Patient data were retrieved from the Surveillance, Epidemiology, and End Results (SEER) database, supported by the National Cancer Institute, as of May 1, 2023. Clinicopathological information was obtained using SEER*Stat software (version 8.4.3). The SEER-18 registries database (November 2022 Submission), covering approximately 28% of the U.S. population, was used. The study period spanned from January 2010 to December 2015, with the 2015 cutoff ensuring a minimum of 5 years of follow-up for all patients.

Patients with BCa were selected based on the histological code from the International Classification of Diseases for Oncology, Third Edition (ICD-O-3). Inclusion criteria included: a) a pathological diagnosis of UC and b) BCa as the primary cancer. Exclusion criteria were: a) diagnoses outside the study period, b) incomplete basic information (e.g., surgery, chemotherapy, tumor size, metastasis), and c) the presence of other tumors. The SEER data used in this study are publicly available and exempt from Ethics Committee approval. The patient selection flowchart is shown in Figure 1 .

Additionally, demographic data and clinical characteristics of patients with BCa admitted to the Department of Urology at the Affiliated Hospital of Guangdong Medical University from January 2021 to December 2023 were collected. The inclusion and exclusion criteria for the Chinese cohort mirrored those applied to the SEER cohort. Informed consent was obtained from all participants. This retrospective study adhered to the Declaration of Helsinki and was approved by the Clinical Research Ethics Committee of the Affiliated Hospital of Guangdong Medical University (ethical approval number: PJKT2024-105).

This study incorporated factors such as age, gender, race, tumor size, tumor grade, T stage, N stage, M stage, radiotherapy status, chemotherapy status, and surgery. To ensure cross-database compatibility, all variables were harmonized between SEER and the external cohort using standardized coding. Tumor size was categorized as < 3 cm or ≥ 3 cm based on preoperative imaging (CT/MRI) or pathological reports.

Treatment variables (surgery, chemotherapy, and radiotherapy) were classified as “Yes” if any modality was administered, regardless of specific regimens, due to limitations in SEER data. Metastasis status (M stage) was confirmed via imaging (CT/MRI/PET-CT) or biopsy according to AJCC 8^th edition criteria. Data processing was standardized: continuous variables (e.g., age) were categorized before analysis, and missing data led to exclusion based on the inclusion criteria. Bone, brain, and lung metastases were used as the primary outcome factors.

Continuous variables were converted into categorical variables for analysis. For nomogram construction, enrolled patients were randomly divided into a training group and a validation group in a 7:3 ratio. Univariate logistic regression analysis was performed to identify risk factors for predicting DM in patients with BCa. Factors with a P-value < 0.05 from the univariate analysis were further assessed using multivariate logistic regression to identify independent risk factors. These independent risk factors were then used to develop a predictive model for DM risk in patients with BCa using R software, with the results visualized in a nomogram. A calibration curve was generated to depict the relationship between actual and predicted probabilities. Additionally, the receiver operating characteristic (ROC) curve was plotted, and the area under the curve (AUC) was calculated to evaluate the model’s prediction accuracy. An independent external validation cohort was used to calibrate and validate the prediction model.

PubMed was searched for studies on BCa utilizing the SEER database to investigate DM. The final outcomes and predictive models from these studies were extracted for comparison with the current research. Six relevant studies were identified.

Four BCa transcriptomic datasets (GSE13507, GSE37817, GSE166716, GSE256292) were obtained from the GEO database (https://www.ncbi.nlm.nih.gov/geo/). After batch effect removal and data normalization, these datasets were used as the training cohort. Additionally, the latest TCGA-BCa transcriptomic data from the TCGA-BLCA database (https://portal.gdc.cancer.gov/) was acquired to serve as the validation cohort.

To identify robust biomarkers for BCa prognosis, a comprehensive computational pipeline was developed, integrating multiple transcriptomic datasets and advanced machine learning. Rigorous preprocessing and harmonization of all datasets ensured data quality and comparability. For microarray-based GEO datasets, background correction and quantile normalization were performed using the Robust Multi-array Average (RMA) algorithm in R. Batch effects across different studies and platforms were corrected using the ComBat algorithm from the sva package. RNA-seq data from TCGA were processed through a standardized pipeline, which included quality control with FastQC, adapter trimming with Trim Galore, and alignment to the GRCh38 reference genome with STAR. Gene-level counts were generated using featureCounts and normalized via the DESeq2 variance stabilizing transformation. Quality control was verified through PCA, which demonstrated that the corrected biological factors accounted for over 80% of the variance.

For feature selection, LASSO regression was applied via the glmnet package, analyzing all 18,543 genes from batch-corrected data against binary metastasis status using 10-fold cross-validation to determine the optimal λ parameter at 1 standard error from the minimum mean squared error. Genes with non-zero coefficients at this threshold were considered candidate biomarkers, and their stability was validated through 100 bootstrap iterations. Only genes selected in more than 80% of the iterations were advanced.

The final validation incorporated a random forest implementation via the ranger package, with 1,000 trees and a node size of 5, ranking features by permutation-based importance scores. The integration framework prioritized consensus genes that appeared in multiple algorithm outputs, requiring consistent directional effects and validation in the independent TCGA cohort with AUC > 0.75. Biological relevance was further assessed through KEGG pathway analysis, protein-protein interaction networks, and literature verification of cancer associations.

Statistical processing was conducted using R software. Count data were expressed as numbers or percentages, while continuous variables following a normal distribution were presented as mean ± standard deviation. Categorical variables were expressed as percentages. Univariate and multivariate logistic regression analyses were performed on SEER database data to identify significant risk factors (P < 0.05). These risk factors were used to construct a predictive model for DM in patients with BCa. The model’s calibration curve, ROC curve, and AUC were generated using Hmisc and ROCR packages in R.

For external validation, the identical nomogram scoring algorithm from the SEER training cohort was applied to the external cohort without recalibration. Model performance, including AUC and calibration curves, was evaluated consistently across all cohorts to eliminate methodological bias. A P-value < 0.05 was considered statistically significant.

A total of 27,484 patients with BCa from January 2010 to December 2015 were sourced from the SEER database for this study. After excluding 8,267 patients with other cancers and 16,904 patients with incomplete information, 2,313 eligible cases were selected. These cases were randomly divided into a training cohort of 1,619 cases and an internal validation cohort of 694 cases. The external validation cohort data (n = 112) were sourced from the Affiliated Hospital of Guangdong Medical University, covering the period from January 2021 to December 2023, based on the same inclusion and exclusion criteria.

No statistically significant differences were observed in the demographic and clinical characteristics between the training and internal validation cohorts, as shown in Table 1 . The majority of patients were aged 65 years or older, comprising 61.05% of the sample. In terms of T stage, patients with T2-T4 tumors accounted for 94.16%, while those with Tis, Ta, or T1 tumors made up only 5.84%. The training cohort included 63 patients diagnosed with DM, and the internal validation cohort had 27 such patients. A substantial proportion of patients underwent surgery (99.78%), and a significant portion received chemotherapy (52.75%).

In the external validation cohort, 15 cases (13.39%) were in the < 65 years group, and 97 cases (86.61%) were in the ≥ 65 years group. Regarding T stage, 71 cases (63.39%) were in the Tis, Ta, T1 group, while 41 cases (36.61%) were in the T2-T4 group. Of the cohort, 35 cases (31.25%) did not undergo chemotherapy, while 77 cases (68.75%) received chemotherapy. In terms of surgery, 34 cases (30.36%) did not undergo surgery, and 78 cases (69.64%) did.

Univariate and multivariate logistic regression analyses were performed on the training cohort to assess each prognostic factor, as shown in Table 2 . The univariate analysis identified grade, T stage, N stage, radiotherapy, chemotherapy, and surgery as significant factors (P < 0.05). In the multivariate regression analysis, a maximum tumor diameter of ≥ 3 cm, N1-N3 stage, and the absence of surgery (P < 0.05) emerged as independent prognostic factors. These factors were subsequently included in the nomogram, as depicted in Figure 2 .

Three independent risk factors—maximum tumor diameter ≥ 3 cm, N1-N3 stage, and absence of surgery—identified through logistic regression analysis were used as the final predictors in this study. A clinical prediction model, developed using R software, was presented as a nomogram to predict DM in patients with BCa ( Figure 2 ). Surgery had the most significant contribution to prognosis, followed by N stage and tumor size. In the nomogram, each factor is aligned with a fractional axis, and a vertical line is drawn to obtain the corresponding score. The total score is derived by summing the scores of each factor on the total axis, reflecting its influence on the progression of DM in BCa.

ROC curve analysis showed that the optimal diagnostic cut-off point for the training group was 0.045, with a sensitivity of 66.70% and specificity of 69.00%. For the internal validation group, the cut-off point was 0.058, with sensitivity at 63.00% and specificity at 80.70%. The external validation group showed a cut-off point of 0.070, with sensitivity and specificity at 95.70% and 93.30%, respectively. Calibration curve and standard curve fitting for the training group were moderate, with an AUC of 0.732, indicating strong discriminative ability of the prediction model. The predicted calibration curve for the internal validation group closely matched the standard curve, indicating the nomogram’s robust predictive capacity. The AUC for the internal validation group was 0.750, and for the external validation group, it was 0.968 ( Figure 3 ).

The external validation cohort demonstrated a higher AUC (0.968) compared to the training (0.732) and internal validation (0.750) cohorts. This discrepancy is likely due to the distinct clinicopathological profile of the external cohort ( Table 1 ), which included a significantly larger proportion of non-muscle invasive tumors (Tis/Ta/T1: 63.39% vs. 5.84% in SEER), a higher rate of DM (M1: 17.86% vs. 3.89% in SEER), and a substantially lower surgery rate (non-surgical: 30.36% vs. 0.25% in SEER).

Given that the absence of surgery and advanced N stage were the strongest predictors in our nomogram, the higher concentration of these high-risk features in the external cohort likely enhanced the model’s discriminative power. Although the sample size was limited (n = 112), the rigorous inclusion criteria ensured data quality. These differences reflect real-world patient allocation and provide a robust test of the model’s performance across diverse settings.

The predictive performance of our model for DM in BCa was compared with several recently published models developed using SEER data. As summarized in Table 3 , the internally validated C-index for our model, incorporating the predictors tumor size ≥ 3 cm, nodal stage N1-N3, and non-receipt of surgery, was 0.750. This performance is comparable to, or exceeds, the predictive accuracy of several SEER-based studies on BCa DM, which reported C-indices/AUCs ranging from 0.688 to 0.722 ( Table 3 ).

Notably, when externally validated using our independent, prospectively collected cohort, the predictive power of our model significantly improved, achieving an outstanding C-index of 0.968. This result far surpasses the performance metrics of all other SEER-based studies, including the highest previously reported AUC of 0.877. This comparative analysis highlights the robust discriminatory power of our model, especially its exceptional performance in the external validation cohort, suggesting strong generalizability beyond the derivation cohort.

PCA demonstrated excellent data quality across datasets, with corrected biological factors explaining more than 80% of the variance ( Figure 4A ). Batch effect analysis identified 253 differentially expressed genes in BCa that met the criteria of logFC > 1 and P < 0.05 ( Figure 4B ). Integrated machine learning approaches, including LASSO regression and Random Forest algorithms, identified 15 and 6 potential BCa biomarkers, respectively, with SRPX and ADH1B emerging as consensus candidates from both methods ( Figures 4D, E ). Comparative expression analysis revealed that both SRPX and ADH1B were significantly downregulated in tumor tissues compared to normal bladder tissues (P < 0.05) across both training and validation cohorts, suggesting their potential protective roles in BCa pathogenesis ( Figure 4C ). ROC curve analysis demonstrated robust diagnostic performance, with AUC values exceeding 0.7 for both biomarkers in all cohorts ( Figures 4F, G ). These findings position SRPX and ADH1B as promising diagnostic biomarkers with potential clinical utility for BCa prognosis prediction. Notably, ADH1B exhibited particularly strong performance in the validation cohort (AUC = 0.983 in the TCGA dataset), highlighting its potential as a superior biomarker candidate.

GO enrichment analysis of ADH1B-associated genes in BCa revealed significant functional associations across biological processes, cellular components, and molecular functions. The most prominent findings were centered around muscle system-related activities, with strong enrichment for muscle contraction, actomyosin structure organization, and smooth muscle contraction ( Figure 5 ). Cellular component analysis showed notable associations with contractile fibers, myosin filaments, and focal adhesions, while molecular function analysis revealed ADH1B’s involvement in calmodulin binding and actin binding. ADH1B may influence BCa progression through modulation of muscle-related functions and extracellular matrix organization, potentially affecting tumor stiffness and metastatic potential.

KEGG pathway enrichment analysis of ADH1B-associated genes revealed significant involvement in several key biological pathways related to the BCa progression. The most significantly enriched pathway was cytoskeletal organization in muscle cells, with four genes (FHL1, MYH11, DES, ACTG2) showing strong associations. Vascular smooth muscle contraction and motor protein function also demonstrated significant enrichment, further supporting ADH1B’s potential role in muscle-related cellular processes ( Figure 5 ). Additionally, the regulation of the actin cytoskeleton pathway was notably represented, suggesting that ADH1B may influence cancer cell motility and metastasis.