Search Results (3,340)

The bone marrow tumor microenvironment (TME) is critical for acute myeloid leukemia (AML) progression, immune evasion, and treatment resistance. SRC, a non-receptor tyrosine kinase involved in multiple oncogenic pathways, has not been systematically characterized in AML in relation to prognosis and immune regulation. We integrated bulk transcriptomic and single-cell RNA-sequencing datasets from TCGA, BeatAML, and GEO. Immune-related targets were identified using xCell-based immune scoring and weighted gene co-expression network analysis (WGCNA), followed by protein–protein interaction analysis and multi-algorithm machine-learning screening. We then evaluated SRC expression patterns, prognostic associations, immune microenvironment features, predicted drug sensitivity, single-cell differentiation dynamics, intercellular communication, and in silico virtual knockout perturbation (scTenifoldKnk). SRC emerged as the most robust hub gene after integration of WGCNA, PPI analysis, machine-learning feature selection, and survival screening. SRC was significantly upregulated in AML compared with normal controls and was independently associated with poor overall survival (HR = 1.231, p = 0.037). High SRC expression was linked to adverse ELN/FAB features, increased immune checkpoint expression, enrichment of inflammatory and immunoregulatory pathways, and a higher proportion of primitive leukemia-associated cell states. Single-cell analyses further suggested that SRC was enriched in CD34⁺ progenitor compartments, associated with altered cell–cell communication, and accompanied by distinct mutation and pathway profiles. Drug-response prediction and in silico network perturbation analysis further supported the potential biological and translational relevance of SRC-centered programs. SRC is a prognostically relevant and immune-associated hub linked to AML microenvironment remodeling, and may serve as a candidate biomarker and potential therapeutic target that warrants further experimental validation. Full article

Assessing disease severity and prognosis in canine acute pancreatitis (AP) remains a major clinical challenge. This study evaluated the clinical utility of the Adapted Modified Canine Activity Index (aMCAI), a clinical scoring system refined from the original MCAI. A prospective observational study was conducted on 42 dogs diagnosed with AP, with aMCAI scores assessed on Days 1, 3, and 5. A linear mixed model (LMM) was used to analyze score progression over time and differences between survivors and non-survivors. Receiver operating characteristic (ROC) curves evaluated the prognostic accuracy for 30 day survival. The LMM analysis revealed that non-survivors had significantly higher aMCAI scores than survivors (p = 0.035), and overall scores decreased significantly over time (p < 0.001). ROC analysis showed poor discrimination on Days 1 and 3. However, on Day 5 the aMCAI demonstrated good prognostic performance (AUC = 0.813, p < 0.001). A cutoff value of ≥2.5 on Day 5 yielded 100% sensitivity, a negative likelihood ratio of 0.00 and a 100% negative predictive value, providing clinically relevant prognostic information. These findings suggest that the aMCAI is a practical tool for monitoring disease progression and may support the identification of dogs with a high likelihood of survival. Full article

Background: Sepsis remains a leading cause of mortality in the intensive care unit (ICU). Platelets (PLTs) are central to coagulation, inflammation, and the maintenance of endothelial integrity. Although thrombocytopenia is an established prognostic marker in sepsis, alterations in PLT function and morphology may provide additional insight into disease progression. Methods: This retrospective pilot study examined adult ICU patients diagnosed with sepsis or septic shock. Extracted data included demographic characteristics, clinical variables, and laboratory parameters. Platelet function was evaluated using impedance aggregometry and rotational thromboelastometry (ROTEM), while PLT morphology metrics were obtained from complete blood counts. Statistical analyses comprised Spearman’s rank correlation and logistic regression. Results: Twenty patients were included. Platelet aggregation was impaired across ASPI, ADP, and TRAP-6 assays despite normal PLT counts and morphology. ROTEM-derived measure of PLT contribution to clot strength was within normal ranges. No correlations were observed between PLT function and PLT morphology parameters. An inverse correlation was identified between ROTEM-derived PLT contribution to clot strength and SOFA score (r = −0.60, p = 0.03). Neither PLT function nor PLT morphology was associated with ICU mortality. Conclusions: Functional PLT deficits may occur in sepsis in the absence of structural abnormalities. ROTEM-derived PLT contribution to clot strength may inversely reflect sepsis severity. Platelet function parameters appear unlikely to predict short-term mortality in septic patients. Full article

Neonates undergoing surgery for congenital heart disease (CHD) are at high risk for brain function impairment. Reliable early predictors of postoperative neurological complications are lacking. We examined a retrospective cohort of 55 surgically treated CHD neonates systematically monitored by concomitant conventional electroencephalography (cEEG) and amplitude-integrated EEG (aEEG). Neonates underwent cEEG/aEEG at three time points: T0 (preoperative, duration: 90–120 min); T1 (24–48 h after cardiac surgery, duration: ≥11 h); and T2 (7–10 days post-surgery, duration: 90–120 min). For each patient, aEEG background activity was evaluated and scored, and clinical and surgical data were retrieved to establish short-term post-surgical outcomes. Patients with normal T0 monitoring had significantly higher aEEG bandwidths in T1. A lower Aristotle basic score was associated with an improvement in aEEG at T1. Inversely, a narrower aEEG bandwidth in T1 was associated with post-surgical neurological deterioration. The aEEG bandwidth accurately predicted short-term neurological outcome; in particular, a minimal aEEG amplitude above 17.5 µV excluded poor neurological outcome with a negative predictive value of 81.48%. Our results demonstrated that aEEG bandwidth and trend dynamics may be associated with surgical complexity and neurological outcomes. aEEG background trend monitoring may provide relevant prognostic information on neurological outcomes in surgically treated CHD neonates. Full article

Background/Objectives: Non-variceal upper gastrointestinal bleeding (NVUGIB) remains a major clinical emergency, particularly among patients receiving antiplatelet or anticoagulant therapy, whose use has increased substantially in recent years. This study aimed to evaluate the clinical characteristics, endoscopic findings, risk stratification, and outcomes of NVUGIB in patients receiving antithrombotic therapy, and to compare the predictive performance of commonly used prognostic scores. Methods: This prospective cohort study included 89 patients receiving antithrombotic therapy who presented with NVUGIB at Beni-Suef University Hospitals between March 2023 and March 2025. Clinical presentation, laboratory findings, and endoscopic characteristics were recorded. Risk stratification was assessed using Glasgow–Blatchford (GBS), Rockall, Baylor, AIMS65, ABC, and PNED scores. The optimal cut-off values for prediction of rebleeding and mortality were determined using receiver operating characteristic (ROC) analysis and the Youden index. Area under the curve (AUC) values were reported with 95% confidence intervals. Results: Endoscopy revealed that peptic ulcers were the most common lesion (41/89, 46%), followed by erosive disease (27/89, 30%), with the stomach being the most frequently involved site (76.5%). Rebleeding occurred in 16 patients (18.0%), while mortality was observed in 2 patients (2.2%). The Glasgow–Blatchford score demonstrated the most consistent performance for predicting rebleeding, with an optimal cutoff value of 5.5 (derived using the Youden index), yielding 92.9% sensitivity and 78.8% specificity. For mortality prediction, AIMS65, ABC, and PNED scores showed very high AUC values, although these findings should be interpreted cautiously due to the small number of mortality events (n = 2). No statistically significant difference in rebleeding or mortality was observed between single and dual antithrombotic therapy, although patients receiving dual therapy required longer hospitalization and more transfusion units. Conclusions: In patients with antithrombotic-related GI bleeding, ulcers and erosions predominate, with minimal differences between single and dual therapy outcomes. Concomitant NSAID use trends toward higher mortality. Glasgow–Blatchford score offers optimal performance for both rebleeding and mortality prediction, with a cutoff of 5.5 providing excellent sensitivity (92.9%) and specificity (78.8%) for rebleeding risk assessment. Full article

Background: Pulmonary embolism (PE) is a major cause of morbidity and in-hospital mortality and requires rapid clinical assessment for diagnosis, risk stratification, and management. Several validated clinical prediction tools, including the Pulmonary Embolism Severity Index (PESI), WELLS, PADUA, and IMPROVE scores, are commonly used to evaluate thromboembolic risk and predict clinical outcomes. Methods: A retrospective observational study was conducted on a cohort of 538 patients diagnosed with PE, all recruited between January 2020–December 2025. Group comparisons between survivors and non-survivors were performed using independent samples t-tests and Mann–Whitney U tests. Receiver operating characteristic (ROC) curve analysis was used to evaluate the predictive performance of clinical scores. Multivariable logistic regression analysis was performed to identify independent predictors of in-hospital mortality. Results: Mean age was 69 years, and overall death-rate was 18.4%. Significant differences between survivors and non-survivors were observed for age and clinical scores. White blood cell count, neutrophils, lymphocytes, platelet count, total bilirubin, aspartate aminotransferase (AST), alanine aminotransferase (ALT), gamma-glutamyl transferase (GGT), procalcitonin, and international normalized ratio (INR), were significantly associated with in-hospital mortality. ROC curve analysis demonstrated predictive performance of the evaluated clinical scores. Logistic regression identified PESI score, procalcitonin levels, and white blood cell count as independent predictors of unfavorable outcome. mortality. Conclusions: Clinical risk scores and specific laboratory biomarkers were associated with in-hospital mortality in patients with pulmonary embolism. The PESI score, procalcitonin, and white blood cell count showed independent predictive value for death rate in this cohort. Full article

Background/Objectives: Twin pregnancies carry substantially elevated perinatal risks, yet tools for first-trimester risk stratification remain limited. This retrospective cohort study evaluated the predictive value of crown–rump length (CRL) and nuchal translucency (NT) discordance for adverse perinatal outcomes in 184 twin pregnancies at Ankara Etlik City Hospital, Turkey (October 2022–January 2024). Methods: CRL discordance ≥ 10% and NT discordance ≥ 20% were assessed for a birth-weight-based proxy of fetal growth restriction (FGR), preeclampsia, and neonatal outcomes using multivariable logistic regression adjusted for chorionicity, body mass index (BMI), and conception mode. Results: CRL discordance ≥ 10% was independently associated with the birth-weight-based FGR proxy (adjusted odds ratio [OR] 7.79, 95% confidence interval [CI] 3.95–20.12, p < 0.001; area under the curve [AUC] 0.736). NT discordance ≥ 20% was also independently associated with the birth-weight-based FGR proxy (OR 3.74, 95% CI 1.91–8.39, p < 0.001; AUC 0.612). Both parameters were associated with lower Apgar scores. IVF conception was independently associated with preeclampsia in an exploratory analysis (OR 5.31, 95% CI 1.41–28.66, p = 0.016). Continuous modelling confirmed a dose–response relationship for CRL discordance (OR per 1% increase = 1.20, 95% CI 1.13–1.32). Conclusions: These findings suggest that first-trimester CRL discordance may provide useful early prognostic information for birth-weight-based adverse growth outcome in twin pregnancies, pending prospective validation in cohorts with Doppler-based FGR ascertainment. Full article

Background/Objectives: Carcinoembryonic antigen (CEA) is widely used in colon cancer management; however, its diagnostic and prognostic accuracy is limited by biological variability, as well as false-positive or false-negative results. Radiomics provides quantitative descriptors of tumor heterogeneity and offers objective assessment of tumor characteristics. This study aimed to evaluate the potential of computed tomography (CT)-based radiomic features to distinguish between CEA-positive and CEA-negative colon cancer patients. Methods: In this retrospective study, 150 patients with histopathologically confirmed colon cancer were screened, and 109 were eligible after image-quality assessment (53 CEA-positive, 56 CEA-negative). A total of 107 radiomic features were extracted from preoperative contrast-enhanced CT images. After z-score normalization, feature robustness was assessed using intra- and inter-observer agreement. Correlation-based feature selection (|ρ| ≥ 0.7) was applied. Five machine-learning classifiers—Support Vector Machine (SVM), Decision Tree, Ensemble, k-Nearest Neighbor (k-NN), and Neural Network (NN)—were trained using stratified 5-fold cross-validation. Performance was evaluated using accuracy, recall, specificity, F1-score, and ROC-AUC. Results: The best performance was obtained with 41 selected features. The k-NN classifier achieved the highest accuracy (77.4 ± 2%) and ROC-AUC (0.8523 ± 0.013), while SVM and NN achieved the highest recall (83.0 ± 0.3). These models showed balanced and robust performance in distinguishing CEA-positive from CEA-negative patients. Conclusions: CT-based radiomic analysis combined with machine learning—particularly k-NN, SVM, and neural network classifiers—showed promising performance in differentiating colon cancer patients according to serum CEA status. Radiomic features may provide imaging-based information associated with serum biomarkers such as CEA, potentially enhancing tumor characterization and supporting more personalized decision-making. Full article

Background: Neoadjuvant chemotherapy (NAC) is widely used in the management of stage I–III breast cancer, with tumor regression serving as an important surrogate for long-term outcome. Identifying reliable pathological biomarkers predictive of residual disease remains clinically relevant. Methods: We conducted a retrospective cohort study of 165 patients with non-metastatic breast cancer treated with neoadjuvant chemotherapy followed by surgery between 2019 and 2022. Pathological response was assessed using the Residual Cancer Burden (RCB) index. The primary study endpoint was extensive residual disease (RCB-III), defined as the poorest category of tumor regression, indicating treatment resistance. Associations between the Nottingham Score together with other histopathological parameters, immunohistochemical markers (ER, PR, HER2), Ki67 proliferation index, and RCB were analyzed using univariate and multivariable logistic regression. Results: In univariate analysis, higher Nottingham scores (OR = 1.807, p = 0.0017), negative ER expression (OR = 3.017, p = 0.0255), the absence of lymphovascular invasion (OR = 0.1877, p = 0.0069) and elevated Ki67 (OR = 1.034, p = 0.0003) were significantly associated with RCB III. In multivariable analysis, only Ki67 and lymphovascular invasion remained independent predictors of RCB III, while Nottingham score and ER expression lost statistical significance. Correlation analysis demonstrated strong associations between Nottingham score, Ki67, hormone receptor loss, and tumoral necrosis. Conclusions: Ki67 is an independent predictor of poor tumor regression following neoadjuvant chemotherapy and appears to capture much of the prognostic information traditionally attributed to histologic grade and Nottingham score. However, the absence of lymphovascular invasion remains a significant positive prognostic factor. These observations support further investigation into the integration of proliferation markers into multivariable predictive models to improve response stratification in breast cancer. Full article

Background: Frailty is a major determinant of adverse outcomes in older adults with cardiovascular disease. Automated digital tools may facilitate routine frailty assessment in hospital settings; however, their validity and prognostic relevance in acutely hospitalized patients remain insufficiently established. Methods: In this prospective cohort study, 113 hospitalized cardiology patients underwent frailty assessment using both manual Short Physical Performance Battery (mSPPB) and an automated electronic SPPB (eSPPB) system. Agreement between methods was evaluated using Pearson correlation, intraclass correlation coefficients (ICCs), and Bland–Altman analysis. Frailty was defined as SPPB < 5. The association between frailty and 30-day mortality was assessed using logistic regression and Kaplan–Meier survival analysis. Results: Seventeen patients (15.0%) were classified as frail. Automated and manual SPPB scores were highly correlated (r = 0.994, p < 0.001) and demonstrated good agreement (ICC = 0.80). Bland–Altman analysis showed a mean difference of −1.63 points (95% limits of agreement −4.41 to 1.16). Frailty was associated with significantly higher 30-day mortality (17.6% vs. 2.1%, p = 0.009), corresponding to a tenfold increase in mortality odds (OR 10.07; 95% CI 1.5–67.5). An exploratory model showed apparent discriminative performance (AUC 0.83; 95% CI 0.71–0.95). Conclusions: Automated eSPPB demonstrated good agreement with manual assessment and was significantly associated with short-term mortality in hospitalized cardiovascular patients. These findings support the validity and potential clinical utility of automated frailty assessment for risk stratification in acute cardiology settings. Full article

Objective: To evaluate overall survival and to identify clinical, pathological, and demographic factors associated with survival in patients with stage III colon cancer. Methods: This retrospective cross-sectional study included 452 patients with stage III colon cancer who were followed at Ankara Bilkent City Hospital between 2005 and 2025. Patient data, including age, sex, ECOG performance status, comorbidities, tumor characteristics, treatment-related toxicities, and recurrence, were analyzed using PASW Statistics 18.0 (SPSS Inc., Chicago, IL, USA). Kaplan–Meier and log-rank tests were used for survival analysis. Prognostic factors, survival, mortality, and recurrence predictions were evaluated using machine learning algorithms, including coarse tree, bagged trees, support vector machines, and k-nearest neighbors. Furthermore, an explainable artificial intelligence framework was incorporated to improve model transparency and reveal clinically meaningful feature contributions. Model performance was assessed using accuracy, sensitivity, specificity, and F-score. Results: According to statistical analyses, older age, ECOG performance score ≥ 2, stage IIIC disease, N2-level lymph node metastasis, and the presence of comorbidities—particularly diabetes mellitus—were significantly associated with worse survival (p < 0.05). Machine learning analyses identified key prognostic factors, including positive surgical margins, rash, mucositis, thrombocytopenia, number of chemotherapy cycles, pathological tumor subtype, diarrhea, age at diagnosis, and anemia. SHAP analysis further demonstrated that treatment-related variables, particularly surgical margin positivity and chemotherapy-associated toxicities, were among the most influential predictors of survival. Several machine learning models outperformed traditional statistical methods in predicting mortality and recurrence, with the highest accuracy observed in ensemble methods such as coarse tree (87%) and bagged trees. Conclusions: This study identifies key prognostic factors influencing survival in stage III colon cancer and demonstrates that machine learning-based approaches can complement conventional statistical methods. The integration of clinical and treatment-related variables may improve individualized risk stratification and support clinical decision-making. These findings may also guide future large-scale, multicenter, and prospective studies. Full article

Multiple sclerosis (MS) is a chronic autoimmune and neurodegenerative disease characterized by marked clinical heterogeneity. While the genetic architecture underlying disease susceptibility is well established, the role of genetic factors in shaping disease prognosis remains clearly defined. In this structured narrative review, we examine available evidence on genetic contribution to key MS prognostic domains. This includes clinical outcomes, such as age at onset, relapse rate, disability progression, neurological sequelae, and cognitive impairment. We also consider radiological measures like brain and spinal cord lesion burden, gadolinium-enhancing lesions, and atrophy, as well as laboratory biomarkers, such as oligoclonal bands and Immunoglobulin G (IgG) index. Overall, current evidence suggests that genetic influences on prognosis are modest and highly heterogeneous. Only a limited number of associations—primarily from genome-wide association studies (GWAS)—have shown consistent replication, whereas many reported findings come from small candidate-gene studies and remain unconfirmed. Among these, the largest GWAS on age-related Multiple Sclerosis Severity Score (MSSS) identified a locus in the DYSF–ZNF638 region reaching genome-wide significance. The strongest evidence from GWAS relates to relapse rate, magnetic resonance imaging (MRI) measures (e.g., thalamic atrophy) and intrathecal IgG synthesis, the latter also reaching genome-wide significance. Interpretation of genotype–phenotype associations is further limited by small sample sizes, limited replication, heterogeneity in study design with the predominance of candidate-gene approaches, variability in outcome definitions, treatment exposure, and population ancestry. These limitations currently preclude the routine use of genetic markers for prognostic stratification in clinical practice. Larger studies and collaborative genetic consortia efforts are needed to improve statistical power and reproducibility. Additionally, emerging epigenetic studies may provide valuable insights into prognosis and disease management. Understanding which genetic factors can predict diverse MS courses could enhance patient management and enable personalized treatment approaches. Full article

Background: Malnutrition is highly prevalent among critically ill patients and has been associated with worse clinical outcomes, particularly in sepsis. Several nutritional risk scores have been proposed to identify patients at increased risk of mortality in the intensive care unit (ICU). This study aimed to evaluate the prognostic value of three commonly used nutritional indices—modified Nutrition Risk in the Critically Ill (mNUTRIC), Prognostic Nutritional Index (PNI), and Controlling Nutritional Status (CONUT)—for predicting mortality in septic ICU patients. Methods: In this prospective observational cohort study conducted at two ICUs, 155 critically ill patients at nutritional risk were evaluated, including 105 patients with sepsis and 50 without sepsis. The primary endpoint was ICU mortality. Nutritional risk scores (mNUTRIC, PNI, and CONUT) were calculated at ICU admission. Survival analysis was performed using Kaplan–Meier (KM) curves and log-rank tests to compare survival probabilities across nutritional risk categories. Cox proportional hazards regression analysis was used to assess the association between nutritional scores and ICU mortality. Of note, only 24 mortality events were recorded in the septic cohort, which limits the statistical power of the findings. Results: KM analysis revealed significantly reduced survival among patients with severe malnutrition, as measured by the PNI score (log-rank p = 0.044). Patients with high mNUTRIC scores showed a tendency toward lower survival probability compared with those with low nutritional risk, approaching statistical significance (log-rank p = 0.059). No significant survival differences were observed between CONUT categories (log-rank p = 0.380). In univariate Cox regression analysis, the mNUTRIC score was significantly associated with ICU mortality (HR 1.67, 95% CI 1.17–2.38, p = 0.005). Conclusions: In this selected cohort, mNUTRIC demonstrated the strongest univariate prognostic signal for ICU mortality; however, this association was attenuated and did not reach statistical significance after limited multivariable adjustment. These findings are exploratory and apply specifically to a cohort of septic ICU patients with confirmed nutritional risk and therefore should not be generalized to the broader population of critically ill septic patients. Full article

Objectives: Soft tissue sarcomas (STS) are biologically heterogeneous malignancies with unpredictable clinical behavior. Although tumor size, histological grade, and surgical margin status remain the main determinants of prognosis, additional biomarkers that integrate tumor biology and host-related factors are needed. The hemoglobin × albumin × lymphocyte/platelet (HALP) score reflects systemic inflammation and nutritional status. This study aimed to evaluate the association between preoperative HALP score and oncological as well as surgical outcomes in patients undergoing curative resection for STS. Materials and Methods: A retrospective cohort study was conducted including 46 consecutive patients who underwent surgery for STS between 2017 and 2025. HALP scores were calculated using preoperative laboratory parameters, and patients were stratified into low- and high-HALP groups according to the cohort median (24.9). Overall survival (OAS) and disease-free survival (DFS) were analyzed using the Kaplan–Meier method and Cox proportional hazards models. Surgical margin status and postoperative complications were also compared. Results: Patients with low HALP scores had significantly larger tumors, higher rates of non-R0 resection, and increased major complications (p < 0.05). Recurrence and mortality were more frequent in the low-HALP group. Kaplan–Meier analysis demonstrated significantly shorter OAS (log-rank p = 0.0034) and DFS (log-rank p = 0.0318) in patients with low HALP scores. In univariate Cox analysis, HALP was significantly associated with survival outcomes; however, in multivariate analysis, histological grade and surgical margin status remained independent prognostic factors, while HALP lost independent significance. Conclusions: A low preoperative HALP score is associated with aggressive tumor characteristics, increased surgical morbidity, and poorer survival in STS patients. Although HALP did not retain independent significance in multivariable analysis, its strong association with tumor aggressiveness and survival suggests that it may reflect the systemic manifestation of high-risk tumor biology. As a simple and cost-effective biomarker derived from routine laboratory parameters, HALP may support preoperative risk stratification and help identify patients with biologically aggressive disease. Full article

Background/Objectives: We developed and validated a multimodal magnetic resonance imaging (MRI) framework combining deep learning segmentation with radiomics to predict local recurrence in nasopharyngeal carcinoma (NPC). Methods: This retrospective two-center study included 1074 NPC patients treated between 2015 and 2019. Center 1 cases were split 8:2 into training and internal test sets, while Center 2 served for external validation. A multimodal Swin UNet model automatically segmented tumors from pretreatment T1-weighted, T2-weighted, and contrast-enhanced T1 (CET1) images. Radiomics features were extracted from expert-reviewed regions of interest, selected, and modeled using extreme gradient boosting for recurrence prediction. Results: The multimodal segmentation model maintained consistent but moderate Dice similarity coefficients (0.737, 0.666, and 0.726 for T1WI, T2WI, and CET1 in external validation). These values reflect the moderate overlap typical for nasopharyngeal carcinoma, given its highly infiltrative growth and ill-defined boundaries along complex anatomic interfaces. For local recurrence prediction, single-modality models reached external AUCs between 0.754 and 0.781. Importantly, the multimodal fusion model demonstrated numerical improvement over single modalities in the external validation set (e.g., vs. T1WI, p = 0.141), achieving an AUC of 0.910, accuracy of 0.908, sensitivity of 0.805, specificity of 0.946, and F1-score of 0.825. Conclusions: The multimodal MRI radiomics model, developed alongside a deep learning segmentation module, demonstrated favorable multicenter performance for evaluating NPC recurrence risk. The primary prognostic analysis was based on expert-reviewed regions of interest; a supplementary analysis using fully automatic segmentation masks yielded comparable, non-significantly different performance across all cohorts (Training AUC: 0.887; Internal Test AUC: 0.892; External Validation AUC: 0.885 vs. 0.910, p = 0.145), supporting the feasibility of future end-to-end deployment. Fusing multimodal features yielded numerical improvements over single-sequence models in external validation, providing a basis for post-treatment surveillance planning. Full article