Machine learning applications in the detection and treatment of esophageal cancer Authors: Fangyuan Long, Hongru Zhang, Shungeng Zhang, Zongfu Dong, Xupeng Huang, Ronghang Hu Published: 2026 ABSTRACT Esophageal cancer (EC) is a gastrointestinal malignancy associated with a poor prognosis worldwide. It is characterized by an insidious onset, marked tumor heterogeneity, and complex diagnostic and therapeutic pathways. Consequently, clinical management continues to face challenges in early screening, accurate diagnosis, and the assessment of treatment response and prognosis. In recent years, the rapid advancement of artificial intelligence (AI) and machine learning (ML) techniques in medical image analysis and multidimensional data modeling has provided novel technological approaches for the precision diagnosis and treatment of EC. This narrative review summarizes recent advances in the application of AI in EC, with a particular focus on endoscopic and imaging-assisted diagnosis, prediction of treatment response and prognostic assessment. Existing studies have reported promising diagnostic performance of deep learning–based endoscopic image analysis models. In some controlled or retrospective settings, their performance has approached that of expert readers in the detection of EC and precancerous lesions, although prospective validation remains limited. ML models integrating clinical information, radiomic features, and selected biomarkers have also shown promising performance in predicting treatment response and supporting risk stratification. However, most existing studies are based on retrospective, single-center, or regionally confined datasets and are subject to several limitations, including population heterogeneity, insufficient data standardization, limited external validation, and poor model interpretability, which collectively hinder clinical translation. Overall, AI holds great promise for the precision management of EC; however, its clinical translation still depends on multicenter collaboration, the establishment of high-quality datasets, prospective validation studies, and the parallel development of ethical and regulatory frameworks. Background Esophageal cancer (EC) is one of the most common malignancies of the gastrointestinal tract worldwide. According to global statistics, it ranks as the 11th most frequently diagnosed cancer and the seventh leading cause of cancer-related mortality worldwide [1]. In recent years, with improvements in public health conditions and better control of major risk factors, the overall incidence of EC has declined in some regions. However, in high-incidence areas such as China, the disease continues to exhibit high incidence and mortality rates, and newly diagnosed cases and cancer-related deaths still account for a substantial proportion of the global burden [2]. The major histopathological subtypes of EC are esophageal squamous cell carcinoma (ESCC) and esophageal adenocarcinoma, with ESCC being the predominant subtype in China. Because the clinical manifestations of early-stage esophageal cancer are often subtle or nonspecific, most patients are diagnosed at a locally advanced or advanced stage when obvious symptoms, such as dysphagia and retrosternal discomfort, appear. As a result, the optimal window for treatment is often missed, leading to a poor prognosis. Currently, neoadjuvant therapy followed by radical surgical resection has become the standard treatment strategy for patients with resectable locally advanced EC [3]. With the successful application of immune checkpoint inhibitors across multiple solid tumors, neoadjuvant chemotherapy combined with immunotherapy has increasingly become an important first-line treatment option for patients with locally advanced esophageal squamous cell carcinoma. Multiple randomized controlled trials and real-world studies have suggested that this treatment strategy may improve the R0 resection rate and the pathological complete response rate, and improve progression-free survival and overall survival (OS) to some extent. However, in clinical practice, there is substantial heterogeneity in different patients’ responses to neoadjuvant therapy. Some patients derive limited benefit or even experience disease progression, which may fail to improve outcomes but may also expose them to unnecessary treatment-related toxicities and delay the optimal timing of surgery. Therefore, accurately identifying patients who are likely to benefit before treatment or at an early stage of therapy remains a critical challenge for precision medicine in EC. Although histopathological examination remains the gold standard for diagnosis and treatment response assessment, its invasiveness and procedural complexity limit its utility for dynamic evaluation. In clinical practice, endoscopic and imaging modalities are commonly used for assessment. However, these approaches remain limited in detecting early-stage lesions and in differentiation between immunotherapy-related pseudoprogression and hyperprogression. Moreover, the results are highly dependent on operator experience, leading to substantial interobserver variability. Therefore, there is a need to develop more objective, stable, and reproducible auxiliary assessment techniques to improve diagnostic accuracy and assist therapeutic assessment. Against this background, AI, particularly ML techniques, has rapidly advanced in the medical field and shown broad application potential. By training on large-scale clinical, imaging, pathological, and molecular datasets, ML algorithms can automatically extract latent features and capture complex nonlinear relationships, thereby supporting applications in disease detection, classification, treatment response prediction, and prognostic assessment. Compared with traditional statistical models, ML approaches may offer advantages in handling high-dimensional data, multivariable interactions, and complex pattern recognition tasks. In recent years, with advances in computational power and the accumulation of large-scale medical data, ML has been increasingly applied to multiple aspects of EC research and clinical practice, including endoscopy-assisted detection, imaging analysis, biomarker discovery, and the prediction of treatment response and survival outcomes. Based on these considerations, a comprehensive narrative review of recent advances in ML applications for EC detection, treatment response prediction, prognostic assessment, and imaging analysis is important for understanding the current landscape, key challenges, and future directions of this field. This review aims to summarize and critically analyze the major applications of ML in esophageal cancer–related research, thereby providing a reference for future research and potential clinical translation. Methods This study was conducted as a narrative review. Relevant literature on ML applications in esophageal cancer was identified through searches of PubMed, Web of Science, and Embase. The search included studies published between January 1, 2010, and January 31, 2026, using keywords such as “esophageal cancer”, “machine learning”, “artificial intelligence”, “deep learning”, “radiomics”, and “endoscopy”. The search was restricted to English-language publications in peer-reviewed journals. In addition, the reference lists of included studies were manually screened to identify potentially relevant articles. Studies were selected based on their relevance to the topic, with an emphasis on original research investigating clinical applications of machine learning in detection, treatment response prediction, and prognostic assessment. Studies were excluded if they were conference abstracts, editorials, letters, duplicate publications, animal studies, or studies lacking sufficient methodological or outcome information. When overlapping datasets from the same institution or research group were identified, priority was given to studies with larger sample sizes, more comprehensive analyses, or more recent data. The screening and selection of studies were performed manually by the authors based on title, abstract, and full-text assessment. As a narrative review, this study was not intended to provide a systematic or exhaustive synthesis of all available evidence, and no formal systematic review protocol, meta-analysis, or risk-of-bias assessment was performed. Applications of ML in EC detection The clinical outcomes of patients with EC are closely related to disease stage; therefore, early and accurate detection is essential for improving prognosis. However, conventional diagnostic methods still have limitations in sensitivity, specificity, and repeatability, and their performance is highly dependent on operator experience. In this context, ML enables the integrated analysis of multimodal data and facilitates feature extraction, thereby providing a new approach to improving the reliability and accuracy of EC detection. As shown in Fig. 1, existing studies have applied ML models to multiple detection steps, including imaging, endoscopic image analysis, pathology, and the modeling of clinical and molecular features, with the aim of achieving accurate identification of EC and lymph node metastasis across different clinical settings. The following sections provide a structured overview of the research progress and current applications of ML in EC detection, organized according to the detection approaches described above. Fig. 1Machine learning-based detection of esophageal cancer Machine learning-based detection of esophageal cancer Imaging-based detection Imaging examinations play a fundamental role in the initial screening, detection, and staging of EC. Commonly used modalities include CT, MRI, PET/CT, and conventional barium esophagography. Non-contrast CT is commonly used for the initial imaging evaluation of patients with suspected EC. Previous studies have shown that focal or diffuse esophageal wall thickening on non-contrast CT can serve as an important imaging marker of EC, based on which clinicians may recommend further endoscopic examination. However, due to its limited ability to visualize lesion boundaries, the extent of tumor infiltration, and small lesions, diagnosis based on manual interpretation of non-contrast CT remains uncertain. He et al. developed an ML model based on esophageal wall thickness features that was able to delineate the esophageal wall on non-contrast chest CT images [4]. In a test set consisting of 414 images, the model achieved a sensitivity of 0.888 and a specificity of 0.909. Notably, the study further included a validation set comprising 52 previously missed EC cases and 48 normal cases. With the aid of this model, improvements were observed in the sensitivity, specificity, and accuracy of three radiologists. Lin et al. employed nnU-Net to delineate the esophageal wall and trained an ML model for EC detection, which achieved a sensitivity of 0.900, specificity of 0.880, and accuracy of 0.882. The study showed that with the aid of this model, the diagnostic performance of junior radiologists improved in retrospective evaluations and approached that of senior clinicians within the study setting [5]. Contrast-enhanced CT is the most widely used imaging modality for the evaluation of EC and can effectively demonstrate irregular esophageal wall thickening, focal or diffuse enhancement, lymph node enlargement, and invasion of surrounding tissues. In recent years, several studies have explored the application of ML models for the automated analysis of contrast-enhanced CT images. In 2021, Chen et al. and Takeuchi et al. independently developed DL models based on contrast-enhanced CT images for the automatic detection of EC. The former improved lesion detection accuracy by introducing the online hard example mining (OHEM) into the Faster R-CNN framework, while the latter employed a fine-tuned VGG16 convolutional neural network (CNN) and integrated the Grad-CAM method to improve the interpretability of the model’s decision-making process. Both studies suggested the potential of DL for automated EC detection [6, 7]. Subsequently, in 2023, Yasaka et al. further evaluated the potential clinical utility of a DL model based on contrast-enhanced CT images in real clinical reading settings. The results showed that the model maintained favorable detection performance and improved the diagnostic accuracy of less experienced readers [8]. Building on these studies, in 2024, Zhang et al. applied ML to the automatic segmentation of primary EC tumor volume on contrast-enhanced CT images, enabling stable identification of tumor extent and potentially providing a basis for subsequent staging assessment and treatment planning [9]. By providing both anatomical and metabolic information, PET/CT offers distinct advantages in the initial staging of EC and is commonly used to assess distant and lymph node metastases after diagnosis. However, manual interpretation of PET/CT remains limited in consistency, efficiency, and the detection of small metastatic lymph nodes, creating a need for the application of machine learning‑assisted analysis. Early PET/CT-based ML studies mainly focused on the automatic detection and quantitative analysis of primary tumors, including automatic segmentation of ESCC volume and prediction of tumor invasion depth (T2/T3 stage) [10, 11]. Building on this, recent studies have further extended to the automated identification and prediction of metastatic lymph nodes. Yuan et al. developed a predictive model based on PET/CT that integrated ML, radiomic features, and clinical variables to assess the risk of cervical lymph node metastasis in patients with ESCC. The model demonstrated favorable discriminative performance in internal validation, achieving an area under the receiver operating characteristic curve (AUC) of 0.955 and an accuracy of 0.951. However, these findings should be interpreted cautiously because the results were derived from retrospective data and internal validation only. These findings suggest potential clinical utility, potentially assisting in the assessment of lymph node involvement and potentially supporting preoperative decision-making, pending further validation [12]. Shinozaki et al. developed a multimodal ML model by integrating non-contrast CT, contrast-enhanced CT, and PET/CT for the diagnosis of lymph node metastasis. The results suggested that the multimodal ML model may provide improved diagnostic performance compared with the individual imaging modalities evaluated in the study. In retrospective analyses, its performance approached that of experienced radiologists. These findings indicate that integrating multimodal imaging with ML methods may have potential to improve the accuracy of lymph node staging in EC [13]. AI-assisted detection in endoscopy Currently, endoscopic examination combined with pathological biopsy remains the gold standard for EC diagnosis. Recently, AI-assisted endoscopic diagnosis has emerged as a promising technical approach for improving the detection of early-stage EC. A meta-analysis involving 28 studies systematically evaluated the performance of DL models in the endoscopic diagnosis of EC and high-grade dysplasia. The results demonstrated high pooled sensitivity, specificity, and AUC. The overall diagnostic accuracy was reported to be higher than that of conventional manual interpretation in several included studies, with the models showing promising generalization in external validation settings [14]. Building on this evidence, numerous primary studies have further investigated the application of AI across various endoscopic imaging modalities and clinical scenarios. Building on previous retrospective studies, Nakao et al. conducted a prospective randomized controlled trial (RCT) to evaluate the potential clinical utility of AI-assisted endoscopist detection of ESCC in real-world settings [15]. This study enrolled high-risk individuals undergoing screening or surveillance esophagogastroduodenoscopy. These participants were then randomly assigned to either the AI-assisted group or the conventional endoscopy group. The AI system was developed based on a DL architecture. It provided real-time alerts for suspicious lesions during white light imaging (WLI), narrow-band imaging (NBI), and iodine staining. Results demonstrated that AI assistance did not significantly improve the detection rate of ESCC among non-expert endoscopists (47% vs. 45%, P = 0.93). Similarly, no statistically significant differences were observed in the expert group or the overall population. Nevertheless, a numerical improvement in detection rates was observed across all subgroups in the AI-assisted group. These findings suggest that AI assistance may help improve diagnostic consistency and stability. In contrast to these findings, a multicenter study focusing on early neoplasia in Barrett’s esophagus (BE) developed and validated a computer-aided detection (CAD) system within real-world clinical settings [16]. Trained on a large-scale dataset of endoscopic images from 15 centers, the CAD system improved the detection sensitivity for neoplasia among general endoscopists at both image and video levels, without a significant decrease in specificity. Notably, in a benchmark test set containing a high proportion of subtle neoplasia, the diagnostic performance of the CAD system exceeded that of non-expert endoscopists but also approached the performance of international experts in retrospective benchmark testing, indicating that AI-assisted systems may have potential utility in specific lesion types and controlled clinical scenarios. On the other hand, research focusing on the limitations of AI-assisted endoscopic systems has increasingly gained attention. Chang et al. systematically compared the performance of various object detection models in WLI and introduced artifact labels to minimize non-lesion interference [17]. Results demonstrated that although the overall precision improved, the recall for early ESCC and dysplasia remained limited, with errors primarily stemming from missed detections rather than misclassification. This suggests that under conventional imaging conditions, relying solely on object detection frameworks remains insufficient to fully address the under-recognition of subtle early-stage lesions, which partially explains why no significant AI benefit was observed in prospective studies. In addition to conventional WLI and NBI, hyperspectral imaging (HSI) has gained increasing attention as an emerging endoscopic imaging technology in recent years. By acquiring tissue reflectance information across multiple continuous spectral bands, HSI enhances spectral differences between tissues without relying on staining, thereby providing a richer feature set to facilitate the identification of subtle lesions and early-stage cancer. Based on this principle, several studies have attempted to integrate HSI with AI models to enhance the automated recognition of esophageal diseases, particularly early-stage EC. Yang et al. directly utilized HSI data to train deep learning models and compared their diagnostic performance with conventional RGB-WLI or NBI images [18]. Results showed that models trained on HSI data achieved relatively improved performance compared with models using conventional endoscopic images in terms of accuracy, precision, and recall, with this advantage being particularly evident in early-stage EC prediction tasks. Recent research has proposed the spectral-assisted visual enhancer (SAVE) technology, which converts conventional WLI images into pseudo-hyperspectral images without altering endoscopic hardware, thereby enhancing the spectral differences between tissues. In a study involving approximately 2,400 endoscopic images, researchers compared the performance of various ML and DL models across different imaging modalities. The results showed that the combination of SAVE and NBI demonstrated more favorable overall performance in disease classification tasks compared with WLI alone, demonstrating advantages in metrics such as precision and recall and suggesting that imaging modality significantly affects AI model performance [19]. Recent lesion detection research has demonstrated that in the recognition of normal tissue, dysplasia, and ESCC, SAVE-based models demonstrated comparatively favorable detection performance compared to NBI across various object detection frameworks, with a more pronounced advantage particularly in identifying early-stage or poorly defined lesions [20]. These studies demonstrate that the integration of HSI and AI technologies may contribute to improving the detection of early-stage ESCC within existing endoscopic imaging systems, providing a new technical direction for computer-aided diagnosis. However, it should be noted that most currently available endoscopic AI studies are based on ESCC cohorts from East Asian populations, and their applicability to EAC or Western screening settings remains uncertain. Beyond spectral enhancement, other studies have attempted to achieve the “virtualization” of traditional chromoendoscopy through DL. Toya et al. developed a virtual Lugol chromoendoscopy (V-LCE) system based on CycleGAN, which converts WLI images into simulated Lugol staining effects to facilitate the recognition of superficial ESCC [21]. Trained on paired WLI and authentic chromoendoscopy images from 170 patients, the model demonstrated in independent testing that its performance in lesion detection and boundary visualization was better than that of conventional WLI, although its overall efficacy remained lower than that of actual Lugol chromoendoscopy. Notably, V-LCE provides a diagnostic performance intermediate between WLI and actual chromoendoscopy without increasing procedural time or the risks and discomfort associated with Lugol staining, suggesting that generative DL-driven “virtual enhanced imaging” may serve as a supplementary detection tool for early-stage ESCC. Pathological and recognition of early-stage lesions Pathological examination serves as the gold standard for the diagnosis and grading of EC, and plays an indispensable role in the recognition of early-stage lesions, histological typing, and risk stratification. However, conventional pathological diagnosis relies on manual assessment by pathologists, and is not only time-consuming but also prone to false-negative results when lesion volumes are extremely small. Recently, non-destructive 3D pathology has demonstrated significant advantages in enhancing the sensitivity of early-stage lesion detection by providing holistic visualization of intact biopsy tissues. However, the large datasets generated also pose new challenges for manual image review. Barner et al. introduced an intelligent DL-based triage strategy to identify the most diagnostically valuable 2D slices from 3D pathological data, which significantly reduced the workload of pathologists while enhancing diagnostic sensitivity, providing a new technical paradigm for the deep integration of pathology and AI [22]. Furthermore, with the advancement of digital pathology and whole slide imaging (WSI) technologies, DL has also been introduced for the automated analysis and precise segmentation of EC pathological images. U-Net-based segmentation models have demonstrated promising segmentation performance and encouraging external generalizability in recognizing mucosal and tumor regions, generating clear ROI masks even with limited sample sizes. These models may assist in pathological grading and slide screening and may support clinical decision-making, thereby further extending the application of AI in EC pathology [23]. The application of digital pathology and DL has expanded beyond simple morphological recognition to the prediction of tumor molecular characteristics and biological behavior. The OncoMet framework proposed by Aalam et al., based on WSI, showed promising performance in predicting metastatic risk in patients with EC while simultaneously inferring the activation status of multiple key signaling pathways [24]. The study reported AUC values of 0.92 and 0.84 for internal and external validation, respectively, suggesting that DL models based on pathological images may provide potentially useful information for prognostic assessment and therapeutic decision support in EC without requiring additional molecular testing. Detection models based on clinical and molecular data Clinical information and molecular data also play vital roles in the early detection, metastatic risk assessment, and individualized management of EC. Traditional risk assessment methods based on single indicators struggle to fully characterize the complex biological processes of tumor development, when dealing with complex genomic data. In contrast, ML technologies can efficiently integrate multidimensional and heterogeneous data, providing new insights for constructing more robust detection and prediction models. In recent years, relevant research has primarily focused on liquid biopsy data, key gene or molecular feature selection, and the integration of clinical variables. Liquid biopsy has gained extensive attention in cancer diagnosis and treatment due to its minimally invasive nature and reproducibility. Qiao et al. developed an ML-based framework for the early detection of ESCC patients based on plasma cell-free DNA (cfDNA) methylation signatures. By identifying tumor-specific differentially methylated regions, the model was able to distinguish ESCC from healthy individuals and benign esophageal diseases [25]. The model achieved a sensitivity of approximately 75% and a specificity of 96% in an independent validation cohort, while its detection performance remained relatively stable across different age, sex, and pathological subgroups. These results suggest that cfDNA methylation combined with ML methods may serve as a non-invasive tool for early screening and risk stratification of EC. As research has progressed, the advantages of ML in biomarker screening and integrated analysis have gradually become evident. Through strategies such as feature selection, dimensionality reduction, and model training, researchers can identify representative genes, methylation sites, microbial DNA, or autoantibody combinations from high-dimensional and noisy molecular data. These features can then be used to construct relatively stable prediction models, thereby improving model interpretability and potential clinical usability while maintaining favorable diagnostic performance [26–32]. Previous studies have applied various ML algorithms (such as LASSO, SVM, and XGBoost) to analyze circulating microbial DNA, serum glycopeptide profiles, and autoantibody profiles, reporting promising diagnostic performance in the early detection and risk stratification of EC and its precancerous lesions. These findings suggest that ML-driven biomarker discovery not only helps overcome the limited stability of single indicators but also provides methodological support for multi-omics data integration and the construction of individualized detection models. On the other hand, ML models based on clinical data have also been utilized to assess the risk of lymph node or distant metastasis in patients with EC. By integrating routine clinical variables such as age, tumor location, stage, and laboratory indicators, these models facilitate the early identification of high-risk populations [33]. Overall, ML models based on clinical and molecular data provide a complementary perspective to imaging and morphological analysis for EC detection, helping to reveal the biological heterogeneity of tumors at a systemic level. However, such studies still face challenges such as limited sample sizes, data imbalance, and insufficient external validation; their clinical applicability and real-world utility remain to be validated in prospective multicenter studies. Application of ML in the treatment and prognostic assessment of EC In the clinical management of EC, precise evaluation of treatment response and long-term prognosis is of critical importance for optimizing therapeutic strategies and improving patient survival outcomes. However, traditional assessment methods based on TNM staging, pathological features, and single clinical indicators struggle to fully reflect tumor biological heterogeneity and individual variation. With the advancement of ML technologies, researchers have begun to integrate imaging, pathology, molecular omics, and clinical data to construct models for predicting therapeutic efficacy, assessing prognostic risks, and supporting individualized clinical decision-making. Although an increasing number of studies have reported promising performance, the current evidence remains largely based on retrospective analyses with heterogeneous study designs and limited external validation, which may introduce bias and limit the generalizability of the findings. Such approaches not only may help identify patients who are more likely to benefit before or during early treatment stages but also provide new technical pathways for dynamic risk stratification and treatment adjustment. Consequently, this section reviews the research progress of ML in treatment response prediction, prognosis and survival assessment, and individualized treatment decision support. Prediction of treatment response In recent years, neoadjuvant therapy has become a key component of the comprehensive treatment system for locally advanced EC, with therapeutic strategies rapidly evolving from traditional nCRT to neoadjuvant immunotherapy combined with chemotherapy (nICT). Several recent studies have suggested that nICT may offer advantages in major pathological response (MPR) and pathological complete response (pCR) compared with nCRT. However, the therapeutic response to neoadjuvant treatment still exhibits significant individual heterogeneity, and the high variability in clinical outcomes makes reliance on traditional staging or empirical decision-making insufficient for precision medicine. This highlights the importance of establishing reliable efficacy prediction tools before or during the early stages of treatment. In this context, ML methods provide a new technical pathway for the precise prediction of neoadjuvant therapy response by performing high-dimensional feature extraction and pattern recognition on multimodal data, including imaging, pathology, and endoscopic images. Specifically, the ability to predict pCR may have potential clinical significance: for patients with a high predicted probability of achieving pCR, future studies may explore the feasibility of non-surgical management strategies such as “watch-and-wait.” Such strategies may help reduce unnecessary surgical intervention and related complications while potentially contributing to organ preservation in carefully selected patients. Consequently, ML-based research is increasingly being explored as a potential tool for supporting individualized treatment decision-making in EC. Table 1 systematically summarizes representative studies of ML models based on imaging, pathology, and endoscopic images for predicting neoadjuvant therapy efficacy, providing a comprehensive framework and comparative basis for future research. Table 1Application of machine learning in predicting responses to neoadjuvant therapyCase, First author (year)Study designData sourceModelSample sizeValidation typeEndpointPerformanceZheng (2025) [34]Multicenter retrospectivePET/CTHybrid Deep Learning(DL) Architecturen = 831Internal + external validationtreatment decision classificationAUC = 0.9935Liu (2025) [35]Multicenter retrospectiveCTDELRN (Deep Ensemble)n = 485Internal validation (LOOCV)pathological complete response (pCR)AUC = 0.796(0.661–0.930)Hirata (2025) [36]Single-center retrospectiveDWI-MRIKNeighborsDist + BAGn = 70Internal cross-validationpCRAUC = 0.85(0.73–0.93)Li (2024) [37]Single-center retrospectiveCTSupport Vector Machine (SVM)n = 102Internal validationpCRAUC = 0.937(0.856–1.000)Liu (2024) [38]Multicenter retrospectiveCT + MRI(T2WI )+ DWIExtra Treesn = 151External validationpCRAUC = 0.868 (0.766–0.959)Yang (2025) [39]Single-center retrospectiveMRI(T2WI)Extra Treesn = 132Internal validationtumor regression grade (TRG)AUC = 0.932(0.883–0.981)Li (2024) [40]Multicenter retrospectiveContrast-enhanced CTRandom Forest(RF)n = 231Internal + external validationpCRAUC = 0.918(0.878–0.957)Wang (2023) [41]Single-center retrospectiveCT+ Clinical dataLogistic Regression (LR)n = 112Internal validation (LOOCV)pCRAUC = 0.891(0.823–0.950)Oda (2023) [42]Single-center retrospectiveContrast-enhanced CTLRn = 144Internal cross-validationTRAUC = 0.815(0.626-1.000)Hu (2021) [43]Multicenter retrospectiveCTResNet50-SVMn = 231Internal + external validationpCRAUC = 0.805(0.696–0.913)Qi (2025) [44]Multicenter retrospectiveContrast-enhanced CT + H&E ImageMultimodal Intermediate Fusion Model + XGBoostn = 335Internal + external validationpCRAUC = 0.97(0.94–0.99)Ling (2025) [45]Single-center retrospective and prospectiveCTXGBoostn = 157Internal + external validationmajor pathological response(MPR), pCRAUC = 0.915(0.844–0.981)Liu (2025) [46]Retrospective+ Prospective, single centerCT + MRIXGBoostn = 66Internal validation (LOOCV)pCRAUC = 0.961Ye (2025) [47]Multicenter retrospectiveCTXGBoostn = 60Internal validationMPRAUC = 0.870(0.706-1.000)Gu (2025) [48]Single-center retrospectiveCTDLn = 293Internal + external validationpCRAUC = 0.926(0.886–0.934)Zhang (2025) [49]Single-center retrospectiveContrast-enhanced CTLRn = 741Internal + external validationpCRAUC = 0.92(0.90–0.95)Wang (2024) [50]Multicenter retrospectiveContrast-enhanced CTVision-Mamban = 82Internal + external validationpCRAUC = 0.93(0.87–0.99)Qi (2024) [51]Single-center retrospectiveCTLRn = 126Internal validationpCRAUC = 0.852(0.824–0.876)Matsuda (2025) [52]Retrospective+ Prospective、single centerPET/CT + Contrast-enhanced CT + Clinical featuresSVMn = 1041Internal validationpCRAUC = 0.90(0.84–0.96)Yuan (2025) [53]Multicenter retrospectiveEndoscopic imageResNet-50n = 220Internal validationpCRAUC = 0.99(0.96–1.00)Yuan (2025) [54]Retrospective+ Prospective、single centerEndoscopic imageMultivariate LRn = 300Internal validationpCRAUC = 0.99(0.98–0.99)Kawahara (2022) [55]Single-center retrospectiveEndoscopic imageCNNn = 98Internal validationpCRAUC = 0.83Wang (2025) [56]Single-center retrospectiveEndoscopic imageCNNn = 114Internal validationpCRAUC = 0.913Qi (2025) [57]Single-center retrospectiveClinical features+Endoscopic ultrasound Blood biomarkersRFn = 223Internal + external validationpCRAUC = 0.89(0.75-1.00)Miura (2025) [58]Single-center retrospectiveCT + WSIResNet-50 + SVMn = 170Internal validationTRGAUC = 0.77(0.73–0.83)Xie (2025) [59]Single-center retrospectiveEndoscopic image + Clinical featuresResNet50n = 309Internal + external validationpCRAUC = 0.831(0.753–0.908)Sasagawa (2022) [60]Multicenter retrospectiveCTExtra Treesn = 121Internal cross-validationTreatment responseAUC = 0.73Imaging data represent one of the most widely utilized modalities in current treatment response prediction research, with CT, MRI, and PET/CT constituting the core data sources for model construction. Early research primarily focused on predicting pCR after nCRT, mostly based on radiomic features from CT or MRI, combined with traditional ML algorithms or DL frameworks to construct predictive models. Several studies have suggested that models based on contrast-enhanced CT or multiparametric MRI may achieve moderate-to-high discriminative performance (with AUCs typically ranging from 0.80 to 0.93). These models have demonstrated encouraging generalizability in a limited number of multicenter studies [35, 37–39]. For instance, Liu et al. conducted a multicenter retrospective study to construct a DL-radiomics model based on pre-treatment CT, T2, and DWI images for predicting pCR in ESCC patients following nCRT. The results demonstrated that the integrated model achieved an AUC of 0.868 in the independent test cohort, showing improved performance compared with single-modality imaging and models based solely on clinical features. This suggests that the fusion of multimodal imaging features can enhance predictive performance for neoadjuvant therapy response [38]. Table 1Application of machine learning in predicting responses to neoadjuvant therapyCase, First author (year)Study designData sourceModelSample sizeValidation typeEndpointPerformanceZheng (2025) [34]Multicenter retrospectivePET/CTHybrid Deep Learning(DL) Architecturen = 831Internal + external validationtreatment decision classificationAUC = 0.9935Liu (2025) [35]Multicenter retrospectiveCTDELRN (Deep Ensemble)n = 485Internal validation (LOOCV)pathological complete response (pCR)AUC = 0.796(0.661–0.930)Hirata (2025) [36]Single-center retrospectiveDWI-MRIKNeighborsDist + BAGn = 70Internal cross-validationpCRAUC = 0.85(0.73–0.93)Li (2024) [37]Single-center retrospectiveCTSupport Vector Machine (SVM)n = 102Internal validationpCRAUC = 0.937(0.856–1.000)Liu (2024) [38]Multicenter retrospectiveCT + MRI(T2WI )+ DWIExtra Treesn = 151External validationpCRAUC = 0.868 (0.766–0.959)Yang (2025) [39]Single-center retrospectiveMRI(T2WI)Extra Treesn = 132Internal validationtumor regression grade (TRG)AUC = 0.932(0.883–0.981)Li (2024) [40]Multicenter retrospectiveContrast-enhanced CTRandom Forest(RF)n = 231Internal + external validationpCRAUC = 0.918(0.878–0.957)Wang (2023) [41]Single-center retrospectiveCT+ Clinical dataLogistic Regression (LR)n = 112Internal validation (LOOCV)pCRAUC = 0.891(0.823–0.950)Oda (2023) [42]Single-center retrospectiveContrast-enhanced CTLRn = 144Internal cross-validationTRAUC = 0.815(0.626-1.000)Hu (2021) [43]Multicenter retrospectiveCTResNet50-SVMn = 231Internal + external validationpCRAUC = 0.805(0.696–0.913)Qi (2025) [44]Multicenter retrospectiveContrast-enhanced CT + H&E ImageMultimodal Intermediate Fusion Model + XGBoostn = 335Internal + external validationpCRAUC = 0.97(0.94–0.99)Ling (2025) [45]Single-center retrospective and prospectiveCTXGBoostn = 157Internal + external validationmajor pathological response(MPR), pCRAUC = 0.915(0.844–0.981)Liu (2025) [46]Retrospective+ Prospective, single centerCT + MRIXGBoostn = 66Internal validation (LOOCV)pCRAUC = 0.961Ye (2025) [47]Multicenter retrospectiveCTXGBoostn = 60Internal validationMPRAUC = 0.870(0.706-1.000)Gu (2025) [48]Single-center retrospectiveCTDLn = 293Internal + external validationpCRAUC = 0.926(0.886–0.934)Zhang (2025) [49]Single-center retrospectiveContrast-enhanced CTLRn = 741Internal + external validationpCRAUC = 0.92(0.90–0.95)Wang (2024) [50]Multicenter retrospectiveContrast-enhanced CTVision-Mamban = 82Internal + external validationpCRAUC = 0.93(0.87–0.99)Qi (2024) [51]Single-center retrospectiveCTLRn = 126Internal validationpCRAUC = 0.852(0.824–0.876)Matsuda (2025) [52]Retrospective+ Prospective、single centerPET/CT + Contrast-enhanced CT + Clinical featuresSVMn = 1041Internal validationpCRAUC = 0.90(0.84–0.96)Yuan (2025) [53]Multicenter retrospectiveEndoscopic imageResNet-50n = 220Internal validationpCRAUC = 0.99(0.96–1.00)Yuan (2025) [54]Retrospective+ Prospective、single centerEndoscopic imageMultivariate LRn = 300Internal validationpCRAUC = 0.99(0.98–0.99)Kawahara (2022) [55]Single-center retrospectiveEndoscopic imageCNNn = 98Internal validationpCRAUC = 0.83Wang (2025) [56]Single-center retrospectiveEndoscopic imageCNNn = 114Internal validationpCRAUC = 0.913Qi (2025) [57]Single-center retrospectiveClinical features+Endoscopic ultrasound Blood biomarkersRFn = 223Internal + external validationpCRAUC = 0.89(0.75-1.00)Miura (2025) [58]Single-center retrospectiveCT + WSIResNet-50 + SVMn = 170Internal validationTRGAUC = 0.77(0.73–0.83)Xie (2025) [59]Single-center retrospectiveEndoscopic image + Clinical featuresResNet50n = 309Internal + external validationpCRAUC = 0.831(0.753–0.908)Sasagawa (2022) [60]Multicenter retrospectiveCTExtra Treesn = 121Internal cross-validationTreatment responseAUC = 0.73Imaging data represent one of the most widely utilized modalities in current treatment response prediction research, with CT, MRI, and PET/CT constituting the core data sources for model construction. Early research primarily focused on predicting pCR after nCRT, mostly based on radiomic features from CT or MRI, combined with traditional ML algorithms or DL frameworks to construct predictive models. Several studies have suggested that models based on contrast-enhanced CT or multiparametric MRI may achieve moderate-to-high discriminative performance (with AUCs typically ranging from 0.80 to 0.93). These models have demonstrated encouraging generalizability in a limited number of multicenter studies [35, 37–39]. For instance, Liu et al. conducted a multicenter retrospective study to construct a DL-radiomics model based on pre-treatment CT, T2, and DWI images for predicting pCR in ESCC patients following nCRT. The results demonstrated that the integrated model achieved an AUC of 0.868 in the independent test cohort, showing improved performance compared with single-modality imaging and models based solely on clinical features. This suggests that the fusion of multimodal imaging features can enhance predictive performance for neoadjuvant therapy response [38]. Application of machine learning in predicting responses to neoadjuvant therapy Imaging data represent one of the most widely utilized modalities in current treatment response prediction research, with CT, MRI, and PET/CT constituting the core data sources for model construction. Early research primarily focused on predicting pCR after nCRT, mostly based on radiomic features from CT or MRI, combined with traditional ML algorithms or DL frameworks to construct predictive models. Several studies have suggested that models based on contrast-enhanced CT or multiparametric MRI may achieve moderate-to-high discriminative performance (with AUCs typically ranging from 0.80 to 0.93). These models have demonstrated encouraging generalizability in a limited number of multicenter studies [35, 37–39]. For instance, Liu et al. conducted a multicenter retrospective study to construct a DL-radiomics model based on pre-treatment CT, T2, and DWI images for predicting pCR in ESCC patients following nCRT. The results demonstrated that the integrated model achieved an AUC of 0.868 in the independent test cohort, showing improved performance compared with single-modality imaging and models based solely on clinical features. This suggests that the fusion of multimodal imaging features can enhance predictive performance for neoadjuvant therapy response [38]. With the evolution of treatment paradigms, research focus has gradually shifted toward the nICT population in recent years. Numerous studies have constructed ML or DL models based on contrast-enhanced CT to predict pCR or major pathological response (MPR) following nICT. Several studies have reported favorable predictive performance for nICT response prediction, with some multicenter studies reporting AUC values exceeding 0.90. However, these findings should be interpreted in the context of retrospective study designs and varying validation strategies [44, 48, 49]. In 2025, Zhang et al. constructed Vision-Mamba, a DL model integrating radiomic features, based on pre-treatment contrast-enhanced CT images from 741 patients in a multicenter study. The model reported AUC values of 0.92 and 0.84 for the internal and external validation cohorts, respectively, showing improved performance compared with traditional radiomics and conventional DL frameworks within this study [49]. The model demonstrated favorable performance in pCR prediction and showed potential utility for risk stratification, providing imaging-based information that may help identify patients more likely to benefit from treatment and support individualized management strategies. Other studies have introduced PET/CT or multimodal imaging features, combined with clinical variables for joint modeling. The results suggest that these models may provide improved predictive performance compared with single-modality models for assessing immunotherapy-related response [51]. These findings suggest that imaging features can, to a certain extent, reflect the variations in biological responses of tumors to chemoradiotherapy and immunotherapy. In addition to imaging data, pathological information, particularly digital whole-slide WSI, has been increasingly incorporated into treatment response prediction models in recent years. Compared to single-modality imaging, multimodal models may provide a more comprehensive characterization of tumor heterogeneity by fusing imaging and pathological features. Qi et al. conducted a multicenter study that performed intermediate-level fusion of contrast-enhanced CT and H&E-stained WSI features. Combined with the XGBoost model to predict pCR following nICT, this approach showed improved predictive performance in both internal and external validation cohorts [44]. Similarly, other studies have attempted to jointly model CT with WSI or MRI features for evaluating the efficacy of immunotherapy combination treatments, demonstrating more stable predictive performance compared to single-modality models [57]. These studies suggest that the architectural and microenvironmental information embedded in pathological images may serve as an important complement to imaging features, providing a more biologically meaningful basis for the prediction of neoadjuvant therapy response. Endoscopic images can directly reflect post-treatment changes in tumor surface morphology and mucosal structures, potentially offering complementary advantages for neoadjuvant therapy response assessment. In recent years, several studies have constructed DL models based on WLI to preoperatively predict pCR or major pathological response (MPR) in ESCC patients after neoadjuvant therapy. These studies have reported promising discriminative performance of ML models based on endoscopic images for pCR prediction. In some retrospective studies, their performance in patient-level prediction was reported to exceed that of traditional endoscopic biopsy results, and the predicted response categories were associated with long-term survival outcomes. This suggests that endoscopy-based AI methods hold potential to provide supportive information for future exploration of “watch-and-wait” strategies and organ preservation strategies after neoadjuvant therapy [52, 54]. Beyond the direct utilization of endoscopic images, studies have also extracted quantitative endoscopic features or integrated endoscopic ultrasound (EUS) and clinical indicators to construct predictive models, which have demonstrated promising predictive performance in nCT or nICT populations [53, 56]. These findings suggest that endoscopy-related information may contribute to response assessment and may potentially assist pre-treatment decision support. In addition, most treatment response prediction studies included predominantly ESCC patients from East Asian cohorts, which may limit the generalizability of these findings to EAC populations commonly observed in Western countries. Prognostic assessment and survival prediction In the clinical management of EC, accurate assessment of long-term prognosis and survival risk is essential for determining treatment intensity, optimizing follow-up strategies, and implementing risk-stratified management. However, traditional prognostic assessment systems, which rely on TNM staging, pathological grading, and single clinical indicators, may not fully capture tumor biological heterogeneity and inter-patient variability. In recent years, the increasing application of ML methods in medical data analysis has enabled researchers to integrate multidimensional clinical and imaging data for more refined prognostic assessment. Against this background, various ML models have been developed to predict OS and prognostic risk in EC. Table 2 summarizes representative studies in recent years. Table 2Applications of machine learning in prognostic assessment and survival predictionCase, First author (year)Study designData sourceModelSample sizeValidation typeEndpointPerformanceZhang (2023) [61]Single-center retrospectiveClinical featuresDeepSurvn = 6020Internal validationOverall survival (OS)AUC = 0.805C-index = 0.732(0.714–0.750)Xu (2025) [62]Single-center retrospective cohort studyClinical featuresDeepSurvn = 5276Internal validationOSCI = 0.83Tuersun (2025) [63]Multicenter retrospectiveClinical featuresNeural Multi-Task Logistic Regressionn = 6002Internal + external validationOS1 years AUC = 0.9533 years AUC = 0.9105 years AUC = 0.866Sheng (2024) [64]Single-center retrospectiveClinical featuresXGBoost; SVMn = 1897Internal validationEarly mortalityXGBoost: AUC = 0.822(0.792–0.852)Sensitivity = 0.667Specificity = 0.895SVM: AUC = 0.824, Sensitivity = 0.684Specificity = 0.882Xie (2025) [65]Multicenter retrospectivePET/CT + Clinical features3D ResNet10n = 331Internal + external validationImmunotherapy survival stratification (PFS/OS)AUC = 0.882(0.792–0.971)Yu (2025) [66]Single-center retrospectiveTranscriptome dataRandom Survival Forest(RSF)n = 323Internal + external validationOS400 days AUC = 0.971800 days AUC = 0.8631200 days AUC = 0.857Huang (2022) [67]Multicenter retrospectiveClinical featuresDeepSurvn = 9069Internal validationOSC-index = 0.753Mutevelizade (2025) [68]Single-center retrospectivePET/CTXGBoostn = 39Internal validationMortality, treatment response, lymph node metastasisLymph Node Metastasis Prediction AUC = 0.92Prediction of treatment response AUC = 0.90Pathological Subtype Prediction AUC = 0.95Guo (2025) [69]Single-center retrospective3D CTDL model integrating NFAM and GSIM modulesn = 352Internal + external validationOSC-index = 0.725AUC = 0.692Liu (2025) [70]Single-center retrospectiveCT+ Clinical featuresLogistic Regression(LR); SVMn = 526Internal validation3-year recurrence and recurrence patternsLRAUC = 0.830SVMAUC = 0.825Huang (2025) [71]Multicenter retrospectiveCT+ Clinical featuresCNN-Transformern = 482Internal + external validationPFS, OSAUC = 0.741Yu (2025) [72]Multicenter retrospectiveContrast-enhanced CT + Clinical featuresSVMn = 264Internal + external validationLocoregional recurrence (LRR)AUC = 0.744(0.7437–0.7443)Kasai (2024) [73]Single-center retrospectiveCTRFn = 44Internal validationPFS, OSAUC = 0.92(0.71–0.99)Cui (2022) [74]Single-center retrospectiveContrast-enhanced CT + Clinical featuresLASSO Coxn = 204Internal validationPFS, OSPFS prediction model AUC = 0.833C-index = 0.79OS prediction model AUC = 0.768C-index = 0.71Liu (2025) [75]Single-center retrospectiveContrast-enhanced CT+ blood inflammatory markers + Clinical featuresLASSO Coxn = 105Internal validationPFS, OSAUC = 0.872(0.676-1.000)C-index = 0.864Kawahara (2024) [76]Multicenter retrospectiveCT + PET/CT + Clinical featuresLASSO Coxn = 142Internal + external validationOSAUC = 0.93(0.840-1.000)C-index = 0.92Wang (2025) [77]Multicenter retrospectiveClinical features + Clinical featuresXGBoostn = 323Internal + external validationOSInternal validation set AUC = 0.794(0.720–0.860)External validation set AUC = 0.689(0.61–0.76)Xu (2025) [78]Single-center retrospectiveClinical featuresSVMn = 310Internal + external validationPostoperative recurrenceAUC = 0.847(0.768–0.927)Lu (2024) [79]Single-center retrospectiveClinical features + blood inflammatory markersRSFn = 2521Internal validationOS3 years AUC = 0.7615 years AUC = 0.771Zhang (2023) [80]Single-center retrospectiveClinical featuresElastic Net, Random Forest;n = 2441Internal validationOSC-index = 0.7311 years AUC = 0.7253 years AUC = 0.7205 years AUC = 0.752Liu (2025) [81]Single-center retrospectiveCT + Clinical featuresLR;n = 212Internal validationOSAUC = 0.910(0.880–0.940)Gujjuri (2023) [82]Multicenter retrospectiveClinical features + Pathological DataCox Proportional Hazards (CPH);n = 4719Internal + external validationOS, DFSAUC = 0.794(0.785–0.802)Xie (2024) [83]Multicenter retrospectivePathological Data + Clinical features + Tumor markersRF;n = 485Internal + external validationOSAUC = 0.975(0.962–0.987)Li (2025) [84]Multicenter retrospectiveGenomic features + Clinical featuresElastic Net;n = 100Internal + external validationOSC-index = 0.72 Table 2Applications of machine learning in prognostic assessment and survival predictionCase, First author (year)Study designData sourceModelSample sizeValidation typeEndpointPerformanceZhang (2023) [61]Single-center retrospectiveClinical featuresDeepSurvn = 6020Internal validationOverall survival (OS)AUC = 0.805C-index = 0.732(0.714–0.750)Xu (2025) [62]Single-center retrospective cohort studyClinical featuresDeepSurvn = 5276Internal validationOSCI = 0.83Tuersun (2025) [63]Multicenter retrospectiveClinical featuresNeural Multi-Task Logistic Regressionn = 6002Internal + external validationOS1 years AUC = 0.9533 years AUC = 0.9105 years AUC = 0.866Sheng (2024) [64]Single-center retrospectiveClinical featuresXGBoost; SVMn = 1897Internal validationEarly mortalityXGBoost: AUC = 0.822(0.792–0.852)Sensitivity = 0.667Specificity = 0.895SVM: AUC = 0.824, Sensitivity = 0.684Specificity = 0.882Xie (2025) [65]Multicenter retrospectivePET/CT + Clinical features3D ResNet10n = 331Internal + external validationImmunotherapy survival stratification (PFS/OS)AUC = 0.882(0.792–0.971)Yu (2025) [66]Single-center retrospectiveTranscriptome dataRandom Survival Forest(RSF)n = 323Internal + external validationOS400 days AUC = 0.971800 days AUC = 0.8631200 days AUC = 0.857Huang (2022) [67]Multicenter retrospectiveClinical featuresDeepSurvn = 9069Internal validationOSC-index = 0.753Mutevelizade (2025) [68]Single-center retrospectivePET/CTXGBoostn = 39Internal validationMortality, treatment response, lymph node metastasisLymph Node Metastasis Prediction AUC = 0.92Prediction of treatment response AUC = 0.90Pathological Subtype Prediction AUC = 0.95Guo (2025) [69]Single-center retrospective3D CTDL model integrating NFAM and GSIM modulesn = 352Internal + external validationOSC-index = 0.725AUC = 0.692Liu (2025) [70]Single-center retrospectiveCT+ Clinical featuresLogistic Regression(LR); SVMn = 526Internal validation3-year recurrence and recurrence patternsLRAUC = 0.830SVMAUC = 0.825Huang (2025) [71]Multicenter retrospectiveCT+ Clinical featuresCNN-Transformern = 482Internal + external validationPFS, OSAUC = 0.741Yu (2025) [72]Multicenter retrospectiveContrast-enhanced CT + Clinical featuresSVMn = 264Internal + external validationLocoregional recurrence (LRR)AUC = 0.744(0.7437–0.7443)Kasai (2024) [73]Single-center retrospectiveCTRFn = 44Internal validationPFS, OSAUC = 0.92(0.71–0.99)Cui (2022) [74]Single-center retrospectiveContrast-enhanced CT + Clinical featuresLASSO Coxn = 204Internal validationPFS, OSPFS prediction model AUC = 0.833C-index = 0.79OS prediction model AUC = 0.768C-index = 0.71Liu (2025) [75]Single-center retrospectiveContrast-enhanced CT+ blood inflammatory markers + Clinical featuresLASSO Coxn = 105Internal validationPFS, OSAUC = 0.872(0.676-1.000)C-index = 0.864Kawahara (2024) [76]Multicenter retrospectiveCT + PET/CT + Clinical featuresLASSO Coxn = 142Internal + external validationOSAUC = 0.93(0.840-1.000)C-index = 0.92Wang (2025) [77]Multicenter retrospectiveClinical features + Clinical featuresXGBoostn = 323Internal + external validationOSInternal validation set AUC = 0.794(0.720–0.860)External validation set AUC = 0.689(0.61–0.76)Xu (2025) [78]Single-center retrospectiveClinical featuresSVMn = 310Internal + external validationPostoperative recurrenceAUC = 0.847(0.768–0.927)Lu (2024) [79]Single-center retrospectiveClinical features + blood inflammatory markersRSFn = 2521Internal validationOS3 years AUC = 0.7615 years AUC = 0.771Zhang (2023) [80]Single-center retrospectiveClinical featuresElastic Net, Random Forest;n = 2441Internal validationOSC-index = 0.7311 years AUC = 0.7253 years AUC = 0.7205 years AUC = 0.752Liu (2025) [81]Single-center retrospectiveCT + Clinical featuresLR;n = 212Internal validationOSAUC = 0.910(0.880–0.940)Gujjuri (2023) [82]Multicenter retrospectiveClinical features + Pathological DataCox Proportional Hazards (CPH);n = 4719Internal + external validationOS, DFSAUC = 0.794(0.785–0.802)Xie (2024) [83]Multicenter retrospectivePathological Data + Clinical features + Tumor markersRF;n = 485Internal + external validationOSAUC = 0.975(0.962–0.987)Li (2025) [84]Multicenter retrospectiveGenomic features + Clinical featuresElastic Net;n = 100Internal + external validationOSC-index = 0.72 Applications of machine learning in prognostic assessment and survival prediction AUC = 0.805 C-index = 0.732(0.714–0.750) 1 years AUC = 0.953 3 years AUC = 0.910 5 years AUC = 0.866 XGBoost: AUC = 0.822(0.792–0.852) Sensitivity = 0.667 Specificity = 0.895 SVM: AUC = 0.824, Sensitivity = 0.684 Specificity = 0.882 400 days AUC = 0.971 800 days AUC = 0.863 1200 days AUC = 0.857 Lymph Node Metastasis Prediction AUC = 0.92 Prediction of treatment response AUC = 0.90 Pathological Subtype Prediction AUC = 0.95 C-index = 0.725 AUC = 0.692 LR AUC = 0.830 SVM AUC = 0.825 PFS prediction model AUC = 0.833 C-index = 0.79 OS prediction model AUC = 0.768 C-index = 0.71 AUC = 0.872(0.676-1.000) C-index = 0.864 AUC = 0.93(0.840-1.000) C-index = 0.92 Internal validation set AUC = 0.794(0.720–0.860) External validation set AUC = 0.689(0.61–0.76) 3 years AUC = 0.761 5 years AUC = 0.771 C-index = 0.731 1 years AUC = 0.725 3 years AUC = 0.720 5 years AUC = 0.752 In existing research, ML models based on large-scale clinical databases constitute an important component of prognostic prediction studies. Several studies leveraging national cancer databases such as SEER, have integrated clinical variables including demographic characteristics, tumor stage, and pathological features to construct models for predicting OS or progression-free survival (PFS) [61–64, 67]. Compared with traditional Cox regression models, ML algorithms such as Random Forest, Support Vector Machines, and Gradient Boosting may offer advantages in capturing nonlinear relationships and high-order interactions among variables. These algorithms have shown promising predictive performance across multiple studies. Furthermore, several studies have reported consistent discriminative performance across different stages, population subgroups, and external validation cohorts. For instance, Huang et al. constructed a DL-based survival prediction network leveraging the SEER database. They compared the prognostic performance of this model with that of the traditional TNM staging system in over 9,000 EC patients, and subsequently performed external validation in an independent cohort [67]. The results showed that the DL model achieved a higher C-index for predicting cancer-specific survival than the TNM staging system in both the internal and external validation cohorts. Furthermore, the researchers explored potential treatment recommendation strategies based on model outputs; Kaplan–Meier analysis showed that patients receiving model-recommended treatments showed improved survival outcomes compared with those who did not within this retrospective analysis. These findings suggest that ML-based survival models may have utility for risk assessment and may provide supportive information for clinical decision support. The advantages of such database-derived models lie in their large sample sizes and long-term follow-up, which may better reflect the survival distribution characteristics of real-world populations, making them suitable for population-level risk stratification and prognostic assessment. In addition to clinical variables, some studies have further incorporated CT imaging or radiomic features into the construction of survival prediction models. Relevant studies primarily extract high-dimensional quantitative features from pre-treatment or postoperative CT images and jointly model them with clinical indicators. The resulting models are used to predict OS, PFS, or recurrence risk. Current findings suggest that models based solely on clinical variables generally provide moderate predictive performance. After incorporating imaging or radiomic features, the discriminative ability of the models may be improved, with most studies reporting an increase in C-index or AUC to approximately 0.74–0.85. Some studies have reported AUC or C-index values exceeding 0.9 in internal and external validation cohorts, although these findings should be interpreted cautiously because of differences in sample size and study design [65, 68–76]. For instance, Mingyu et al. conducted a single-center retrospective study involving 526 ESCC patients who underwent radical surgery. Based on preoperative contrast-enhanced CT radiomic features, they integrated clinicopathological variables such as tumor length, lymph node status, and differentiation grade to construct ML models for predicting 3-year postoperative recurrence risk and recurrence patterns [70]. The results demonstrated that the model achieved an AUC of 0.82–0.83 for recurrence prediction in both the training and validation cohorts. Furthermore, the model showed potential utility for postoperative risk stratification, thereby potentially helping to identify patient subgroups that may benefit from adjuvant therapy. Notably, with the increasing application of immunotherapy in unresectable ESCC, research has begun to explore joint modeling strategies combining imaging features with immune-related biomarkers. In 2025, Liu et al. included 105 unresectable ESCC patients receiving PD-1 inhibitors combined with concurrent chemoradiotherapy. Based on preoperative arterial-phase CT radiomic features and integrated with peripheral blood inflammatory markers, they constructed ML survival models to assess 2-year PFS [75]. Results demonstrated that the combined model, integrating clinical features, radiomics, and inflammatory markers, achieved an AUC of 0.834 and 0.872 in the training and validation cohorts, respectively. This performance compared favorably with models based solely on clinical or imaging features within the study. This suggests that the synergistic integration of imaging phenotypes and host immune-inflammatory status may provide additional prognostic information in the setting of immunotherapy combination treatments. In addition, other studies have expanded the feature dimensions of prognostic prediction at the molecular level. The study by Li et al. constructed ML models based on multi-omics genomic data to characterize the survival risk profiles of ESCC patients. The results suggested that multi-omics features may provide additional information for risk stratification and survival prediction [84]. However, such models require substantial sample sizes, standardized data processing, and relatively high testing costs, and currently remain primarily in the exploratory stage. Moreover, substantial heterogeneity exists across studies regarding histological subtype and geographic population distribution, with most imaging- and treatment-related studies primarily focusing on ESCC patients from East Asia. Overall, current applications of ML in the prognostic assessment of EC remain dominated by single-modality models and relatively simple multimodal approaches, with considerable variability in reported model performance across studies. Most models are constructed based on retrospective databases, making them susceptible to factors such as selection bias, heterogeneity in treatment strategies, and missing variables. Their generalizability and reproducibility across diverse populations and clinical settings still require further validation. Current challenges and future perspectives Current challenges A key limitation of current ML models in the detection and therapeutic assessment of EC lies in significant data heterogeneity and selection bias. Existing studies often rely on single-center or regionally confined retrospective databases to develop models or analyze data from national cancer registries such as the U.S. SEER database. Study populations are predominantly derived from East Asian cohorts, with only limited evidence from European populations, resulting in a strong focus on ESCC, which predominates in these regions. In contrast, esophageal adenocarcinoma is the more common subtype in Western countries and differs substantially in terms of etiology, tumor biology, and anatomical characteristics. Therefore, the generalizability of current ML models to EAC remains uncertain, and caution is warranted when extrapolating these findings to Western populations. Although several emerging studies have begun to investigate AI-assisted detection and prediction models for Barrett’s esophagus–related neoplasia and EAC, currently available datasets remain relatively limited compared with those for ESCC. Future studies should place greater emphasis on establishing multicenter EAC-dominant cohorts and conducting external validation in Western populations to improve the broader clinical applicability and generalizability of ML models. Substantial differences exist among studies in terms of case composition, staging distribution, treatment strategies, and follow-up endpoints. Such inconsistencies in population characteristics and clinical pathways often result in decreased predictive performance during external validation, thereby limiting the generalizability of models across different healthcare settings. Moreover, the lack of standardized protocols for imaging acquisition, radiomics feature extraction, and clinical variable definitions further exacerbates inconsistencies in data distribution, thereby affecting the reliability of prognostic predictions in real-world settings. Beyond population heterogeneity, insufficient data quality and standardization also constitute key factors limiting model stability. In clinical datasets, missing variables, incomplete follow-up, and inconsistent definitions of survival endpoints are common. Imaging and radiomics data are particularly susceptible to variations in scanning devices, imaging parameters, and operational protocols. Variations in feature selection methods, modeling workflows, and evaluation metrics across studies lead to a lack of comparability of model performance between different datasets. In studies involving multimodal data, the integration of imaging features, molecular information, and clinical indicators still faces challenges such as limited sample size and redundant information, which may lead to models demonstrating strong performance in training cohorts while exhibiting limited generalizability in external validation settings. These factors partially undermine the reproducibility and stability of ML models in EC detection and prognostic prediction. Currently, most ML models remain at the stage of retrospective analyses and internal or limited external validation, with a lack of prospective studies or randomized controlled trials systematically evaluating their clinical utility. Most studies primarily report statistical performance metrics such as AUC or C-index, while evidence supporting their integration into real-world clinical workflows and their impact on patient management remains limited. These limitations highlight the remaining gaps in both technological maturity and clinical translation. Importantly, high model performance reported in controlled or internally validated datasets should not be interpreted as evidence of immediate clinical utility or implementation readiness. Many studies are conducted under relatively homogeneous imaging protocols, selected patient populations, and retrospective study conditions, which may overestimate model generalizability in real-world settings. Although some models have demonstrated promising results in external validation cohorts, evidence regarding their impact on clinical decision-making, workflow efficiency, patient outcomes, and cost-effectiveness remains limited. Therefore, the transition of ML models from technical feasibility to clinically deployable tools still requires prospective multicenter validation and real-world implementation studies. Although DL models demonstrate high performance in prognostic prediction, their “black-box” nature limits clinicians’ understanding of and confidence in the predicted results. The lack of clear explanations for feature contributions makes it difficult for models to serve as reliable decision-support tools when faced with complex clinical scenarios or when predictions conflict with clinical experience. Meanwhile, variations in hardware infrastructure, computational resources, and digital maturity across healthcare institutions mean that the deploy ability and stability of some models in resource-limited settings remain to be evaluated. Collectively, these factors affect the acceptability and widespread adoption of ML models in real-world clinical settings. The development and application of ML models heavily rely on large-scale clinical, imaging, and omics datasets, which involve patient privacy and sensitive information. Strict data protection regulations, while safeguarding patient rights, also impose constraints on cross-institutional data sharing and multicenter collaboration. Furthermore, the absence of clear guidelines regarding liability for erroneous model predictions introduces additional regulatory and institutional barriers to clinical implementation. Future directions In the future, a key factor for achieving more reliable application of ML models in EC clinical management lies in establishing large-scale, collaborative datasets spanning multiple regions and populations. Through multicenter collaboration, integrating real-world data across different ethnicities, disease subtypes, and healthcare resource settings may help mitigate the population limitations and selection biases commonly observed in existing studies. Simultaneously, establishing standardized protocols for data acquisition, annotation, and quality control is of fundamental importance for enhancing the stability of model training and the reproducibility of results. Building on this, the introduction of semi-supervised or self-supervised learning strategies can partially reduce reliance on high-quality annotated data while fully leveraging long-term follow-up and real-world clinical datasets, thereby providing more robust support for prognostic prediction models. Beyond data-level improvements, ML development in the field of EC is gradually evolving from single-task models toward multimodal integration and applications embedded within clinical workflows. Recent studies have begun exploring the direct integration of AI technologies into screening, diagnostic, treatment and surgical workflows to address critical bottlenecks in real-world clinical settings. For example, DL-based endoscopic image super-resolution methods have been applied to expand the effective field of view of microendoscopes, enhancing screening efficiency without compromising diagnostic accuracy [85]. Studies focusing on treatment process management have also demonstrated that ML models may assist in identifying radiotherapy or chemoradiotherapy patients’ potential need for proactive supportive interventions, thereby enabling the early optimization of individualized treatment strategies [56]. Meanwhile, the emergence of surgical data science has enabled AI to participate in the workflow analysis of complex surgical procedures. In robot-assisted minimally invasive esophagectomy, models trained on multicenter video datasets have shown the ability to identify surgical phases and key procedural steps, demonstrating potential cross-center adaptability [86]. Although most of the aforementioned studies remain at exploratory or retrospective validation stages, they share the common characteristic of progressively approaching real-world clinical settings, emphasizing model feasibility across different device, operator practices, and institutional conditions. This trend suggests that future research should place greater emphasis on prospective, multicenter clinical validation, incorporating clinically relevant endpoints such as survival outcomes, quality of life, and workflow efficiency, to facilitate the translation of ML models from “technically feasible” toward clinically applicable tools. From a clinical implementation perspective, future research should place greater emphasis on how ML models can be effectively incorporated into existing diagnostic and treatment pathways. Rather than functioning as standalone tools, ML systems are more likely to be adopted as decision-support components embedded within routine clinical workflows. For example, in endoscopic screening, ML algorithms may be integrated into real-time imaging systems to provide lesion detection prompts and risk alerts during examination. In radiology, such models may assist in highlighting suspicious regions or generating quantitative risk estimates, thereby improving diagnostic consistency, particularly among less experienced clinicians. At the level of treatment planning and prognostic evaluation, ML outputs—such as predicted probabilities of treatment response or survival risk—may be incorporated into multidisciplinary team discussions as supplementary information to support clinical judgment. However, the practical implementation of these approaches remains challenging. Key barriers include the lack of prospective evidence demonstrating clinical benefit, limited model interpretability, heterogeneity in data acquisition across institutions, and the absence of standardized guidelines for integrating AI outputs into clinical decision-making processes. In addition, regulatory requirements, data privacy concerns, and issues related to clinical responsibility further complicate real-world adoption. Addressing these challenges will be essential for translating ML models from research settings into clinically meaningful applications. Conclusion Overall, ML techniques demonstrate significant potential in the screening, diagnosis, treatment assessment, and prognostic prediction of EC, with particularly rapid advancements in endoscopic image analysis, radiomics-based modeling, and risk stratification. Existing studies indicate that AI-based models have achieved performance approaching that of experienced clinical experts in controlled or retrospective settings across multiple tasks, and may serve as assistive tools to enhance diagnostic consistency and the granularity of risk assessment. However, a review of the current literature reveals that AI research in the field of EC remains predominantly based on retrospective analyses and single-center datasets, with model generalizability, clinical utility, and real-world applicability yet to be fully validated. Significant differences in population composition, data quality, feature selection, and endpoint definitions across studies limit the comparability and reproducibility of model performance. From a clinical translation perspective, although some endoscopy AI systems are gradually moving toward practical application, most prognostic prediction and treatment decision-support models remain at an early stage of technological maturity, lacking prospective studies and evidence-based support based on clinical outcomes. Therefore, the value of AI in EC management should be regarded as a supplement to clinical decision-making rather than a replacement. In the future, through multicenter and multi-ethnic data collaboration, standardized dataset development, and the gradual advancement of prospective clinical validation, AI may gradually transition from methodological research to clinically applicable tools, providing increasingly stable and sustainable support for EC risk stratification, individualized treatment, and long-term follow-up management.