Complex biological systems analysis and deep learning for prognostic prediction of esophageal squamous cell carcinoma

Summary

Esophageal squamous cell carcinoma (ESCC) prognosis remains poor, and traditional models often fail to capture complex nonlinear interactions between clinical and molecular features. We integrated transcriptomic data from public datasets and an independent clinical cohort to identify prognostic biomarkers. Using weighted gene co-expression network analysis (WGCNA) and Lasso-Cox regression, we identified 16 key genes to construct a deep learning-based survival model, DeepSurv, integrating clinical and genetic features. DeepSurv demonstrated superior predictive performance compared to conventional machine learning models in both internal and external validation cohorts. SHAP value analysis highlighted the contribution of specific genes, including FAM155B and CFHR4, alongside TNM stages. This study establishes a robust, multidimensional prognostic tool that enhances risk stratification and offers insights into the molecular mechanisms driving ESCC progression.

Highlights

• •WGCNA and Lasso-Cox identified 16 robust prognostic genes for ESCC
• •DeepSurv integrates clinical and transcriptomic data for precise risk prediction
• •The model outperforms traditional machine learning methods in multi-cohort validation
• •SHAP analysis highlights FAM155B and CFHR4 as key molecular drivers of prognosis

Cancer; Machine learning

Introduction

Esophageal squamous cell carcinoma (ESCC) accounts for more than 80% of all esophageal cancer cases worldwide.¹ Despite the availability of multiple therapeutic strategies—including surgery, radiotherapy, chemotherapy, and immunotherapy—the prognosis of ESCC remains poor, with a 5-year survival rate of only 20%.¹

In recent years, numerous studies have demonstrated the enormous potential of gene-based biomarkers in prognostic prediction. By identifying key driver genes and genetic alterations, clinicians can better estimate patient survival and tailor individualized treatment strategies. For example, Liu et al.² extracted prognostic gene signatures from complex biological data using bioinformatics approaches, identifying adverse prognostic genes and altered metabolic pathways in ESCC, thereby providing potential therapeutic targets. Similarly, Shao et al.³ applied weighted gene co-expression network analysis (WGCNA) to identify core prognostic genes in ESCC by clustering genes with similar expression patterns and correlating modules with clinical phenotypes, thereby systematically uncovering molecular determinants of tumor initiation and progression.

Despite these advances in identifying candidate genes, transforming gene expression alterations into clinically actionable prognostic tools remains a major challenge. Building predictive models that integrate multidimensional data—such as clinical variables and gene expression—may enable more accurate forecasting of disease progression and patient outcomes, thus providing stronger evidence for clinical decision-making. Traditional linear models have been widely used for constructing prognostic models, but they assume linear relationships between input features and outcomes, and are often sensitive to outliers, which limits their predictive accuracy and generalizability in complex biological datasets.^4,5,6

To address these limitations, Katzman et al.⁷ proposed DeepSurv, an innovative deep learning-based survival analysis framework. DeepSurv captures nonlinear interactions between risk factors and survival outcomes, and has demonstrated superior performance in survival prediction across various cancers.^8,9

Therefore, in this study, we combined complex biological systems analysis with deep learning strategies to identify prognostic genes associated with ESCC using publicly available transcriptomic datasets. We then integrated these gene signatures with clinical variables to construct survival prediction models. Furthermore, we performed comprehensive analyses of the selected genes to explore their potential clinical implications. Our goal was to accelerate the discovery of potential therapeutic targets and provide more accurate prognostic evaluation for patients with ESCC. To the best of our knowledge, this is the first study to apply the DeepSurv framework by integrating both transcriptomic features and clinical risk factors for prognostic prediction in ESCC.

Results

This study employed a multi-stage systems biology strategy to identify prognostic biomarkers for ESCC and to construct multiple machine learning models for survival prediction. The overall study design is illustrated in Figure 1. The general characteristics of the datasets used in this study are summarized in Table 1.

Figure 1.

Workflow diagram of the study

WGCNA, weighted correlation network analysis; GO, Gene Ontology; GSVA, gene set variation analysis.

Table 1.

Clinical characteristics of patients

Characteristics	GSE53625 cohort (N = 179)	TCGA cohort (N = 94)	External validation cohort (N = 30)
Sex (No. (%))
Male	146 (81.6)	80 (85.1)	27 (90.0)
Female	33 (18.4)	14 (14.9)	3 (10.0)
Age (mean (SD))	59.4 (9.02)	58.4 (10.34)	68.1 (5.34)
Lesion location (No. (%))
Lower	20 (11.2)	39 (41.5)	15 (50.0)
Median	97 (54.2)	50 (53.2)	13 (43.3)
Upper	62 (34.6)	5 (5.3)	2 (6.7)
Tumor grade (No. (%))
Well	49 (27.4)	16 (17.0)	1 (3.3)
Moderately	98 (54.7)	57 (60.6)	22 (73.3)
Poorly	32 (17.9)	21 (22.3)	7 (23.3)
pT stage (No. (%))
T1	12 (6.7)	8 (8.5)	4 (13.3)
T2	27 (15.1)	32 (34.0)	7 (23.3)
T3/T4	140 (78.3)	54 (57.5)	19 (63.3)
pN stage (No. (%))
N0	83 (46.4)	56 (59.6)	11 (36.7)
N1	62 (34.6)	29 (30.9)	13 (43.3)
N2	22 (12.3)	6 (6.4)	5 (16.7)
N3	12 (6.7)	3 (3.2)	1 (3.3)
Pathologic stage (No. (%))
I/II	87 (48.6)	63 (67.0)	11 (36.7)
III/IV	92 (51.4)	31 (33.0)	19 (63.3)

Identification of differentially expressed genes (DEGs)

Differential expression analysis between ESCC tumor and adjacent normal tissues yielded 2,225 differentially expressed genes (DEGs), including 981 upregulated and 1,244 downregulated genes. A volcano plot illustrated the overall distribution of DEGs (Figure 2A), and a heatmap displayed the top 10 upregulated and top 10 downregulated genes ranked by adjusted p-values (Figure 2B).

Figure 2.

Volcano plot and heatmap of differentially expressed genes (DEGs) from GSE53625

(A) Volcano plot identifies DEGs. The x axis represents the log2 fold change, and the y axis represents the -log10 false discovery rate (FDR). The vertical dashed lines represent the threshold of |log2FC| > 1, and the horizontal dashed line represents FDR <0.05.

(B) Heatmap of the top 10 upregulated and top 10 downregulated DEGs. The color scale from blue to red indicates the relative gene expression levels, where red represents high expression, and blue represents low expression.

Identification of key gene modules by WGCNA

WGCNA was performed on 165 ESCC samples after removing 14 outliers, using 3,374 genes with variance above the 75th percentile. A soft threshold power of β = 5 was chosen to ensure a scale-free network topology (Figure 3A). Genes were clustered into modules via dynamic tree cutting, followed by module merging (mergeCutHeight = 0.25, deepSplit = 2, minModuleSize = 30) (Figure 3B). A total of 15 modules were identified (Figure 3C; Table S1).

Figure 3.

Identification of key modules related to tumorigenesis through WGCNA

(A) Scatterplot of soft threshold power (left) and average connectivity (right) in the WGCNA network.

(B) Clustering dendrogram and module colors of the WGCNA network.

(C) Heatmap of the correlation between different module colors and tumor.

(D) Scatterplots show the relationship between gene significance (GS) and module membership (MM) in the MEbrown, MEturquoise, MEyellow, and MEblue modules.

(E) Venn diagram shows the intersection of tumor-related genes identified by the “limma” package and WGCNA.

Module-trait correlation analysis revealed that the brown, turquoise, yellow, and blue modules showed the strongest associations with tumor status (|correlation| > 0.5) (Figure 3D). Genes from these modules were intersected with DEGs identified by the “limma” package, yielding 1,446 tumor-related genes (Figure 3E).

Functional enrichment analysis further revealed 394 Gene Ontology (GO) terms (315 biological processes, 36 cellular components, and 43 molecular functions), including epidermis development, collagen-containing extracellular matrix, and actin binding (Figure S1A). KEGG pathway analysis identified 319 enriched pathways, such as “cytokine-cytokine receptor interaction,” “ECM-receptor interaction,” and “protein digestion and absorption” (Figure S1B), indicating their potential involvement in ESCC progression.

Selection of prognostic gene signatures

The 1,446 tumor-related genes were subjected to univariate Cox regression, which identified 29 genes significantly associated with overall survival (p < 0.01; Figure 4A). Representative genes with hazard ratio (HR) > 1 (suggesting increased risk) included KBTBD12, MAMDC2, ADGRB3, OBP2A, CST1 and CST2, while representative genes with HR < 1 (suggesting protective association) included FAM155B, NELL2, MAGEA11, MAL2, B3GNT3, SLCO1B3, IL36RN, LDHD, GRHL3, HPSE, HOOK1, CARD18, MAGEB5, POF1B, PNPLA1, CFHR4, PRR9 and TMPRSS11E.

Figure 4.

Screening of prognostic genes and construction of the risk score model

(A) Forest plot of 29 prognosis-related genes identified by univariate Cox regression analysis (p < 0.01). The squares represent the hazard ratio (HR), and the horizontal lines represent the 95% confidence interval (CI). HR > 1 indicates risk genes, while HR < 1 indicates protective genes.

(B) LASSO coefficient profiles of the 29 candidate genes. Each curve corresponds to a gene.

(C) Selection of the optimal parameter (lambda) in the LASSO model using 10-fold cross-validation. The vertical dotted lines represent the optimal lambda value and one standard error from the minimum.

(D) Distribution of risk scores for patients in the training cohort.

(E) Kaplan-Meier survival analysis of overall survival (OS) between high-risk (yellow) and low-risk (blue) groups. The p-value was calculated using the log rank test (p < 0.0001, n = 273).

These 29 genes were further subjected to Lasso regression with 10-fold cross-validation, yielding an optimal penalty parameter λ = 0.052917 (Figures 4B and 4C). Sixteen key genes were retained as predictors (Table S2). Risk scores based on these genes effectively stratified patients into high- and low-risk groups, with high-risk patients showing significantly higher mortality (Figure 4D). Kaplan-Meier analysis confirmed that low-risk patients had significantly better OS compared to high-risk patients (p < 0.01; Figure 4E).

Survival analysis of individual genes further demonstrated that expression levels of ADGRB3, ANGPTL7, B3GNT3, CFHR4, CST1, FAM155B, HOOK1, HPSE, KBTBD12, MAGEA11, MAL2, NELL2, POF1B, PRR9, and SLCO1B3 significantly distinguished patient outcomes (Figure S2).

Correlation analysis showed that these prognostic genes were significantly associated with tumor grade, and some were related to T stage and N stage (Figure S3A). Univariate Cox regression further identified age, N stage, TNM stage, tumor grade, and the composite risk score as significant predictors of OS (p < 0.05) (Figure S3B). Multivariate Cox regression demonstrated that the risk score was an independent prognostic factor (Figure S3C).

Gene set variation analysis (GSVA) revealed differences in enriched GO terms between high- and low-risk groups. In high-risk patients, pathways such as the “positive regulation of WNT signaling pathway” were upregulated, whereas processes such as the “regulation of water loss via skin” were downregulated (Figures S4A and S4B), suggesting that prognostic genes may influence outcomes via these pathways.

Construction of deep learning models for survival prediction

To construct prognostic models, we integrated five clinical features (age, T stage, N stage, TNM stage, and tumor grade) with the 16 gene signatures. Although T stage was not significantly correlated with OS in this study, it was included given its established role in clinical decision-making.

Among the tested models, DeepSurv achieved the highest predictive performance, with a C-index of 0.857 in the training cohort and 0.705 in the validation cohort (mean C-index = 0.781), outperforming RSF (mean C-index = 0.730), SSVM (mean C-index = 0.727), and CoxPH (mean C-index = 0.672) (Figure 5A). These values represent the average C-indices across the training and validation cohorts. Time-dependent ROC analysis further demonstrated the stable and superior discriminative ability of DeepSurv across follow-up time points (Figure 5B).

Figure 5.

Construction and validation of machine learning models

(A) Heatmap compares the concordance index (C-index) of DeepSurv with CoxPH, RSF, and SSVM models in both trainieng (n = 191) and validation (n = 82) cohorts. Higher C-index values (red) indicate better predictive accuracy.

(B) Time-dependent ROC curves show the area under the curve (AUC) of the DeepSurv model at different follow-up time points.

(C and D) Kaplan-Meier survival curves for risk stratification in the internal validation cohort (C) and the external validation cohort (D, n = 30). p-values were calculated using the log rank test. Shaded areas represent the 95% confidence intervals.

(E) SHapley Additive exPlanations (SHAP) summary plot visualizes feature importance. The y axis lists the features ranked by importance. Each dot represents a single patient. The color indicates the value of the feature (red = high expression/value, blue = low expression/value). The x axis (SHAP value) shows the impact on the model output; positive SHAP values indicate a higher risk of mortality, while negative values indicate a lower risk.

Risk stratification based on the DeepSurv-derived score effectively distinguished OS between high- and low-risk groups in both the internal (Figure 5C) and external validation cohorts (Figure 5D).

Model interpretability analysis using SHAP highlighted genes such as FAM155B, CFHR4, CST1, and PRR9, along with TNM and N stages, as the most influential predictors (Figure 5E). These features likely play pivotal roles in determining patient prognosis.

Discussion

In this study, we integrated clinical risk factors with transcriptomic features and applied the DeepSurv model to predict prognosis in patients with ESCC. Compared with traditional approaches, the DeepSurv model demonstrated superior predictive performance in both internal and external cohorts, achieving the highest concordance index (C-index). These findings highlight the potential of deep learning-based survival analysis to enable precise and individualized risk stratification in ESCC.

The primary innovation of our work lies in the deep integration of clinical variables with gene signatures derived from high-throughput data. Most previous prognostic models for ESCC relied either on clinical variables or single-layer molecular features, typically analyzed using Cox regression or other linear models.^10,11,12 Such approaches are constrained by the proportional hazards assumption and are unable to capture complex nonlinear interactions. By contrast, DeepSurv effectively models high-dimensional nonlinear interactions between clinical and gene features, thereby enhancing predictive stability and clinical applicability.¹³ Furthermore, the use of SHAP-based interpretation alleviates the “black box” limitation of deep learning models by quantifying the contribution of each feature.¹⁴

Several studies have attempted to predict ESCC prognosis using WGCNA, risk score models, or machine learning methods.^15,16,17,18 For example, Zhang et al.¹⁵ developed a machine learning model for survival prediction based on clinicopathological factors, while other studies have focused on ferroptosis-, cuproptosis-, or autophagy-related gene signatures.^16,17,18,19 However, these single-dimensional approaches often fail to fully reflect the biological complexity of ESCC progression, thereby limiting predictive accuracy. In contrast, our findings show that DeepSurv achieved superior risk stratification compared with RSF, SSVM, and CoxPH models, consistent with prior reports across other cancer types.^20,21 These results suggest that integrating multi-source data with advanced deep learning algorithms is critical for developing robust prognostic models.

Beyond predictive performance, our analysis also identified several genes closely linked to ESCC progression. For instance, CST1 has been shown to promote ESCC cell proliferation and invasion by enhancing mitochondrial complex I activity and activating the OXPHOS/MEK/ERK signaling axis.²² CFHR4, a member of the factor H/CFHR family, plays a key role in the tumor immune microenvironment, and its downregulation has been associated with poor prognosis in multiple malignancies.^23,24 FAM155B, as an auxiliary subunit of the NALCN channel complex, regulates NALCN localization at the cell membrane.²⁵ Since NALCN directly modulates epithelial cancer cell migration, FAM155B may indirectly influence metastasis.²⁶ Additionally, HOOK1 has been linked to prolonged survival through regulation of necroptosis, whereas high expression and hypomethylation of MAGEA11 are associated with poor outcomes.^27,28 Collectively, these genes are implicated in tumor cell migration, immune regulation, and metabolic reprogramming, underscoring the biological relevance of our model.

To facilitate clinical translation, this model holds significant potential for personalized medicine. By accurately stratifying patients into high- and low-risk groups, the DeepSurv-based model can assist clinicians in identifying high-risk individuals who may benefit from more aggressive follow-up schedules or adjuvant therapies. Conversely, low-risk patients might avoid overtreatment, thereby optimizing resource allocation and improving quality of life.

In conclusion, we developed and validated a DeepSurv-based prognostic model for ESCC by integrating 16 gene signatures with clinical risk factors. This model not only improved survival prediction accuracy but also revealed potential molecular mechanisms underlying disease progression. Our findings suggest that deep learning-based, multi-source integrative approaches hold promise for advancing personalized treatment strategies and identifying potential therapeutic targets in ESCC.

Limitations of the study

This study has several limitations. First, the sample size of the external validation cohort remained relatively small, which may introduce selection bias. However, despite this limitation, the results demonstrated strong consistency across both internal and external cohorts, reinforcing the model’s potential generalizability. Larger multicenter prospective studies are still required for further validation. Second, we only incorporated transcriptomic features, while proteomic, epigenetic, or multi-omics integration may provide additional prognostic insights that were not captured in this analysis. Future work may explore multi-omics integration and advanced architectures such as graph neural networks (GNNs) or Transformers to further improve predictive accuracy and interpretability.^29,30 Finally, translating the DeepSurv-based model into a clinically applicable decision-support tool represents an important direction for future research.

Resource availability

Lead contact

Further information and requests for resources should be directed to and will be fulfilled by the lead contact, Yongjian Chang (yongjianchang@seu.edu.cn).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• •Data: This paper analyzes existing, publicly available data. The TCGA-ESCA datasets were retrieved from the Genomic Data Commons (GDC) portal (https://portal.gdc.cancer.gov/). The independent clinical cohort data supporting the findings of this study have been deposited in Zenodo under the DOI: https://doi.org/10.5281/zenodo.18627126.
• •Code: All original code for the DeepSurv model and data analysis has been deposited at GitHub and is publicly available as of the date of publication. The DOI for the code is: https://github.com/ChngYJ/ESCC-Survival.
• •Additional information: Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Acknowledgments

This work received no specific funding from any funding agency in the public, commercial, or not-for-profit sectors. We thank Prof. Wenwu Yu (affiliated with Southeast University) for their valuable guidance on the study design and data analysis.

Author contributions

Conceived and designed the study: Y.C. Performed the experiments: Y.C. Analyzed the data: Y.C., W.Z. Wrote the paper: W.Z. All authors read and approved the final manuscript.

Declaration of interests

The authors declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Deposited data
https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE53265	GSE53625	GEO dataset
https://portal.gdc.cancer.gov	TCGA	TCGA dataset
https://doi.org/10.5281/zenodo.18627126	This paper	Processed transcriptomic data
Software and algorithms
DeepSurv model code	This paper	https://github.com/ChngYJ/ESCC-Survival
R (v4.4.1)	The R Foundation	https://cran.r-project.org/
Python (v 3.10.11)	Python Software Foundation	https://www.python.org/
Limma R package (v 3.56.2)	Bioconductor	https://bioconductor.org/packages/limma
WGCNA R package (v 1.72–5)	R CRAN	https://cran.r-project.org/web/packages/WGCNA/
ClusterProfiler R package (v 4.8.3)	Bioconductor	https://bioconductor.org/packages/clusterProfiler
GSVA R package (v 1.48.3)	Bioconductor	https://bioconductor.org/packages/GSVA
Survival R package (v 3.7–0)	R CRAN	https://cran.r-project.org/web/packages/survival/
Glmnet R package (v 4.1–8)	R CRAN	https://cran.r-project.org/web/packages/glmnet/
Survminer R package (v 0.4.9)	R CRAN	https://cran.r-project.org/web/packages/survminer/
Scikit-learn (v 1.0.2)	Python Software Foundation	https://scikit-learn.org/

Experimental model and study participant details

Human tumor samples

This study was designed as a retrospective analysis, and no new patients were recruited specifically for this research. Instead, paired tumor and adjacent non-tumor tissues were obtained from the archival biological sample bank of the Department of Pathology at Zhongda Hospital, Southeast University. These samples were originally collected from 30 patients with primary ESCC who underwent surgery between January 2020 and January 2023. The study was approved by the Ethics Committee of Zhongda Hospital, Southeast University (Approval No.2025ZDSYLL247-P01).

The external validation cohort included 27 male and 3 female patients (Table 1). Detailed clinical and demographic characteristics for each individual patient are provided in Table S3. The influence of sex on the results was not analyzed due to the limited sample size and the significant male predominance in ESCC prevalence.

The inclusion criteria were as follows: (1) patients histologically diagnosed with primary ESCC; (2) patients who underwent radical esophagectomy without preoperative chemotherapy or radiotherapy; and (3) availability of complete clinicopathological and follow-up data. Patients with other malignancies or incomplete medical records were excluded.

Method details

Data acquisition and preprocessing

The GSE53625 dataset, including overall survival (OS) and survival time (OS.time) information, was downloaded from the Gene Expression Omnibus (GEO) database.³¹ This dataset contains mRNA expression profiles and clinical data from 179 paired ESCC and adjacent normal tissues. Probe sequences were re-annotated against RefSeq coding sequences (www.ncbi.nlm.nih.gov/refseq/). RNA-seq data and clinical information for 94 ESCC tumor samples were obtained from The Cancer Genome Atlas (TCGA, https://portal.gdc.cancer.gov/). Both datasets were normalized and used for model training.

Samples lacking follow-up data were excluded. 179 ESCC samples from the GSE53625 dataset and 94 ESCC samples from the TCGA-ESCC dataset were finally included for model training. Missing clinical variables were imputed: binary factors were imputed using logistic regression, while categorical factors with more than two levels were imputed using proportional odds models or multivariate regression.

Differential expression analysis

Differentially expressed genes (DEGs) between tumor and adjacent normal tissues in GSE53625 were identified using the “limma” package (version 3.56.2).³² The thresholds were set as |log2 fold change| > 1, p < 0.05, and false discovery rate (FDR) < 0.05. Heatmaps and volcano plots were generated using “ggplot2” to visualize expression differences.

Weighted gene co-expression network analysis (WGCNA)

WGCNA was performed using the “WGCNA” R package (version 1.72-5).³³ After hierarchical clustering, 14 outlier samples were removed, leaving 172 ESCC cases for downstream analysis. Genes with low expression or poor sequencing quality were filtered out, and those with variance above the 75th percentile were retained. Based on the scale-free topology criterion, a soft threshold power β was selected to construct an adjacency matrix and transformed into a topological overlap matrix (TOM). Modules were identified via dynamic tree cutting, merged if similar, and correlated with clinical traits to identify modules most strongly associated with tumor status.

Genes from the most tumor-associated modules were intersected with DEGs from the “limma” analysis using a Venn diagram to obtain tumor-related candidate genes.

Functional enrichment analysis

Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses were conducted using the “ClusterProfiler” R package (version 4.8.3).^34,35 GO terms included biological process (BP), cellular component (CC), and molecular function (MF). Significance was set at adjusted p < 0.05.

Commercial reproduction rights for KEGG pathway diagrams were secured through a licensing agreement with Kanehisa Laboratories. All KEGG-derived content complies with the terms outlined at http://www.kegg.jp/kegg/legal.html.

Prognostic gene selection using Lasso-Cox regression

Candidate tumor-related genes were subjected to univariate Cox regression using the “survival” package (version 3.7–0). Genes significantly associated with OS (p < 0.01) were then reduced using least absolute shrinkage and selection operator (Lasso) regression with the “glmnet” package (version 4.1–8). This combined approach facilitated both prognostic power and feature selection. A risk score for each patient was calculated based on the expression levels of the final gene set. Patients were stratified into high- and low-risk groups according to the median risk score. Kaplan–Meier (K–M) survival curves were generated using the “survminer” package (version 0.4.9).

Gene Set Variation Analysis (GSVA)

GSVA was performed using the “GSVA” package (version 1.48.3) with gene sets “c2.cp.kegg.v7.4.symbols.gmt” and “h.all.v7.5.1.symbols.gmt” as references.³⁶ Differences between high- and low-risk groups were evaluated to identify biological pathways associated with survival outcomes.

Correlation with clinical factors

Associations between key prognostic genes and clinicopathological features were analyzed, and risk scores were further evaluated for their correlation with survival. Heatmaps were used for visualization. Risk scores and clinical features were further analyzed using univariate Cox regression to identify variables associated with OS.

Construction and validation of deep learning models

Deep learning–based models—including Cox proportional hazards (CoxPH), random survival forest (RSF), survival support vector machine (SSVM), and DeepSurv. Input features included five clinical risk factors (age, T stage, N stage, TNM stage, and tumor grade). While T stage was not identified as statistically significant in univariate analysis, it was included due to its established importance in clinical decision-making for ESCC. The selection of clinical features thus combined statistical evidence with clinical relevance.

Gene expression data from GSE53625 (n = 179) and TCGA-ESCC (n = 94) were separately normalized and then combined into a single dataset comprising 273 tumor samples. This integrated dataset was randomly split into a training set (70%, n = 191) and a test set (30%, n = 82). To ensure model robustness and minimize overfitting, we employed a rigorous internal validation strategy using repeated random shuffling and cross-validation on the integrated public datasets. Furthermore, external validation was performed using RNA-seq data derived from the independent clinical cohort (n = 30) to test the model’s generalizability in a real-world setting. Model performance was assessed by the concordance index (C-index). Patients were stratified into high- and low-risk groups based on the median risk score derived from the best-performing model. Survival differences between groups were evaluated using Kaplan–Meier (K–M) curves and log rank tests. Model interpretability was assessed using SHapley Additive exPlanations (SHAP), with results visualized in beeswarm plots.

Quantification and statistical analysis

All statistical analyses were performed using R software (version 4.4.1) and Python (version 3.10.11). Continuous variables were compared between two groups using the Wilcoxon rank-sum test. For survival analysis, Kaplan–Meier curves were generated to visualize overall survival, and differences were assessed using the log rank test. Univariate and multivariate Cox proportional hazards regression analyses were conducted to identify independent prognostic factors and to calculate hazard ratios (HRs) with 95% confidence intervals (CIs). The predictive performance of the prognostic models was evaluated using the C-index and ROC curves. Statistical significance was defined as a two-tailed p value <0.05.

Published: April 1, 2026