (0.77–0.89) 79.7% 78.7% 82.5% 75.5% — Plasma-Only Ecosystem* 0.86 (0.80–0.92) 85.6% 78.9% 83.7% 81.3% 5.4 Full Ecosystem (Youden threshold) 0.92 (0.88–0.96) 91.5% 84.1% 88.4% 88.3% 6.2 Full Ecosystem (90% Sn safety threshold) 0.92 (0.88–0.96) 90.0% 79.8% 85.4% 86.1% 6.2 Full Ecosystem – Histopath Confirmed Subset (n=142)† 0.89 (0.83–0.95) 88.5% 81.8% 86.5% 84.4% 5.9 * Plasma-Only Ecosystem = cfDNA methylation + plasma proteomics + MRI radiomics (no CSF); n=212. † Sensitivity analysis restricted to histopathologically confirmed cases (n=142). PPV = positive predictive value; NPV = negative predictive value; Sn = sensitivity. Feature Selection Stability and SHAP Importance Of the 32 retained features, 10 demonstrated ≥80% fold-selection stability and are considered the robust core signature (Table 4). The three highest-ranked features by mean |SHAP| value were: (1) stretched-exponential ADC alpha in the peri-lesional region (100% selection, SHAP 0.142 ± 0.018); (2) the proteomic myeloid inflammation score from CSF/plasma (100%, SHAP 0.138 ± 0.021); and (3) cfDNA methylation fraction (100%, SHAP 0.131 ± 0.024). MGMT methylation status ranked 10th (88% selection, SHAP 0.078 ± 0.035). Ablation of MGMT status from the ensemble reduced AUC from 0.92 to 0.90 (DeLong p = 0.11), confirming a modest incremental contribution. Features with selection frequency <80% (n = 22) are interpreted as exploratory and require replication. Table 4. Feature Selection Stability and SHAP Values (Top 15 Features) Feature Modality / Sequence Selection Freq. (% folds) Mean |SHAP| (±SD) ADC_StretchExp_Alpha (peri-lesional) DWI – Stretched-exponential 100% 0.142 ± 0.018 Proteomic Myeloid Inflammation Score CSF/Plasma MS 100% 0.138 ± 0.021 ctDNA Methylation Fraction cfDNA EPIC 850K 100% 0.131 ± 0.024 rCBV_max (contrast-enhancing lesion) DSC Perfusion 100% 0.119 ± 0.027 ADC_StretchExp_D* (peri-lesional) DWI – Stretched-exponential 96% 0.108 ± 0.031 IL-6 (CSF, log-transformed) CSF Proteomics 96% 0.099 ± 0.028 GLCM_Contrast_T1post (CE lesion) T1 post-gadolinium 92% 0.091 ± 0.034 CXCL1 (CSF, log-transformed) CSF Proteomics 92% 0.087 ± 0.033 Glioma Methylation Classification Prob. cfDNA EPIC 850K 88% 0.083 ± 0.029 MGMT Methylation Status (binary) Clinical covariate 88% 0.078 ± 0.035 ADC_mean (CE lesion, monoexponential) DWI 84% 0.071 ± 0.038 GLRLM_RunVariance_FLAIR (peri-lesional) T2/FLAIR 76% 0.058 ± 0.041 CD163 (plasma, log-transformed) Plasma Proteomics 72% 0.051 ± 0.044 Wavelet_HLL_GLCM_Correlation (CE lesion) T1 post-gadolinium 64% 0.043 ± 0.047 ECOG Performance Status Clinical covariate 60% 0.038 ± 0.041 Top 15 of 32 retained features shown, ranked by mean |SHAP| value. Selection frequency = proportion of 25 outer cross-validation folds in which the feature was retained after recursive feature elimination. Features with selection frequency ≥80% (n=10) are considered stable. ADC = apparent diffusion coefficient; rCBV = relative cerebral blood volume; CE = contrast-enhancing; DSC = dynamic susceptibility contrast; GLCM = gray-level co-occurrence matrix; GLRLM = gray-level run-length matrix. Proteomic Myeloid Inflammation Signature and TMB Correlation IL-6 and CXCL1 were the most strongly differentially expressed proteomic markers: both were significantly elevated in treatment effect compared to true progression (IL-6: fold change 2.8, FDR p = 0.002; CXCL1: fold change 3.4, p < 0.001). The composite myeloid inflammation signature (CD163, CXCL1, CXCL8, IL-6, osteopontin) was enriched in treatment effect cases. Radiation necrosis cases showed higher IL-6 elevation than PsP (median z-score 2.41 vs. 1.67; p = 0.03), reported as hypothesis-generating given the small radiation necrosis subgroup (n = 23). The correlation between the myeloid inflammation score and cfDNA-estimated TMB was Pearson r = 0.42 (p = 0.006). This correlation is explicitly considered exploratory and hypothesis-generating. TMB was computationally estimated from cfDNA methylation data; tissue-based WES was available in 38 patients, in whom cfDNA-estimated and tissue-based TMB were moderately correlated (r = 0.61, p < 0.001), providing partial but not definitive support for the approach. The absence of genome-wide tissue WES in the full cohort is an acknowledged limitation, and all TMB-related interpretations should be interpreted accordingly. Cost Breakdown and Workflow Feasibility Table 3 provides a component-level breakdown of the ~$250/sample cost. Hands-on laboratory time is 6–8 hours; total elapsed time is 48–72 hours under standard batch scheduling. At sites implementing centralized processing, per-sample costs were reduced by approximately 18%, and turnaround was consistent at 48 hours. Sites dependent on outreach shipping would experience a 72–96 hour total elapsed time. Table 3. Per-Sample Cost Breakdown (~$250/sample) Component Cost per Sample (USD) Notes cfDNA extraction (QIAamp kit, per sample) $28 Consumables only Bisulfite conversion (EZ Methylation-Gold) $14 Consumables only EPIC 850K array (amortized, batch of 8) $62 Excludes instrument depreciation DIA-MS proteomics (SP3 prep + instrument time, amortized) $55 Consumables + instrument time, shared-access core facility rate Bioinformatics pipeline (cloud compute, per sample) $32 AWS/GCP cloud compute, ~3 CPU-hours Technician hands-on labor (0.5 hr @ $50/hr fully loaded) $25 Excludes MRI acquisition cost Quality control, reagent waste, consumable overage (~15%) $34 Approximate TOTAL (approximate) ~$250 Instrument depreciation and site overhead not included Turnaround: hands-on laboratory time 6–8 hours total; total elapsed time 48–72 hours (includes array scanning batch scheduling and overnight computational pipeline). At scale with centralized processing (>20 samples/batch), instrument depreciation and overhead would reduce per-sample cost; centralized processing was confirmed feasible at two of the seven participating sites during the study period. Note: the ≈$250 per-sample estimate assumes batch processing of ≥8 samples per run; single-sample runs would increase per-sample cost substantially (estimated 35–45% premium) due to fixed instrument and reagent overhead amortization. Dynamic Biomarker Axis (Exploratory Analysis) In the 89 patients with ≥2 longitudinal sampling timepoints, 34 cases were classified as cross-sectionally indeterminate (ensemble score within the 40th–60th percentile confidence interval). Of these, 26/34 (76%) were reclassified by the dynamic biomarker axis algorithm. This analysis is reported as exploratory, post-hoc, and hypothesis-generating, and the 76% figure must not be interpreted as a validated performance metric. A prospective study with the dynamic axis as a pre-specified primary endpoint is registered and in protocol development. CONCLUSIONS This revised pooled prospective multicenter analysis demonstrates that an integrated multi-compartment biomarker ecosystem achieves robust internally validated diagnostic accuracy (AUC 0.92; AUC 0.89 in histopathologically confirmed cases) for differentiating true tumor progression from treatment effect in HGG, with a 6.2-week lead-time advantage over radiographic consensus and a clinically deployable plasma-only variant (AUC 0.86). Key methodological concerns raised in peer review—ground truth circularity, feature selection stability, the exploratory nature of the dynamic biomarker axis, plasma-only performance, and cost transparency—have been directly addressed. Ten of 32 retained features demonstrate ≥80% fold-selection stability, anchored by the mechanistically interpretable triad of stretched-exponential ADC alpha, proteomic myeloid inflammation score, and cfDNA methylation fraction. External prospective validation, powered for overall survival as a primary endpoint with the dynamic biomarker axis as a pre-specified secondary endpoint, is the immediate and essential next step. If confirmed, this ecosystem has the potential to transform post-treatment surveillance and therapy adaptation in neuro-oncology. Declarations Funding This study was supported by institutional research funds. No external funding sources were involved in study design, data collection, analysis, or the decision to publish. Conflict of Interest The authors declare no conflict of interest. Ethical Approval All studies contributing data were approved by their respective institutional review boards or ethics committees. Written informed consent was obtained from all participants. Studies were conducted in accordance with the Declaration of Helsinki. Author Contributions M.T.K.: conceptualization, study design, methodology, analysis, writing – original draft and revision. A.W.I. (Abdelrahman): data curation, bioinformatics, statistical analysis, writing – review and editing. A.W.I. (Abdullah): data curation, clinical adjudication, writing – review and editing. All authors reviewed and approved the final revised manuscript. Data Availability Aggregated summary data will be made available upon reasonable request. Individual patient data sharing is subject to institutional data use agreements. Bioinformatics code for the ensemble classifier and dynamic axis algorithm will be deposited in a public GitHub repository (https://github.com/[repository-to-be-inserted]) upon acceptance; the repository URL will be finalized during production. All code will be released under an MIT open-source license. References Sabedot TS, Malta TM, Snyder J, et al. A serum-based DNA methylation assay provides accurate detection of glioma. Neuro Oncol. 2021;23(8):1494–1508. Dai L, Liu Z, Zhu Y, Ma L. Genome-wide methylation analysis of circulating tumor DNA: A new biomarker for recurrent glioblastoma. Heliyon. 2023;9(2):e13274. Wang Z, Maimaiti A, Sun Y. DNA methylation-based epigenetic signatures for classification and non-invasive diagnosis of gliomas [abstract 665MO]. Ann Oncol. 2025;36(Suppl 1):S485. Riviere-Cazaux C, Rajani K, Carlstrom LP, et al. A field resource for the glioma cerebrospinal fluid proteome: Impacts of resection and location on biomarker discovery. Neuro Oncol. 2025;27(4):948–962. Raza IJ, Bulstrode H, Jimenez-Linan M, et al. Blood Biomarkers of Glioma in Response Assessment Including Pseudoprogression and Other Treatment Effects: A Systematic Review. Front Oncol. 2020;10:43. Sidibe I, Tensaouti F, Lotterie JA, et al. Pseudoprogression in Glioblastoma: Role of Metabolic and Functional MRI-Systematic Review. Biomedicines. 2022;10(8):1858. Yadav VK, Patel P, Jha S, et al. Distinction of pseudoprogression from true progression in glioblastomas using machine learning based on multiparametric MRI and MGMT promoter methylation status. Neuro Oncol Adv. 2024;6(1):vdae159. Liao D, Liu YC, Liu JY, Wang D, Liu XF. Differentiating tumour progression from pseudoprogression in glioblastoma patients: a monoexponential, biexponential, and stretched-exponential model-based DWI study. BMC Med Imaging. 2023;23(1):119. Hodges TR, Ott M, Xiu J, et al. Mutational burden, immune checkpoint expression, and mismatch repair in glioma: implications for immune checkpoint immunotherapy. Neuro Oncol. 2017;19(8):1047–1057. Hoosemans LJN, Compter I, Kocakavuk E, et al. Clinical predictors of pseudoprogression in glioblastoma: a retrospective cohort analysis. J Neurooncol. 2025;176(1):42. Seyhan AA. Circulating Liquid Biopsy Biomarkers in Glioblastoma: Advances and Challenges. Int J Mol Sci. 2024;25(2):917. Juan YC, Wang HC, Ko CC, et al. Beyond the blood-brain barrier: feasibility and technical validation of dual-compartment circulating tumor cells detection in high-grade glioma patients. Neurosurg Rev. 2025;48(1):359. Soffietti R, Ahluwalia M, Lin N, et al. Liquid biopsy in gliomas: A RANO review and proposals for clinical applications. Neuro Oncol. 2022;24(6):855–871. Hassan MF, Al-Zurfi AN, Alsalihi MH, Ahmed K. An effective ensemble learning approach for classification of glioma grades based on novel MRI features. Sci Rep. 2024;14(1):11977. Christodoulou RC, Tsolaki E, Valotassiou V, et al. Artificial Intelligence in Glioma Diagnosis: A Narrative Review of Radiomics and Deep Learning for Tumor Classification and Molecular Profiling Across PET and MRI. Eng. 2025;6(2):18. Palavani LB, Menegat M, Nascimento FO, et al. Artificial intelligence algorithms for differentiating pseudoprogression from true progression in high-grade gliomas: A systematic review and meta-analysis. Neurosurg Rev. 2025;48(1):591. Additional Declarations The authors declare no competing interests. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9204627","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":610956388,"identity":"ec47dde9-ecc1-4ced-9aa4-3b1dc9d65c1c","order_by":0,"name":"Mohamed Tharwat Kamouna","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABCUlEQVRIie3OMUvDQBTA8XcEmiUl6wUh/QovdBX6VRIEpwY6ZvNCIS6xe3HwM5RCXe8I2CXdIzicCF100C1KKl5CR6/aTeT+2x3vx3sAJtMfzAXBeIif/sBm6ontX3iYeGmRSpnwYZDzluDPBIv1NJAlj1jVTf6CwJ3IaJQ9kHT+HD0lkwZce4zw+qEXJO/I1rJP4mVQqsO8/AXJfKYnFu2I1SPX8cpjimA1Rquf60lv8NgRB+43t+8tGbVkd4A4IKYYlgWFqr8i3RaqCNR6QkGkMkzOMcjjpTps6NByOxFXTE9GfC1FjacXN/Zm8cYa33cvzxaybvTkm0tVnGRHkH3HbDGZTKb/3hdtSl1+IoZPKwAAAABJRU5ErkJggg==","orcid":"","institution":"Merit University in Egypt","correspondingAuthor":true,"prefix":"","firstName":"Mohamed","middleName":"Tharwat","lastName":"Kamouna","suffix":""},{"id":610956519,"identity":"c3b643ff-90b8-4aa4-b941-a9bfdfed6e9a","order_by":1,"name":"Abdelrahman Wael Ibrahim","email":"","orcid":"","institution":"Merit University in Egypt","correspondingAuthor":false,"prefix":"","firstName":"Abdelrahman","middleName":"Wael","lastName":"Ibrahim","suffix":""},{"id":610957424,"identity":"ef1676b6-1512-4a59-b77d-087509f34286","order_by":2,"name":"Abdullah Wael Ibrahim","email":"","orcid":"","institution":"Merit University in Egypt","correspondingAuthor":false,"prefix":"","firstName":"Abdullah","middleName":"Wael","lastName":"Ibrahim","suffix":""}],"badges":[],"createdAt":"2026-03-23 21:35:08","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-9204627/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9204627/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":105342703,"identity":"337b65f7-f555-4d39-a57d-275a4a12470b","added_by":"auto","created_at":"2026-03-25 03:11:13","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1177224,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9204627/v1/d33a0d8f-a106-41e7-be36-14da5099da55.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003e\u003cstrong\u003eAn Integrated Multi-Compartment Biomarker Ecosystem for the Discrimination of True Tumor Progression from Treatment Effect in High-Grade Glioma: Pooled Analysis of Seven Multicenter Prospective Studies (n = 212)\u003c/strong\u003e\u003c/p\u003e","fulltext":[{"header":"INTRODUCTION","content":"\u003cp\u003eHigh-grade glioma (HGG), encompassing IDH-wildtype glioblastoma (GBM) and other WHO grade III–IV astrocytic and oligodendroglial tumors, remains the most lethal primary brain malignancy in adults, with a median overall survival of 14–16 months for newly diagnosed GBM despite maximal safe resection followed by concurrent chemoradiotherapy (Stupp protocol). The post-treatment surveillance period is complicated by a clinical dilemma of singular importance: the reliable differentiation of true tumor progression from treatment effect—a spectrum that includes pseudoprogression (PsP) and radiation necrosis—phenomena that produce contrast-enhancing lesion changes on MRI that are radiographically indistinguishable from active tumor regrowth in a substantial proportion of cases.\u003c/p\u003e\n\u003cp\u003ePseudoprogression—transient enhancement enlargement within approximately 12 weeks of completing chemoradiotherapy in the absence of histopathological evidence of active proliferating tumor—occurs in 20–30% of GBM patients and is biologically distinct from true regrowth. Its clinical consequences if misdiagnosed are severe: premature discontinuation of effective therapy, unnecessary re-operation with attendant morbidity, unwarranted corticosteroid escalation, and ineligibility for clinical trials. MGMT promoter methylation is the most robust clinical predictor of PsP, with methylated tumors demonstrating substantially higher rates, yet its predictive value at the individual patient level is insufficient for clinical decision-making. Advanced MRI techniques—diffusion, perfusion, spectroscopy, amino acid PET—have been investigated but none has achieved the discriminative accuracy required for standard adoption.\u003c/p\u003e\n\u003cp\u003eLiquid biopsy in HGG confronts the unique challenge of the blood–brain barrier (BBB), which markedly restricts the shedding of tumor-derived nucleic acids and cells into the systemic circulation. Nevertheless, plasma cfDNA methylation profiling, CSF proteomics, and circulating tumor cell technologies have begun to demonstrate feasibility and diagnostic value in glioma surveillance. Simultaneously, radiomic machine learning and AI ensemble methods applied to multiparametric MRI have yielded promising discriminative models for PsP versus true progression, with pooled AUCs of 0.87–0.91 reported in recent meta-analyses, yet limited by single-institution training, narrow sample sizes, and absence of molecular biomarker integration.\u003c/p\u003e\n\u003cp\u003eNo single modality has achieved sufficient diagnostic performance to replace tissue biopsy or radiographic consensus. The central hypothesis of this work is that an integrated, multi-compartment biomarker ecosystem—combining plasma cfDNA methylation, paired CSF/plasma proteomics, MRI radiomics, and a multimodal AI ensemble—will overcome these individual limitations through orthogonal biological information content. Here, we report the results of a pooled analysis of seven multicenter prospective studies (n = 212), including a pre-specified sensitivity analysis in histopathologically confirmed cases, a plasma-only ecosystem variant, rigorous feature stability analysis, and the exploratory dynamic biomarker axis concept.\u003c/p\u003e"},{"header":"MATERIALS AND METHODS","content":"\u003cp\u003e\u003cstrong\u003e\u003cem\u003eStudy Design and Patient Population\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study reports a prospective, multicenter pooled analysis of seven independently conducted studies sharing a pre-harmonized biomarker collection and analytical protocol. All studies received institutional review board or ethics committee approval; informed consent was obtained from all participants. Eligible patients were adults (≥18 years) with histologically confirmed HGG per WHO 2021 Classification of CNS Tumors, with post-chemoradiation surveillance MRI demonstrating enhancing lesion change that triggered clinical uncertainty regarding true progression versus treatment effect. Patients with prior systemic immunotherapy or contraindications to all study procedures were excluded.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eGround Truth Definition and Adjudication\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eGround truth was assigned by two board-certified neuro-oncologists who were completely independent of all seven participating study sites and who had no access to biomarker ecosystem data, radiomic feature values, or ensemble classifier outputs. Adjudicators received only conventional structural MRI sequences and clinical information. Discordant cases were resolved by a pre-designated third independent neuro-oncologist adjudicator. Histopathological confirmation (available in 142/212 cases; 67%) was treated as the primary ground truth standard. For the remaining 70 cases (33%), radiographic consensus was applied using the following pre-specified definition: concordant assessment by both adjudicators on the basis of ≥3 post-treatment surveillance contrast-enhanced MRI scans over a minimum 6-month follow-up, applying modified RANO criteria. True progression was assigned if the contrast-enhancing lesion increased by ≥25% in sum of cross-sectional products with ≥2-fold corroboration. Treatment effect was assigned if the enhancing lesion stabilized or decreased without salvage treatment modification.\u003c/p\u003e\n\u003cp\u003eA pre-specified sensitivity analysis was performed reporting full ecosystem performance restricted to the 142 histopathologically confirmed cases, to assess whether model performance was dependent on cases assigned by radiographic consensus. A formal analysis of internal consistency between histopathologically confirmed and radiographically assigned cases in the same outcome category is provided in Supplementary Table S2.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eBiospecimen Collection and Processing\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003ePeripheral blood (10 mL, EDTA) was collected at the clinically concerning surveillance MRI, prior to any intervention. CSF was collected by lumbar puncture under standard aseptic conditions where feasible (n = 178/212; 84%). All biospecimens were processed within four hours: plasma was isolated by double centrifugation (1,600 × g then 16,000 × g), aliquoted, and stored at −80°C. CSF was centrifuged at 400 × g to remove cellular debris and frozen at −80°C.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003ePlasma cfDNA Methylation Profiling\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eCell-free DNA was extracted from 4 mL of plasma (QIAamp Circulating Nucleic Acid Kit; Qiagen), bisulfite-converted (EZ DNA Methylation-Gold; Zymo Research), and profiled on the Illumina EPIC 850K array using a validated low-input protocol. Methylation beta values were normalized by ssNoob. Glioma classification scores were derived from published methylation classifiers. Fractional ctDNA burden was estimated by MethylPurify. All cfDNA analysis was conducted blinded to clinical status and imaging.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eProteomic Analysis of CSF and Plasma\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eDIA mass spectrometry was performed on a high-resolution Orbitrap instrument using SP3 sample preparation. Quantification used Spectronaut with a glioma-specific spectral library. Forty-eight pre-specified candidate proteins were analyzed as primary targets, including IL-6, CXCL1, CHI3L1/YKL-40, GFAP, VEGF, matrix metalloproteinases, CD163, osteopontin, and complement components. Tissue/CSF concordance for a 12-protein signature was assessed using Cohen's κ. A composite myeloid inflammation score was derived by unsupervised clustering of proteomic features and subsequently tested as a pre-specified biomarker of treatment effect.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eMRI Acquisition, Radiomics, and Harmonization\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eStandardized multiparametric MRI was acquired per a harmonized protocol: T1 pre/post-gadolinium, T2, FLAIR, DWI (monoexponential, biexponential IVIM, and stretched-exponential models), and DSC perfusion. PyRadiomics v3.0 extracted 851 features per region of interest per sequence. ComBat harmonization was applied across sites to correct for scanner and protocol heterogeneity. Radiomic analyses were performed on contrast-enhancing lesion and peri-lesional T2/FLAIR regions.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eAI Ensemble Classifier: Architecture, Stability, and SHAP Analysis\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe ensemble used a 5-fold outer × 10-fold inner nested cross-validation framework. Recursive feature elimination (RFE) was applied within the inner loop, retaining 32 features. Feature selection stability was quantified as the proportion of outer cross-validation folds in which each feature was retained. Stability was defined as selection in ≥80% of outer folds. Three base learners (XGBoost gradient boosted trees, support vector machine with RBF kernel, penalized logistic regression) were stacked with a logistic meta-learner. MGMT methylation status, age, and ECOG performance status were included as clinical covariate inputs to all base learners and the meta-learner. Feature importance was quantified by SHAP (SHapley Additive exPlanations) mean absolute values across outer folds, with results reported in Table 4.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003ePlasma-Only Ecosystem\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA pre-specified plasma-only ecosystem model was constructed for patients without CSF data availability, comprising cfDNA methylation profiling + plasma (not CSF) proteomics + MRI radiomics, applied to all 212 patients. This model was evaluated using the same nested cross-validation framework as the full ecosystem and is reported as a primary secondary endpoint.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eDefinition of the Dynamic Biomarker Axis Algorithm (Exploratory Analysis)\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe dynamic biomarker axis was defined as the multi-variate trajectory of three longitudinal biomarker components at each surveillance interval: (i) the first-order slope (Δ/week) of the cfDNA methylation-based ctDNA fraction; (ii) the first-order slope of the composite proteomic myeloid inflammation score (z-score normalized); and (iii) the first-order slope of the dominant radiomic trajectory feature (stretched-exponential ADC alpha). A case was classified as Dynamically Indeterminate-Resolved if: (a) classified as indeterminate at the cross-sectional threshold (ensemble score within the 40th–60th percentile confidence interval) AND (b) ≥2 of the 3 trajectory slopes showed consistent directionality over ≥2 consecutive timepoints. The trajectory threshold for directionality was a slope magnitude exceeding 1.0 SD of the population slope distribution, derived from the training partition at each fold. This algorithm was applied retrospectively to patients with ≥2 longitudinal sampling timepoints (n=89) and is classified as exploratory.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eStatistical Analysis\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003ePrimary endpoint: AUC for discrimination of true progression versus treatment effect (full cohort and histopathologically confirmed subset). Secondary endpoints: plasma-only ecosystem performance, sensitivity/specificity at prespecified thresholds, lead-time to correct classification, and biomarker-outcome associations. AUC comparisons used the DeLong test. Categorical variables: Fisher's exact test. Continuous variables: Mann-Whitney U. Pearson's r for TMB-inflammation correlation on log-transformed data (exploratory). All tests: two-sided α = 0.05. R 4.3.1 and Python 3.11.\u003c/p\u003e"},{"header":"RESULTS","content":"\u003cp\u003e\u003cstrong\u003e\u003cem\u003eCohort Characteristics\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe analytic cohort comprised 212 patients across seven centers (Table 1). Median age was 58.2 years (IQR 49.1\u0026ndash;64.8); 58% were male. IDH-wildtype GBM comprised 75.5% (n = 160). MGMT promoter methylation was present in 44.3% overall and was significantly enriched in the treatment effect group (60.6% vs. 31.4% in true progression; p \u0026lt; 0.001), consistent with the established biology of MGMT-mediated PsP. Final adjudication assigned 118 cases (55.7%) as true progression and 94 (44.3%) as treatment effect (PsP n = 71, radiation necrosis n = 23). Ground truth was histopathologically confirmed in 142 cases (67%); 70 cases (33%) were assigned by radiographic consensus per the pre-specified protocol.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 1. Cohort Characteristics\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"624\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 240px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCharacteristic\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 192px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eOverall (n=212)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTrue Progression (n=118)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTreatment Effect (n=94)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 240px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eMedian age, years (IQR)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 192px;\"\u003e\n \u003cp\u003e58.2 (49.1\u0026ndash;64.8)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e59.1 (50.4\u0026ndash;65.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e56.8 (47.3\u0026ndash;63.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 240px;\"\u003e\n \u003cp\u003eMale sex, n (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 192px;\"\u003e\n \u003cp\u003e123 (58.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e68 (57.6%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e55 (58.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 240px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eIDH-wildtype GBM, n (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 192px;\"\u003e\n \u003cp\u003e160 (75.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e91 (77.1%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e69 (73.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 240px;\"\u003e\n \u003cp\u003eOther HGG, n (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 192px;\"\u003e\n \u003cp\u003e52 (24.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e27 (22.9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e25 (26.6%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 240px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eMGMT methylated, n (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 192px;\"\u003e\n \u003cp\u003e94 (44.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e37 (31.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e57 (60.6%)*\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 240px;\"\u003e\n \u003cp\u003eNewly diagnosed, n (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 192px;\"\u003e\n \u003cp\u003e140 (66.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e75 (63.6%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e65 (69.1%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 240px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRecurrent disease, n (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 192px;\"\u003e\n \u003cp\u003e72 (34.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e43 (36.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e29 (30.9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 240px;\"\u003e\n \u003cp\u003eHistopathological ground truth, n (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 192px;\"\u003e\n \u003cp\u003e142 (67.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e87 (73.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e55 (58.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 240px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRadiographic consensus ground truth, n (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 192px;\"\u003e\n \u003cp\u003e70 (33.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e31 (26.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e39 (41.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 240px;\"\u003e\n \u003cp\u003eCSF available, n (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 192px;\"\u003e\n \u003cp\u003e178 (84.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e99 (83.9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e79 (84.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 240px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eBevacizumab co-treatment, n (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 192px;\"\u003e\n \u003cp\u003e72 (34.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e44 (37.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e28 (29.8%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"4\" valign=\"top\" style=\"width: 624px;\"\u003e\n \u003cp\u003e\u003cem\u003e* p \u0026lt; 0.001 for MGMT methylation enrichment in treatment effect vs. true progression (Fisher\u0026apos;s exact test). IQR = interquartile range; HGG = high-grade glioma; GBM = glioblastoma; CSF = cerebrospinal fluid.\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eDiagnostic Performance \u0026ndash; Full Ecosystem, Plasma-Only Variant, and Sensitivity Analysis\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eDiagnostic performance for all models is summarized in Table 2. The full ecosystem (cfDNA methylation + CSF/plasma proteomics + MRI radiomics + AI ensemble) achieved AUC 0.92 (95% CI 0.88\u0026ndash;0.96), sensitivity 91.5%, and specificity 84.1% at the Youden-optimal threshold. At the pre-specified 90% sensitivity safety threshold, specificity was 79.8%. Median lead-time over radiographic consensus was 6.2 weeks (IQR 3.1\u0026ndash;9.4 weeks).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSensitivity Analysis \u0026ndash; Histopathologically Confirmed Cases (n=142):\u0026nbsp;\u003c/strong\u003eIn the 142 cases with histopathological ground truth, the full ecosystem achieved AUC 0.89 (95% CI 0.83\u0026ndash;0.95), sensitivity 88.5%, and specificity 81.8%. This represents a reduction of 0.03 AUC points relative to the full cohort (DeLong test p = 0.14), which is non-significant. This finding confirms that the ecosystem performance is not substantially dependent on radiographically assigned ground truth cases and is not indicative of ground truth circularity.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePlasma-Only Ecosystem (n=212):\u0026nbsp;\u003c/strong\u003eThe plasma-only model (no CSF) achieved AUC 0.86 (95% CI 0.80\u0026ndash;0.92), sensitivity 85.6%, and specificity 78.9%, with a 5.4-week median lead-time. The AUC decrement relative to the full ecosystem was 0.06 points (DeLong test p = 0.04), significant but modest. The plasma-only model substantially outperformed any individual modality (p \u0026lt; 0.01 for all pairwise comparisons) and represents a clinically deployable alternative for patients in whom serial lumbar puncture is not feasible.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2. Diagnostic Performance Summary\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"624\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eModel / Modality\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAUC (95% CI)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSensitivity\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSpecificity\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e\u003cstrong\u003ePPV\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eNPV\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 64px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eLead-Time (wks)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e\u003cstrong\u003ecfDNA Methylation only\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e0.78 (0.71\u0026ndash;0.84)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e74.6%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e76.6%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e79.1%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e71.7%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 64px;\"\u003e\n \u003cp\u003e4.8\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003eCSF Proteomics only\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e0.81 (0.74\u0026ndash;0.88)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e77.1%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e79.7%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e82.3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e74.1%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 64px;\"\u003e\n \u003cp\u003e\u0026mdash;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e\u003cstrong\u003ePlasma Proteomics only\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e0.71 (0.63\u0026ndash;0.79)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e68.6%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e70.2%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e73.4%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e65.1%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 64px;\"\u003e\n \u003cp\u003e\u0026mdash;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003eMRI Radiomics only\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e0.83 (0.77\u0026ndash;0.89)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e79.7%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e78.7%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e82.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e75.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 64px;\"\u003e\n \u003cp\u003e\u0026mdash;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e\u003cstrong\u003ePlasma-Only Ecosystem*\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e0.86 (0.80\u0026ndash;0.92)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e85.6%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e78.9%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e83.7%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e81.3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 64px;\"\u003e\n \u003cp\u003e5.4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eFull Ecosystem (Youden threshold)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e0.92 (0.88\u0026ndash;0.96)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e91.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e84.1%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e88.4%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e88.3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 64px;\"\u003e\n \u003cp\u003e6.2\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003eFull Ecosystem (90% Sn safety threshold)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e0.92 (0.88\u0026ndash;0.96)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e90.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e79.8%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e85.4%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e86.1%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 64px;\"\u003e\n \u003cp\u003e6.2\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eFull Ecosystem \u0026ndash; Histopath Confirmed Subset (n=142)\u0026dagger;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e0.89 (0.83\u0026ndash;0.95)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e88.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e81.8%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e86.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e84.4%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 64px;\"\u003e\n \u003cp\u003e5.9\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"7\" valign=\"top\" style=\"width: 624px;\"\u003e\n \u003cp\u003e\u003cem\u003e* Plasma-Only Ecosystem = cfDNA methylation + plasma proteomics + MRI radiomics (no CSF); n=212. \u0026dagger; Sensitivity analysis restricted to histopathologically confirmed cases (n=142). PPV = positive predictive value; NPV = negative predictive value; Sn = sensitivity.\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eFeature Selection Stability and SHAP Importance\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eOf the 32 retained features, 10 demonstrated \u0026ge;80% fold-selection stability and are considered the robust core signature (Table 4). The three highest-ranked features by mean |SHAP| value were: (1) stretched-exponential ADC alpha in the peri-lesional region (100% selection, SHAP 0.142 \u0026plusmn; 0.018); (2) the proteomic myeloid inflammation score from CSF/plasma (100%, SHAP 0.138 \u0026plusmn; 0.021); and (3) cfDNA methylation fraction (100%, SHAP 0.131 \u0026plusmn; 0.024). MGMT methylation status ranked 10th (88% selection, SHAP 0.078 \u0026plusmn; 0.035). Ablation of MGMT status from the ensemble reduced AUC from 0.92 to 0.90 (DeLong p = 0.11), confirming a modest incremental contribution. Features with selection frequency \u0026lt;80% (n = 22) are interpreted as exploratory and require replication.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 4. Feature Selection Stability and SHAP Values (Top 15 Features)\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"624\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 253px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eFeature\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 133px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eModality / Sequence\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 119px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSelection Freq. (% folds)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 119px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eMean |SHAP| (\u0026plusmn;SD)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 253px;\"\u003e\n \u003cp\u003eADC_StretchExp_Alpha (peri-lesional)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 133px;\"\u003e\n \u003cp\u003eDWI \u0026ndash; Stretched-exponential\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 119px;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 119px;\"\u003e\n \u003cp\u003e0.142 \u0026plusmn; 0.018\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 253px;\"\u003e\n \u003cp\u003eProteomic Myeloid Inflammation Score\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 133px;\"\u003e\n \u003cp\u003eCSF/Plasma MS\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 119px;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 119px;\"\u003e\n \u003cp\u003e0.138 \u0026plusmn; 0.021\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 253px;\"\u003e\n \u003cp\u003ectDNA Methylation Fraction\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 133px;\"\u003e\n \u003cp\u003ecfDNA EPIC 850K\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 119px;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 119px;\"\u003e\n \u003cp\u003e0.131 \u0026plusmn; 0.024\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 253px;\"\u003e\n \u003cp\u003erCBV_max (contrast-enhancing lesion)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 133px;\"\u003e\n \u003cp\u003eDSC Perfusion\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 119px;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 119px;\"\u003e\n \u003cp\u003e0.119 \u0026plusmn; 0.027\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 253px;\"\u003e\n \u003cp\u003eADC_StretchExp_D* (peri-lesional)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 133px;\"\u003e\n \u003cp\u003eDWI \u0026ndash; Stretched-exponential\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 119px;\"\u003e\n \u003cp\u003e96%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 119px;\"\u003e\n \u003cp\u003e0.108 \u0026plusmn; 0.031\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 253px;\"\u003e\n \u003cp\u003eIL-6 (CSF, log-transformed)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 133px;\"\u003e\n \u003cp\u003eCSF Proteomics\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 119px;\"\u003e\n \u003cp\u003e96%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 119px;\"\u003e\n \u003cp\u003e0.099 \u0026plusmn; 0.028\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 253px;\"\u003e\n \u003cp\u003eGLCM_Contrast_T1post (CE lesion)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 133px;\"\u003e\n \u003cp\u003eT1 post-gadolinium\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 119px;\"\u003e\n \u003cp\u003e92%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 119px;\"\u003e\n \u003cp\u003e0.091 \u0026plusmn; 0.034\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 253px;\"\u003e\n \u003cp\u003eCXCL1 (CSF, log-transformed)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 133px;\"\u003e\n \u003cp\u003eCSF Proteomics\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 119px;\"\u003e\n \u003cp\u003e92%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 119px;\"\u003e\n \u003cp\u003e0.087 \u0026plusmn; 0.033\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 253px;\"\u003e\n \u003cp\u003eGlioma Methylation Classification Prob.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 133px;\"\u003e\n \u003cp\u003ecfDNA EPIC 850K\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 119px;\"\u003e\n \u003cp\u003e88%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 119px;\"\u003e\n \u003cp\u003e0.083 \u0026plusmn; 0.029\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 253px;\"\u003e\n \u003cp\u003eMGMT Methylation Status (binary)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 133px;\"\u003e\n \u003cp\u003eClinical covariate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 119px;\"\u003e\n \u003cp\u003e88%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 119px;\"\u003e\n \u003cp\u003e0.078 \u0026plusmn; 0.035\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 253px;\"\u003e\n \u003cp\u003eADC_mean (CE lesion, monoexponential)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 133px;\"\u003e\n \u003cp\u003eDWI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 119px;\"\u003e\n \u003cp\u003e84%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 119px;\"\u003e\n \u003cp\u003e0.071 \u0026plusmn; 0.038\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 253px;\"\u003e\n \u003cp\u003eGLRLM_RunVariance_FLAIR (peri-lesional)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 133px;\"\u003e\n \u003cp\u003eT2/FLAIR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 119px;\"\u003e\n \u003cp\u003e76%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 119px;\"\u003e\n \u003cp\u003e0.058 \u0026plusmn; 0.041\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 253px;\"\u003e\n \u003cp\u003eCD163 (plasma, log-transformed)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 133px;\"\u003e\n \u003cp\u003ePlasma Proteomics\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 119px;\"\u003e\n \u003cp\u003e72%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 119px;\"\u003e\n \u003cp\u003e0.051 \u0026plusmn; 0.044\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 253px;\"\u003e\n \u003cp\u003eWavelet_HLL_GLCM_Correlation (CE lesion)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 133px;\"\u003e\n \u003cp\u003eT1 post-gadolinium\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 119px;\"\u003e\n \u003cp\u003e64%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 119px;\"\u003e\n \u003cp\u003e0.043 \u0026plusmn; 0.047\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 253px;\"\u003e\n \u003cp\u003eECOG Performance Status\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 133px;\"\u003e\n \u003cp\u003eClinical covariate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 119px;\"\u003e\n \u003cp\u003e60%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 119px;\"\u003e\n \u003cp\u003e0.038 \u0026plusmn; 0.041\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"4\" valign=\"top\" style=\"width: 624px;\"\u003e\n \u003cp\u003e\u003cem\u003eTop 15 of 32 retained features shown, ranked by mean |SHAP| value. Selection frequency = proportion of 25 outer cross-validation folds in which the feature was retained after recursive feature elimination. Features with selection frequency \u0026ge;80% (n=10) are considered stable. ADC = apparent diffusion coefficient; rCBV = relative cerebral blood volume; CE = contrast-enhancing; DSC = dynamic susceptibility contrast; GLCM = gray-level co-occurrence matrix; GLRLM = gray-level run-length matrix.\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eProteomic Myeloid Inflammation Signature and TMB Correlation\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIL-6 and CXCL1 were the most strongly differentially expressed proteomic markers: both were significantly elevated in treatment effect compared to true progression (IL-6: fold change 2.8, FDR p = 0.002; CXCL1: fold change 3.4, p \u0026lt; 0.001). The composite myeloid inflammation signature (CD163, CXCL1, CXCL8, IL-6, osteopontin) was enriched in treatment effect cases. Radiation necrosis cases showed higher IL-6 elevation than PsP (median z-score 2.41 vs. 1.67; p = 0.03), reported as hypothesis-generating given the small radiation necrosis subgroup (n = 23).\u003c/p\u003e\n\u003cp\u003eThe correlation between the myeloid inflammation score and cfDNA-estimated TMB was Pearson r = 0.42 (p = 0.006). This correlation is explicitly considered exploratory and hypothesis-generating. TMB was computationally estimated from cfDNA methylation data; tissue-based WES was available in 38 patients, in whom cfDNA-estimated and tissue-based TMB were moderately correlated (r = 0.61, p \u0026lt; 0.001), providing partial but not definitive support for the approach. The absence of genome-wide tissue WES in the full cohort is an acknowledged limitation, and all TMB-related interpretations should be interpreted accordingly.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eCost Breakdown and Workflow Feasibility\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTable 3 provides a component-level breakdown of the ~$250/sample cost. Hands-on laboratory time is 6\u0026ndash;8 hours; total elapsed time is 48\u0026ndash;72 hours under standard batch scheduling. At sites implementing centralized processing, per-sample costs were reduced by approximately 18%, and turnaround was consistent at 48 hours. Sites dependent on outreach shipping would experience a 72\u0026ndash;96 hour total elapsed time.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 3. Per-Sample Cost Breakdown (~$250/sample)\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"624\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 280px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eComponent\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 172px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCost per Sample (USD)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 172px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eNotes\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 280px;\"\u003e\n \u003cp\u003ecfDNA extraction (QIAamp kit, per sample)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 172px;\"\u003e\n \u003cp\u003e$28\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 172px;\"\u003e\n \u003cp\u003eConsumables only\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 280px;\"\u003e\n \u003cp\u003eBisulfite conversion (EZ Methylation-Gold)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 172px;\"\u003e\n \u003cp\u003e$14\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 172px;\"\u003e\n \u003cp\u003eConsumables only\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 280px;\"\u003e\n \u003cp\u003eEPIC 850K array (amortized, batch of 8)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 172px;\"\u003e\n \u003cp\u003e$62\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 172px;\"\u003e\n \u003cp\u003eExcludes instrument depreciation\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 280px;\"\u003e\n \u003cp\u003eDIA-MS proteomics (SP3 prep + instrument time, amortized)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 172px;\"\u003e\n \u003cp\u003e$55\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 172px;\"\u003e\n \u003cp\u003eConsumables + instrument time, shared-access core facility rate\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 280px;\"\u003e\n \u003cp\u003eBioinformatics pipeline (cloud compute, per sample)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 172px;\"\u003e\n \u003cp\u003e$32\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 172px;\"\u003e\n \u003cp\u003eAWS/GCP cloud compute, ~3 CPU-hours\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 280px;\"\u003e\n \u003cp\u003eTechnician hands-on labor (0.5 hr @ $50/hr fully loaded)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 172px;\"\u003e\n \u003cp\u003e$25\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 172px;\"\u003e\n \u003cp\u003eExcludes MRI acquisition cost\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 280px;\"\u003e\n \u003cp\u003eQuality control, reagent waste, consumable overage (~15%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 172px;\"\u003e\n \u003cp\u003e$34\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 172px;\"\u003e\n \u003cp\u003eApproximate\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 280px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTOTAL (approximate)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 172px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e~$250\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 172px;\"\u003e\n \u003cp\u003eInstrument depreciation and site overhead not included\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"3\" valign=\"top\" style=\"width: 624px;\"\u003e\n \u003cp\u003e\u003cem\u003eTurnaround: hands-on laboratory time 6\u0026ndash;8 hours total; total elapsed time 48\u0026ndash;72 hours (includes array scanning batch scheduling and overnight computational pipeline). At scale with centralized processing (\u0026gt;20 samples/batch), instrument depreciation and overhead would reduce per-sample cost; centralized processing was confirmed feasible at two of the seven participating sites during the study period. Note: the \u0026asymp;$250 per-sample estimate assumes batch processing of \u0026ge;8 samples per run; single-sample runs would increase per-sample cost substantially (estimated 35\u0026ndash;45% premium) due to fixed instrument and reagent overhead amortization.\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u0026nbsp;\u003cstrong\u003e\u003cem\u003eDynamic Biomarker Axis (Exploratory Analysis)\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIn the 89 patients with \u0026ge;2 longitudinal sampling timepoints, 34 cases were classified as cross-sectionally indeterminate (ensemble score within the 40th\u0026ndash;60th percentile confidence interval). Of these, 26/34 (76%) were reclassified by the dynamic biomarker axis algorithm. This analysis is reported as exploratory, post-hoc, and hypothesis-generating, and the 76% figure must not be interpreted as a validated performance metric. A prospective study with the dynamic axis as a pre-specified primary endpoint is registered and in protocol development.\u003c/p\u003e"},{"header":"CONCLUSIONS","content":"\u003cp\u003eThis revised pooled prospective multicenter analysis demonstrates that an integrated multi-compartment biomarker ecosystem achieves robust internally validated diagnostic accuracy (AUC 0.92; AUC 0.89 in histopathologically confirmed cases) for differentiating true tumor progression from treatment effect in HGG, with a 6.2-week lead-time advantage over radiographic consensus and a clinically deployable plasma-only variant (AUC 0.86). Key methodological concerns raised in peer review—ground truth circularity, feature selection stability, the exploratory nature of the dynamic biomarker axis, plasma-only performance, and cost transparency—have been directly addressed. Ten of 32 retained features demonstrate ≥80% fold-selection stability, anchored by the mechanistically interpretable triad of stretched-exponential ADC alpha, proteomic myeloid inflammation score, and cfDNA methylation fraction. External prospective validation, powered for overall survival as a primary endpoint with the dynamic biomarker axis as a pre-specified secondary endpoint, is the immediate and essential next step. If confirmed, this ecosystem has the potential to transform post-treatment surveillance and therapy adaptation in neuro-oncology.\u003c/p\u003e\n\u003cp\u003e\u003cbr\u003e\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003e\u003cem\u003eFunding\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was supported by institutional research funds. No external funding sources were involved in study design, data collection, analysis, or the decision to publish.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eConflict of Interest\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare no conflict of interest.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eEthical Approval\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll studies contributing data were approved by their respective institutional review boards or ethics committees. Written informed consent was obtained from all participants. Studies were conducted in accordance with the Declaration of Helsinki.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eAuthor Contributions\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eM.T.K.: conceptualization, study design, methodology, analysis, writing – original draft and revision. A.W.I. (Abdelrahman): data curation, bioinformatics, statistical analysis, writing – review and editing. A.W.I. (Abdullah): data curation, clinical adjudication, writing – review and editing. All authors reviewed and approved the final revised manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eData Availability\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAggregated summary data will be made available upon reasonable request. Individual patient data sharing is subject to institutional data use agreements. Bioinformatics code for the ensemble classifier and dynamic axis algorithm will be deposited in a public GitHub repository (https://github.com/[repository-to-be-inserted]) upon acceptance; the repository URL will be finalized during production. All code will be released under an MIT open-source license.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eSabedot TS, Malta TM, Snyder J, et al. A serum-based DNA methylation assay provides accurate detection of glioma. Neuro Oncol. 2021;23(8):1494\u0026ndash;1508.\u003c/li\u003e\n\u003cli\u003eDai L, Liu Z, Zhu Y, Ma L. Genome-wide methylation analysis of circulating tumor DNA: A new biomarker for recurrent glioblastoma. Heliyon. 2023;9(2):e13274.\u003c/li\u003e\n\u003cli\u003eWang Z, Maimaiti A, Sun Y. DNA methylation-based epigenetic signatures for classification and non-invasive diagnosis of gliomas [abstract 665MO]. Ann Oncol. 2025;36(Suppl 1):S485.\u003c/li\u003e\n\u003cli\u003eRiviere-Cazaux C, Rajani K, Carlstrom LP, et al. A field resource for the glioma cerebrospinal fluid proteome: Impacts of resection and location on biomarker discovery. Neuro Oncol. 2025;27(4):948\u0026ndash;962.\u003c/li\u003e\n\u003cli\u003eRaza IJ, Bulstrode H, Jimenez-Linan M, et al. Blood Biomarkers of Glioma in Response Assessment Including Pseudoprogression and Other Treatment Effects: A Systematic Review. Front Oncol. 2020;10:43.\u003c/li\u003e\n\u003cli\u003eSidibe I, Tensaouti F, Lotterie JA, et al. Pseudoprogression in Glioblastoma: Role of Metabolic and Functional MRI-Systematic Review. Biomedicines. 2022;10(8):1858.\u003c/li\u003e\n\u003cli\u003eYadav VK, Patel P, Jha S, et al. Distinction of pseudoprogression from true progression in glioblastomas using machine learning based on multiparametric MRI and MGMT promoter methylation status. Neuro Oncol Adv. 2024;6(1):vdae159.\u003c/li\u003e\n\u003cli\u003eLiao D, Liu YC, Liu JY, Wang D, Liu XF. Differentiating tumour progression from pseudoprogression in glioblastoma patients: a monoexponential, biexponential, and stretched-exponential model-based DWI study. BMC Med Imaging. 2023;23(1):119.\u003c/li\u003e\n\u003cli\u003eHodges TR, Ott M, Xiu J, et al. Mutational burden, immune checkpoint expression, and mismatch repair in glioma: implications for immune checkpoint immunotherapy. Neuro Oncol. 2017;19(8):1047\u0026ndash;1057.\u003c/li\u003e\n\u003cli\u003eHoosemans LJN, Compter I, Kocakavuk E, et al. Clinical predictors of pseudoprogression in glioblastoma: a retrospective cohort analysis. J Neurooncol. 2025;176(1):42.\u003c/li\u003e\n\u003cli\u003eSeyhan AA. Circulating Liquid Biopsy Biomarkers in Glioblastoma: Advances and Challenges. Int J Mol Sci. 2024;25(2):917.\u003c/li\u003e\n\u003cli\u003eJuan YC, Wang HC, Ko CC, et al. Beyond the blood-brain barrier: feasibility and technical validation of dual-compartment circulating tumor cells detection in high-grade glioma patients. Neurosurg Rev. 2025;48(1):359.\u003c/li\u003e\n\u003cli\u003eSoffietti R, Ahluwalia M, Lin N, et al. Liquid biopsy in gliomas: A RANO review and proposals for clinical applications. Neuro Oncol. 2022;24(6):855\u0026ndash;871.\u003c/li\u003e\n\u003cli\u003eHassan MF, Al-Zurfi AN, Alsalihi MH, Ahmed K. An effective ensemble learning approach for classification of glioma grades based on novel MRI features. Sci Rep. 2024;14(1):11977.\u003c/li\u003e\n\u003cli\u003eChristodoulou RC, Tsolaki E, Valotassiou V, et al. Artificial Intelligence in Glioma Diagnosis: A Narrative Review of Radiomics and Deep Learning for Tumor Classification and Molecular Profiling Across PET and MRI. Eng. 2025;6(2):18.\u003c/li\u003e\n\u003cli\u003ePalavani LB, Menegat M, Nascimento FO, et al. Artificial intelligence algorithms for differentiating pseudoprogression from true progression in high-grade gliomas: A systematic review and meta-analysis. Neurosurg Rev. 2025;48(1):591.\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"Merit university ","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":true,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"glioblastoma; pseudoprogression; liquid biopsy; DNA methylation; radiomics; artificial intelligence; CSF proteomics; dynamic biomarker axis","lastPublishedDoi":"10.21203/rs.3.rs-9204627/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9204627/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cem\u003e\u003cstrong\u003eBackground\u003c/strong\u003e\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eDistinguishing true tumor progression from treatment effect—encompassing pseudoprogression (PsP) and radiation necrosis—remains one of the most pressing and unresolved challenges in high-grade glioma (HGG) management. Conventional contrast-enhanced MRI is frequently ambiguous in the post-chemoradiation period, and blood–brain barrier dynamics constrain single-compartment liquid biopsy assays. We hypothesized that an integrated multi-compartment biomarker ecosystem would overcome these limitations.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u003cstrong\u003eMethods\u003c/strong\u003e\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003ePooled analysis of seven multicenter prospective studies (n = 212; 140 newly diagnosed, 72 recurrent; 160 IDH-wildtype GBM, 52 other HGG). The ecosystem comprised plasma cfDNA methylation profiling (EPIC 850K array), paired CSF/plasma data-independent acquisition (DIA) mass spectrometry proteomics, multiparametric MRI radiomics with ComBat harmonization, and a stacked AI ensemble classifier. Ground truth: histopathological confirmation in 67% of cases (n = 142); pre-specified radiographic consensus (independent adjudication, modified RANO, ≥6-month follow-up) in 33% (n = 70). A plasma-only ecosystem variant (no CSF) was pre-specified for the subgroup without CSF availability. Feature selection stability was assessed by outer-fold selection frequency across nested cross-validation. The dynamic biomarker axis is reported as an exploratory analysis.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u003cstrong\u003eResults\u003c/strong\u003e\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThe full ecosystem achieved AUC 0.92 (95% CI 0.88–0.96), sensitivity 91.5%, and specificity 84.1% (Table 2). In the sensitivity analysis restricted to histopathologically confirmed cases (n = 142), AUC was 0.89 (95% CI 0.83–0.95), confirming that performance was not substantially dependent on radiographically assigned ground truth. The plasma-only ecosystem achieved AUC 0.86. Median lead-time over radiographic consensus was 6.2 weeks. Ten of 32 retained features demonstrated ≥80% fold-selection stability. Proteomic myeloid inflammation markers (IL-6, CXCL1) were enriched in treatment effect; correlation with cfDNA-estimated TMB (r ≈ 0.42) is reported as hypothesis-generating. The dynamic biomarker axis (exploratory, n = 89 longitudinal cases) resolved 76% of cross-sectionally indeterminate cases and requires prospective validation.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u003cstrong\u003eConclusions\u003c/strong\u003e\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThis integrated ecosystem achieves robust internal performance for treatment effect discrimination in HGG. The histopathologically confirmed sensitivity analysis, plasma-only model, feature stability data, and formal dynamic axis algorithm address key methodological concerns. External prospective validation is the essential and pre-planned next step.\u003c/p\u003e","manuscriptTitle":"An Integrated Multi-Compartment Biomarker Ecosystem for the Discrimination of True Tumor Progression from Treatment Effect in High-Grade Glioma: Pooled Analysis of Seven Multicenter Prospective Studies (n = 212)","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-03-25 03:10:52","doi":"10.21203/rs.3.rs-9204627/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"730b00b9-9627-422e-9309-99555d9f799c","owner":[],"postedDate":"March 25th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":65006921,"name":"Oncology"},{"id":65006922,"name":"Neurology"}],"tags":[],"updatedAt":"2026-03-25T03:10:52+00:00","versionOfRecord":[],"versionCreatedAt":"2026-03-25 03:10:52","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9204627","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9204627","identity":"rs-9204627","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}