Recent advances in machine learning for precision diagnosis and treatment of esophageal disorders

Abstract

The complex pathophysiology and diverse manifestations of esophageal disorders pose challenges in clinical practice, particularly in achieving accurate early diagnosis and risk stratification. While traditional approaches rely heavily on subjective interpretations and variable expertise, machine learning (ML) has emerged as a transformative tool in healthcare. We conducted a comprehensive review of published literature on ML applications in esophageal diseases, analyzing technical approaches, validation methods, and clinical outcomes. ML demonstrates superior performance: In gastroesophageal reflux disease, ML models achieve 80%-90% accuracy in potential of hydrogen-impedance analysis and endoscopic grading; for Barrett’s esophagus, ML-based approaches show 88%-95% accuracy in invasive diagnostics and 77%-85% accuracy in non-invasive screening. In esophageal cancer, ML improves early detection and survival prediction by 6%-10% compared to traditional methods. Novel applications in achalasia and esophageal varices demonstrate promising results in automated diagnosis and risk stratification, with accuracy rates exceeding 85%. While challenges persist in data standardization, model interpretability, and clinical integration, emerging solutions in federated learning and explainable artificial intelligence offer promising pathways forward. The continued evolution of these technologies, coupled with rigorous validation and thoughtful implementation, may fundamentally transform our approach to esophageal disease management in the era of precision medicine.

Key Words: Esophageal disorders; Machine learning; Gastroesophageal reflux disease; Esophageal cancer; Barrett’s esophagus; Achalasia; Clinical decision support system

Core Tip: This review synthesizes machine learning (ML) applications in esophageal disorders, emphasizing three critical advances: (1) Automated analysis of multimodal diagnostic data achieving accuracy rates of 80%-95% across different conditions; (2) Integration of deep learning with endoscopic imaging enabling real-time assistance in diagnosis and risk stratification; and (3) Development of novel non-invasive screening approaches through ML-based biomarker identification. The convergence of artificial intelligence with clinical medicine demonstrates transformative potential in addressing current diagnostic challenges and enabling precision medicine in esophageal disease management.

INTRODUCTION

Esophageal disorders represent a significant global health challenge, affecting millions of people worldwide. These disorders comprise a diverse spectrum of conditions affecting the esophagus, ranging from common benign diseases to rare but aggressive ones[1]. The clinical manifestations are heterogeneous, encompassing gastroesophageal reflux disease (GERD) characterized by symptoms such as heartburn and regurgitation[2], esophageal carcinoma (EC) marked by the proliferation of malignant cells[3], inflammatory conditions such as esophageal varices (EV) and eosinophilic esophagitis[4], and mobility disorders such as achalasia[5] that impair swallowing function. Heartburn, dysphagia, and angina are prevalent symptoms across various esophageal disorders. These conditions significantly affect patients’ nutritional status, sleep patterns, and social interactions[6,7], creating substantial socioeconomic burdens through healthcare costs and reduced productivity.

The etiology of esophageal disorders involves complex interactions among multiple risk factors, such as smoking, excessive alcohol and obesity, which are associated with the development of GERD[8], EC[9], and Barrett’s esophagus (BE)[10]. Diets high in processed foods, red meat and saturated fat and low in fruits and vegetables are also linked to EC. In addition, advancing age increases the likelihood of developing most esophageal disorders, while factors such as family history, hiatal hernia, and liver cirrhosis also have significant impacts. Understanding these risk factors is essential for effective prevention strategies as well as early detection efforts.

Early detection and precise diagnosis of esophageal disorders are crucial for optimal outcomes. Timely intervention allows for more efficient and less invasive therapies for conditions such as GERD, BE, and certain esophageal malignancies. Moreover, precise diagnosis allows targeted therapeutic strategies, decreasing the probability of complications including erosive esophagitis, ulceration, strictures and gastrointestinal bleeding[11,12]. However, current diagnostic approaches face several challenges, including the reliance on subjective interpretation and variable accuracy depending on the clinician’s expertise.

The complexity of risk factor interactions, diverse clinical manifestations, and critical importance of early detection necessitate innovative approaches in esophageal disease management. As shown in Figure 1, the emergence of artificial intelligence (AI), particularly machine learning (ML) and deep learning, offers promising solutions to these diagnostic challenges. ML relies on algorithms that acquire knowledge from data by identifying patterns, which are then utilized to make predictions or conclusions; the human investigators are then able to leverage their own ability to analyze the complex data patterns and provide objective, reproducible results. Traditional ML algorithms, such as regression, K-nearest neighbors (KNN), decision trees, and support vector machines (SVMs), are derived from statistical principles[12]. For example, regression algorithms are used to make predictions about continuous outcomes, and KNN employs a proximity-based approach to classify data points by considering their distance to labeled data. These fundamental algorithms work as the basic components for complicated ensemble methods[13] such as random forests (RFs). The robust interpretability of traditional ML algorithms, which are based on statistical theory, makes them particularly suited for assessing clinical and laboratory data[14], making predictions and detecting risk factors in disease research.

Figure 1 Overview of machine learning applications in major esophageal disorders. CT: Computed tomography; HREM: High-resolution esophageal manometry; BE: Barrett’ s esophagus; EV: Esophageal varices; RE: Reflux esophagitis.

Deep learning algorithms, architected with multilayer neuron structures that imitate biological neural networks processing[15], excel at automatically learning and extracting high-level features from complex and unstructured data. Neural networks primarily consist of convolutional neural networks (CNNs)[16] and recurrent neural networks (RNNs)[17,18]. CNNs are frequently used for the purpose of visual image processing. For instance, Haug et al[19] achieved remarkable success with their deep residual learning framework (ResNet) in the ImageNet image recognition competition, demonstrating its robust feature extraction capabilities. RNNs, on the other hand, excel at analyzing sequential data such as text and audio[17]. The transformer model, which has been designed for text processing such as ChatGPT, has shown outstanding abilities in the field of natural language processing. Transformer models employ self-attention mechanisms to simultaneously scan all words in a text sequence and capture their interdependencies[18]. This makes them highly effective for evaluating medical texts and constructing disease knowledge graphs.

The convergence of advancing AI technology and increasing availability of high-quality medical data has catalyzed the integration of ML into clinical medicine[19-21], with initial efforts showing substantial promise in the diagnosis and management of esophageal disorders. The clinical application of ML models requires a robust workflow for data collection, processing, and modeling. We summarized an end-to-end pipeline for AI modeling (Figure 2), starting from the acquisition of diverse data, including medical imaging, clinical and patient-reported information, biomarker assays, and sequencing data, to clinical applications. These raw data are then subjected to distinct processing pipelines. Unstructured image data are enhanced using image augmentation techniques, while structured data undergo cleaning and standardization to prepare them for computational analysis, whereas multi-omics data are processed with mapping and annotations. Modeling then utilizes deep learning approaches for automatic image-based tasks and ML algorithms for pattern identification. Finally, these workflows culminate in diverse clinical applications, including automated lesion segmentation, tumor identification, diagnostic and prognostic modeling, and informed development of targeted therapy strategies.

Figure 2 Integrated data-drive modeling pipeline for esophageal disorders leveraging deep learning and machine learning for improved diagnosis and prognosis. CT: Computed tomography; NLP: Natural language processing; VCF: Variant call format; VGG: Visual geometry group network; YOLO: You only look once; UNet: U-shaped convolutional network; ResNet: Residual neural network; kNN: K-nearest neighbors; LR: Logistic regression; SVM: Support vector machine; RF: Random forest; DT: Decision tree; GERD: Gastroesophageal reflux disease; BE: Barrett’s esophagus; EC: Esophageal carcinoma; EV: Esophageal varices.

This review provides a comprehensive analysis of recent research progress in ML applications for precision diagnosis and treatment of esophageal disorders. We focus on major conditions including GERD, esophageal malignancies, and other common esophageal diseases, examining how ML technologies are transforming diagnostic accuracy, treatment planning, and outcome prediction. This review also explores the challenges and limitations of current ML applications in this specific medical domain, as well as potential future directions for research and development. By synthesizing these latest advancements, we seek to contribute to the ongoing efforts to improve patient outcomes and enhance clinical decision-making in this area of medicine.

AI-ENHANCED DIAGNOSIS FOR GERD

The global prevalence of GERD is rising, highlighting the increasing need for efficient diagnostic and treatment methods[7,8]. While not life-threatening, GERD significantly impacts quality of life and can lead to complications such as reflux esophagitis (RE) and BE due to chronic acid exposure[22]. The current diagnostic approach combines monitoring reflux events [e.g., potential of hydrogen (pH)-impedance monitoring[23], high-resolution esophageal manometry (HREM)], and observing esophageal lesions via esophageal endoscopy[24]. Nevertheless, the invasive nature of these procedures[25] and their intrinsic dependence on operator expertise[26] restrict their application in the clinic[26]. In response to the challenges, ML applications have emerged as promising solutions to enhance GERD diagnosis through two primary approaches: Analysis of pH-impedance monitoring data and automated grading of endoscopic images.

Enhancing GERD diagnosis with ML analysis of pH-impedance monitoring data

pH-impedance monitoring technology[23] provides comprehensive data and facilitates the correlation between reflux occurrences and symptoms, making it a golden standard for diagnosing GERD[27]. However, traditional automated software analyses are less than 65% accurate and have a high false-positive rate. Medical professionals need to carefully review the monitoring signals for reaching a diagnostic conclusion, consequently causing delays in diagnosis.

By contrast, ML-based approaches have achieved remarkable improvements. Rogers et al[28] developed a model capable of precisely identifying 88.5% of reflux events by analyzing impedance fluctuations in a study of 232 participants. Following episodes of reflux, the body initiates a defensive response known as post-reflux swallow-induced peristaltic wave (PSPW). The PSPW index provides a measure of the esophagus’s chemical clearance ability and could serve as a more effective variable for GERD diagnosis than standard pH-impedance measures (P < 0.01)[29]. Nevertheless, this model has limitations. While the analysis of baseline impedance alterations can detect reflux events, the diagnosis of GERD still requires manual calculation of the PSPW index. To address this issue, Wong et al[30] incorporated body mass index (BMI) and PSPW data alongside the previously mentioned parameters. By utilizing 7939 impedance events collected from 106 patients, a refined ML model was developed that achieved an accuracy of 87% in recognizing reflux episodes and 82% accuracy in calculating the PSPW index. While the two studies above utilized an established pH-impedance index, Zhou et al[31] achieved instant and comparable diagnostic accuracy with a smaller dataset (n = 45) by employing real-time physiological markers derived directly from raw pH-impedance monitoring data. Notably, all three of these ML models exhibited higher performance [area under the curve (AUC) = 0.87] than manual assessment[31].

Leveraging ML to analyze PH impedance monitoring data can not only improve the accuracy of GERD diagnosis but allow for values such as PSPW, which previously needed to be calculated manually, to be determined automatically. Furthermore, ML models enable much more rapid GERD diagnosis, eliminate diagnostic discrepancies resulting from variations in medical practices and expertise, and substantially reduce patient waiting times.

Progress in automatic grading of RE endoscopic images

RE severity is traditionally assessed using the Los Angeles classification system, which classifies mucosal damage from grade A to grade D[32]. This system provides a relatively accurate reflection of GERD severity and serves as a clinical standard for determining surgical candidacy[33]. The assessment of RE relies on endoscopic examination, and encompasses traditional red, green, and blue (RGB) endoscopy and narrow band imaging (NBI), which employs specific blue and green light wavelengths to illuminate the target area[34]. Compared to RGB, NBI enhances the visualization of vascular and mucosal patterns, offering advantages in detecting early-stage tumor lesions. Nevertheless, despite the implementation of NBI, the maximum level of accuracy achieved in manually grading RE is only 77%[33].

In recent years, advancements in esophageal endoscopic image recognition have been achieved through the application of deep learning models, particularly deep CNN. Ge et al[33] pioneered the application of deep learning to predict Los Angeles grading from esophageal endoscopic images, and evaluated the performance of deep CNN, including ResNet, EfficientNet, DenseNet, and Visual Geometry Group (VGG), for this task. The results demonstrated that DenseNet121 achieved the highest accuracy (87%) on a test set of 2081 images, exhibiting a 10% enhancement compared to manual grading. The study demonstrated that current models exhibit strong performance in differentiating between mild (grades A and B) and severe (grades C and D) Los Angeles categories. Nevertheless, they face difficulties in the more complex four-category classification test. To solve this problem, Wang et al[35] developed the GERD-VGG Network model, which is based on the VGG16 backbone network. On a dataset of 464 patients’ RGB and NBI endoscopic images, the model obtained 87.9% accuracy. Building on these findings, Yen et al[36] investigated embedding an RF model in various neural network layers and improved accuracy by 5% in the four-category Los Angeles classifications. Li et al[26] constructed the ECA-ResNext50 model based on the ResNext50 backbone network and evaluated it on an expanded dataset of more than 3000 esophageal endoscopic images. In comparison to models suggested by Wang et al[35] and Yen et al[36], the ECA-ResNext50 model showed higher robustness on a larger dataset, achieving a consistent accuracy of 90%. Huang et al[37] went beyond the assessment of Los Angeles grade and instead used esophageal endoscopic images to directly diagnose RE. They divided the endoscopic images of 147 individuals into smaller, independent images and extracted heterogeneous color features from each of them. These features were subsequently incorporated into their heterogeneous descriptor fusion-SVM model, which was trained using SVMs. This model achieved a remarkable 93.2% diagnostic accuracy for RE, surpassing both AI-assisted grading approaches and manual evaluation.

These studies collectively demonstrate that while various ML models achieve higher LA grading accuracy than manual evaluation, opportunities for further advancement remain. Expanding the dataset to include more severe cases can greatly improve the performance of the model and decrease the chances of misclassifying severe patients as having milder disease. Data augmentation would significantly benefit the development of accurate AI-assisted RE assessment approaches that rely on deep learning image recognition.

Advances in invasive diagnosis and non-invasive screening of BE

BE, a frequently occurring complication of GERD, requires timely detection due to its potential progression to EC[38]. Currently, endoscopy is considered the most reliable method for screening BE, but the diagnostic accuracy relies heavily on the physician expertise and its invasive nature often leads to low patient compliance[39]. Thus, it is crucial to investigate the development of non-invasive methods for screening for BE, as endoscopic examinations are invasive and can only indicate the severity of the lesions.

Recent advances in ML have yielded promising results in both invasive and non-invasive BE detection methods. For instance, Faghani et al[40] utilized YOLOv5, a rapid and accurate deep learning model for object detection, to analyze 8596 esophageal endoscopic images from 542 patients, achieving a BE detection rate of 88%-95% and with sensitivity of 81.3% and specificity of 100%. Recognizing the limitations of invasive procedures, several non-invasive approaches have shown promising results. Rosenfeld et al[25] developed a model based on questionnaire surveys, selecting eight BE-related indicators (e.g., age, sex, and waist circumference) from 1299 subjects, achieving diagnostic accuracy of 81% for BE. Expanding on this approach, Iyer et al[41] analyzed the electronic health records (EHRs) of 262219 patients with BE, constructing a model based on BE family history, clinical symptoms, and laboratory test results, achieving overall sensitivity, specificity, and area under the receiver operating characteristic curve of 76%, 76%, and 0.84, respectively. Berman et al[42] explored the cytosponge-trefoil factor 3 (TFF3) approach for non-invasive BE ML-assisted diagnosis. This minimally invasive approach uses a specialized capsule sponge to collect samples of esophageal cells and detect TFF3 protein levels, assessing intestinal metaplasia lesions to aid BE diagnosis. This method achieved an accuracy of 77% with a low false-negative rate, potentially eliminating the need for unnecessary endoscopic examinations in more than half of potential cases of BE.

The evolution of BE detection methods demonstrates a clear trend toward combining high accuracy with patient convenience. ML models based on non-invasive methods such as questionnaire surveys, EHR analysis, and cytosponge-TFF3 offer promising alternatives. Although these models demonstrate slightly lower accuracy compared to endoscopic imaging models, their non-invasive nature, convenience, and higher patient compliance present significant advantages. These characteristics position them as potentially valuable tools for early BE screening. These complementary approaches also suggest a future where initial non-invasive screening could efficiently identify high-risk patients, thereby optimizing healthcare resource allocation.

ML IN EC

EC represents one of the most aggressive digestive system malignancies, characterized by its insidious progression and late-stage diagnosis. The stark contrast in survival rates-90% for patients with early-stage disease vs below 20% for advanced cases-underscores the critical importance of early detection[43]. However, patients with early-stage EC often present with no obvious symptoms, and currently there are no reliable biomarkers for accurate EC screening. Furthermore, endoscopic images of early-stage EC lack distinct morphological features, making accurate and timely diagnosis challenging for clinicians. Therefore, improving the early diagnosis rate of EC and accurately assessing patient prognosis are critical issues that need to be addressed. The advent of ML technology offers promising solutions for improving both early detection and prognosis assessment through various approaches: Analysis of clinical records, imaging data interpretation, and multi-omics integration[44].

Early EC screening models using textual records and clinical data

While endoscopy remains the gold standard for EC diagnosis, its invasive nature limits widespread application for early screening. By contrast, patient medical records (textual) and laboratory data are readily accessible and cover a wide population, providing a rich data foundation for early EC risk assessment. However, the vast amount of data and numerous indicators make manual analysis challenging, underscoring the crucial role of ML technology. By mining and analyzing raw data, ML models can identify key risk factors and develop early EC screening tools, thereby assisting clinicians in early diagnosis and intervention.

The availability of vast EHRs and laboratory data has provided new opportunities for ML-based early screening of EC. Rubenstein et al[45] conducted a systematic analysis of 11.4 million EHRs from United States veterans between 2005 and 2008, and developed the Kettles Esophageal and Cardia Adenocarcinoma prediction (K-ECAN) model based on patient information such as age, sex, smoking history, and reflux disease history. This model can rapidly identify patients with early-stage EC using textual information, and the AUC of the model reaches 0.77-0.81. Gao et al[44] further integrated textual medical records, laboratory data, and 105 cytological features derived from pathological sections to construct an early EC screening model encompassing 17000 individuals. Their model achieved a predictive accuracy approximately 10% higher than the K-ECAN model. In addition to utilizing patient health records and laboratory data, Meng et al[46] recognizing the close association between EC development and oral microecological imbalance, collected and analyzed the relative abundance data of five bacterial species in saliva samples from patients with EC. They employed various ML models to develop EC prediction models. The results indicated that the Gradient Boosting model optimized with a Genetic Algorithm performed best in validation (accuracy = 70.60%, precision = 46.00%, recall = 90.55%, F1-score = 61.01%), and revealed that Bacteroides and Actinomyces are significant biomarkers for predicting esophageal squamous cell carcinoma (ESCC).

While ensuring high accuracy, easily accessible textual data and non-invasive clinical data provide new possibilities for early diagnosis of EC. These diverse screening approaches demonstrate the potential of ML in maximizing the utility of readily available clinical data for early EC detection while reducing unnecessary invasive procedures.

Multidisciplinary treatment, radiotherapy risk, and prognosis in EC

EC management presents significant clinical challenges due to the necessity for complex, multidisciplinary treatment decisions including surgery, neoadjuvant chemotherapy, chemoradiotherapy, targeted therapies, and immunotherapies. Such complexity emphasizes the need for highly individualized treatment strategies. Given the challenges in optimizing complicated clinical decisions and the potential for treatment variability, Thavanesan et al[47] investigated the utility of ML in predicting multidisciplinary team (MDT) recommendations for patients with EC. To achieve this, the authors assessed multiple ML algorithms, including multinomial logistic regression (LR), RFs, extreme gradient boosting, and decision trees, to analyze clinical data from 399 patients with EC. Among the tested models, multinomial LR achieved best performance with an AUC of 0.793. Notably, age was identified as a significant factor in determining the preference for either upfront surgery or neoadjuvant chemotherapy followed by surgery. This study highlights the potential of ML techniques to support MDTs in standardizing EC treatment decisions, while also highlighting potential biases and providing data-driven insights to optimize complex treatment selections for patients with EC. In contrast to Thavanesan et al’s research[47], Huang et al[48] developed a personalized treatment decision model specifically designed for the challenging subgroup of inoperable, elderly patients with EC. This approach integrates computed tomography (CT) images, clinical data, and blood test results to train a multitask learning model termed CTDEPN. CTDEPN demonstrated robust performance, achieving an AUC of 0.91 on the testing set and 0.84 on an external validation set for predicting objective response rate. These studies demonstrate the potential of ML models to serve as valuable tools within clinical decision support systems for EC management. Both models exemplify the capacity of ML to augment clinical expertise, facilitate data-driven decision-making, and advance personalized medicine in EC care.

Radiotherapy remains a crucial modality in the treatment of EC, yet a significant challenge is the development of radiotherapy-induced lymphopenia (RIL). Severe RIL (grade 4 RIL) compromises the patient’s immune function, elevates the risk of infection, and potentially impairs therapeutic efficacy and patient survival. The lack of reliable tools for assessing severe RIL risk presents a critical clinical challenge. Zhu et al[49] addressed this issue by using eXtreme gradient boosting and clustering algorithms to create an individualized risk assessment model for RIL. This model, which incorporates both clinical and radiation treatment factors, successfully identified high-risk patients with RIL with an AUC of 0.783. The study also suggested that proton therapy may offer a means of RIL mitigation compared to photon therapy, offering opportunities for more individualized treatment strategies in EC. Building on this research, Chu et al[50] advanced the field by stratifying patient subgroups with EC according to RIL risk. Through the implementation of unsupervised learning (hierarchical density-based spatial clustering of applications with noise and uniform manifold approximation and projection), which enabled the delineation of subgroups based on clinical stage and radiotherapy modality. This approach facilitated a more granular exploration of heterogeneous risk factors and the development of subgroup-specific predictive models. This refined and integrated approach culminated in a comprehensive risk assessment model, demonstrating an AUC of 0.783. These findings not only demonstrate the potential advantages of proton therapy for RIL prevention but also underscore the paramount importance of individualized risk stratification as a guide for treatment decisions in patients with EC undergoing radiotherapy.

The poor overall prognosis of patients with EC and the lack of reliable diagnostic biomarkers underscore the importance of developing accurate prognosis models. ML models have demonstrated superior performance in predicting the survival of patients with EC compared to traditional statistical methods such as Kaplan-Meier and Cox regression analyses. Nopour[51] developed a prognosis model using the RF algorithm, incorporating data such as medical records and tumor characteristics from 1656 patients. Their model achieved a 5-year survival prediction accuracy of 0.76. Rahman et al[52] employed the random survival forest (RSF) method, incorporating 41 variables from 6399 patients, including patient characteristics, disease information, treatment regimens, postoperative complications, and pathological features. They developed a model for predicting overall survival after EC resection, achieving a 5-year survival prediction accuracy 10% higher than that of Nopour[51]. However, this model lacked external validation. Lu et al[53] optimized the RSF model and conducted a more detailed analysis of 3-year and 5-year survival in 2521 patients, incorporating 27 clinical features, including age, sex, tumor information, and routine blood test parameters. The AUC values for their 3-year and 5-year survival predictions reached 0.76 and 0.77, respectively, comparable to the results of Nopour[51]. These studies demonstrate that ML models can effectively handle non-linear variables, improve prediction accuracy, and generally achieve prediction accuracy rates approximately 6%-10% higher than Cox proportional hazards models.

ML enhances EC therapy management by providing data-driven support across various treatment stages. Firstly, it facilitates personalized treatment planning for complex multidisciplinary decisions, as well as effectively assesses radiotherapy risks, enabling individualized risk stratification and potentially guiding mitigation strategies. Furthermore, ML-based prognostic models improve accuracy in prognosis prediction compared to traditional methods. By integrating diverse clinical data, these ML applications serve as valuable tools for contributing to more precise EC management and will likely improve patient outcomes.

EC progression assessment models based on imaging data

The morphological changes in the mucosa and vasculature of early lesions are often subtle, making accurate identification difficult even for experienced endoscopists[35]. Furthermore, the considerable inter-individual variability in EC lesion presentation[36] further complicates the accurate identification of tumors and the segmentation of lesion areas.

To address these challenges, several studies have explored deep learning-based models for early EC lesion identification, aiming to improve the accuracy of early EC detection and staging. Fockens et al[54] pioneered the development of a neural network-based lesion identification model to assist endoscopists in the diagnosis of early tumors associated with BE. Their results demonstrated that with the assistance of the model, the accuracy of identifying EC lesions from images could be improved from 74% to 88%. Li et al[55] focused their research on lesion segmentation in EC images, particularly for early EC and polyps. Their model, trained on 1230 samples, was able to accurately segment 88% of the lesion area in an image within 4 minutes. Yuan et al[43] utilized endoscopic images and video data from 1112 patients with EC to develop a model capable of both EC identification and lesion segmentation. This model achieved a tumor identification accuracy of 89.9% and a lesion segmentation accuracy of 87%.

Accurate cancer staging is crucial for developing personalized treatment plans and prognosis assessment. However, traditional manual staging methods based on endoscopic images have limited accuracy and efficiency. ML offers new avenues for improving the accuracy of EC staging. Knabe et al[56] developed an EC tumor-node-metastasis (TNM) staging model using the VGG16 algorithm and 1020 endoscopic images, achieving 73% accuracy in EC identification, 71% accuracy in TNM staging, and 85% accuracy in identifying healthy individuals. To further enhance model performance and generalizability, strategies of integrating publicly available datasets and in-house datasets for model optimization were often employed. For instance, Chempak Kumar and Mubarak[57] integrated an in-house dataset comprising 455 endoscopic images with 573 endoscopic images sourced from publicly available datasets such as KVASIR[58], constructing a larger training dataset. By employing an ensemble learning strategy, they improved the accuracy of EC TNM stage prediction to 97%. These studies indicate that ML models hold the potential to surpass traditional manual assessment methods in EC staging, offering both improved accuracy and significantly enhanced analytical efficiency.

In addition to endoscope, CT imaging analysis has provided valuable insights in ML-based EC progression assessment. Researchers commonly extract radiomic features from CT images and convert them into structured data for subsequent analysis. For example, Jia et al[59] used this approach to stratify 546 patients into four subtypes, with significant differences in 3-year progression-free survival among the subtypes (P = 0.013). Given the close association between skeletal muscle depletion and prognosis in patients with EC, Vogele et al[60] analyzed CT scans from 83 patients, extracted skeletal muscle-related features, and developed a model using the RF algorithm to predict tumor progression (AUC = 0.93). Studies suggest that while CT imaging holds significant potential for EC progression assessment, its current application remains limited to CT images from which clear image features can be extracted.

ML leverages imaging data to significantly enhance EC assessment throughout different stages of the disease. By utilizing endoscopic images, ML models have shown higher accuracy and efficiency in assessments of early EC lesion segmentations. Furthermore, in CT imaging, ML effectively predicted tumor progression and differentiated survival prognosis subtypes by extracting radiomics features. Collectively, the application of ML to imaging data substantially improved the accuracy and efficiency of EC progression assessment.

Based upon clinical imaging data, ML models have been able to advance EC assessment. Endoscopy-based ML has improved early lesion detection and staging accuracy, while CT radiomics analysis has been able to predict progression and differentiates prognosis. Overall, ML has enhanced the accuracy and efficiency of EC assessment through imaging data.

Multi-omics-based diagnostic models for EC

Recent genomic advances have enabled more precise EC diagnostics through ML integration. Studies have shown that mutations in specific genes, such as tumor protein p53, cyclin-dependent kinase inhibitor 2A, and retinoblastoma protein 1, are strongly associated with the development and progression of EC and can serve as potential markers for risk assessment and prognosis prediction[61]. Furthermore, gene expression profiling and pathway analysis have become essential tools for identifying key hub genes and signaling pathways in EC[62]. Specific gene expression markers have been shown to improve the detection efficiency of early EC lesions and specific subtypes[63]. For instance, high expression of the family with sequence similarity 135 member B gene is associated with poor prognosis in ESCC, an EC subtype, suggesting its potential as a prognostic biomarker and therapeutic target. However, the high inter-individual heterogeneity in EC expression profiles leads to variable performance of different tumor markers across diverse patient cohorts[64], necessitating further research to improve the generalizability of EC molecular markers.

Traditional statistical methods for identifying differentially expressed genes (DEGs) are easily affected by sample variations, a limitation that ML methods can effectively overcome. This approach enables secondary screening of DEGs to identify key biomarkers that can effectively differentiate between normal and EC samples. Zhou et al[64] combined gene set enrichment analysis and ML to identify key signaling pathways and biomarkers associated with EC development and progression. Compared to traditional methods, their approach yielded biomarkers with superior predictive performance, robustness, and functional interpretability. Zhang et al[62] using 47 paired EC tumor and normal tissue samples from the Gene Expression Omnibus database, employed ML to identify five EC biomarkers. Three of these biomarkers-glutathione peroxidase 3, MMP1, and MMP12, associated with immune infiltration, have demonstrated good diagnostic value in independent validation (AUC > 0.92).

These studies suggest that ML can serve as a valuable complement to conventional DEG analysis. Xu et al[63] collected data from 14 patients with BE and 8 patients with EC. Using differential gene expression analysis on 12 samples to select 160 DEGs. In a test set of the remaining 10 samples, the differential gene expression analysis misclassified two samples, while the artificial neural network correctly classified all 10 samples. Chuwdhury et al[65] used multi-omics sequencing data from 805 gastrointestinal tumors, including EC, to develop an RF model for identifying neoantigens incorporating genomic and transcriptomic features. The model achieved an AUC of 0.83. Sasagawa et al[66] analyzed genomic and transcriptomic data from 143 patients with EC at a Japanese hospital to investigate the association between immunogenomic features and response to neoadjuvant chemotherapy. A decision tree model predicted response to neoadjuvant chemotherapy with an accuracy of 84%. While ML demonstrates superior performance in selecting features compared to differential expression analysis, its limited interpretability often relegates it to a supplementary role following differential expression analysis.

ML has significantly optimized multi-omics-based EC risk stratification. By identifying key genomic and transcriptomic feature, ML has successfully constructed high-precision diagnostic and prognostic models, effectively overcoming the limitations of traditional differential gene analysis. However, despite the improved capabilities of integrating multi-omics data, ML remains limited by heterogeneity and interpretability, setting its use to that of a supplementary tool to compliment conventional methods.

ML APPLICATION IN OTHER ESOPHAGEAL CONDITIONS

Besides GERD and EC, ML has demonstrated significant utility in diagnosing and managing other esophageal disorders, particularly achalasia cardia and EV. Achalasia results from esophageal neuromuscular dysfunction, while EV is primarily caused by portal hypertension, often secondary to cirrhosis. By analyzing the contribution of different variables to the model, ML excels in revealing correlations that are difficult to discern using traditional approaches, offering new perspectives for early disease warning and risk assessment.

ML-assisted diagnosis and subtyping of achalasia

Achalasia, a rare esophageal motility disorder, relies primarily on HREM for diagnosis. HREM provides precise measurements of esophageal pressure and sphincter function, allowing a comprehensive assessment of esophageal motility. Based on HREM findings and using the Chicago classification, achalasia can be categorized into three subtypes (types I, II, and III).

However, interpreting HREM results requires manual landmarking and analysis, in conjunction with patient history, symptoms, and other clinical findings. This process demands considerable expertise and experience from clinicians and increases diagnostic turnaround time. To address these limitations, Popa et al[67] developed an automated, ML-based diagnostic system for HREM images. This system eliminates the need for manual landmarking and can achieves 93% accuracy in identifying esophageal motility disorders. Furthermore, it can generate diagnostic interpretations of HREM images instantaneously, significantly improving the efficiency of esophageal motility disorder diagnosis.

Regarding subtyping, Carlson et al[68] employed a decision tree model, incorporating age, sex, medication history, and HREM parameters from 180 patients with achalasia, to differentiate between subtypes. The model achieved an overall accuracy of 71%. While the model performed well in distinguishing between type II and type III achalasia, its performance in identifying type I achalasia remained suboptimal. To further optimize subtyping, Takahashi et al[69] performed cluster analysis on age, sex, BMI, disease duration, HREM findings, and treatment strategies in 1824 patients. They classified patients into three categories: Type 1, characterized by classic achalasia with esophageal dilation; Type 2, characterized by compressive achalasia with esophageal dilation; and Type 3, characterized by esophageal pressurization with esophageal shortening/spasm. These three types corresponded to Chicago classification types I-III, respectively.

These studies emphasize the capacity of ML in achalasia management. By analyzing HREM images, ML has significantly improved the diagnostic efficiency for esophageal motility without manual annotation. Furthermore, ML methods have been able to aid in identification of differentiating achalasia subtypes based on distinct patterns in patient data, providing a new approach for accurate diagnosis and personalized treatment of achalasia.

EV bleeding risk assessment and classification

EV is a common and serious complication of cirrhosis, as cirrhosis progresses, the incidence of EV and the risk of variceal bleeding increase, leading to a higher risk of mortality. To assess the risk of EV bleeding in patients with cirrhosis, ML could be employed to develop predictive models using patient medical history, laboratory results, and imaging data. Yan et al[70] utilized an SVM algorithm to extract radiomic features including texture, statistical, wavelet, and histogram of oriented gradients features from the CT images of 546 patients, achieving 86% accuracy in differentiating between mild EV and EV with a high risk of bleeding. Gao et al[71] developed a LR model incorporating demographic information (age, sex), clinical data (etiology of cirrhosis, clinical endpoints), laboratory data (liver function, renal function, platelet count, hemoglobin, and coagulation function), and CT-derived features to predict acute EV bleeding events with 78% accuracy. However, considering the challenges of acquiring and interpreting CT data, Agarwal et al[72] and Hou et al[73] developed predictive models using medical history and laboratory data from 828 and 999 patients, respectively, achieving accuracies of 86% and higher, respectively. Furthermore, Bayani et al[74] found that traditional ML algorithms, such as RF, SVM, and LR outperformed neural network-based models (78%) in classifying EV morphology and size using clinical text and laboratory data, achieving higher accuracies (83%-95%). These studies demonstrate that ML models can effectively and accurately predict EV bleeding risk, providing valuable support for clinical decision-making and potentially reducing reliance on CT scans.

Endoscopy is a crucial method for diagnosing EV, but endoscopic assessment of risk factors can be subjective. Researchers are exploring ML-based endoscopic image analysis techniques to improve the objectivity and accuracy of EV diagnosis and risk assessment. Wang et al[75] developed the ENDOANGEL-gastroesophageal varices (GEV) AI Applications Institute system, which integrates multiple deep learning models. Trained and validated on a multicenter dataset of 6034 images from 1156 patients, the system achieved an accuracy of over 88% for EV detection and grading, outperforming experienced endoscopists. Chen et al[76] utilized a real-time deep CNN system, ENDOANGEL, for GEV diagnosis and rupture risk prediction, achieving 97% accuracy on a dataset of nearly 15000 images from 6189 patients. Furthermore, Hong et al[77], using the EfficientNet model, analyzed endoscopic images from 675 patients and achieved 91% accuracy in predicting EV bleeding risk within 12 months. They also found that combining endoscopist assessments with AI-assisted diagnosis could further improve predictive accuracy. Wang et al[78] developed an automated multimodal ML model, visualizing results with Grad-class activation mapping heatmaps, and achieved 86.8% accuracy in predicting EV bleeding risk on 810 images.

ML has reached a high-precision prediction performance for EV bleeding risk assessment in liver cirrhosis by integrating CT imaging features and clinical data. In endoscopic diagnosis, deep learning models have achieved high accuracy in EV automated detection, grading, and bleeding risk prediction. Notably, combining AI with physician evaluation further enhances diagnostic efficiency and significantly improve objectivity and efficiency. Overall, through the analysis of imaging and clinical data, ML offers promising approaches for accurate EV risk assessment and improved patient outcomes.

DISCCUSSION AND FUTURE PERSPECTIVES

As shown in Table 1, the ML models mentioned above have been developed and validated for various esophageal disorders, covering aspects from risk screening and diagnosis to prognosis prediction and clinical decision support. Multiple studies indicate its ability to achieve diagnostic accuracy surpassing that of expert clinicians in certain tasks[31,33,37,75-77]. ML models for GERD and BE diagnosis achieve 80%-90% accuracy, while EC screening models using clinical data show comparable performance to image-based methods. However, the actual performances vary across studies due to limitations in data availability. In survival prediction for EC, the accuracy remains to be improved (around 75%); however, it still offers a 6%-10% improvement compared to non-ML approaches[52]. For EV bleeding risk prediction, ML models have shown promising results using both clinical and endoscopic data, offering the potential for non-invasive risk stratification. However, the limited interpretability of ML hinders its wider application in multi-omics research. Despite this limitation, studies have shown that ML-selected genes can better differentiate patients with and without cancer compared to those identified through traditional way[63]. Despite these achievements, its widespread clinical implementation faces several challenges. Data limitations represent a primary bottleneck; preprocessing medical data (cleaning, quality control, structuring) demands substantial time and resources[79,80]. Additionally, restrictions imposed by ethical review and data sharing hinder the development of large, high-quality datasets, thereby impacting model generalizability.

Table 1 Performance of machine learning models for diagnosis, risk stratification, and prognosis in esophageal diseases.

Ref.	ML classifier	Disorder	Outcomes	Data type (n)	Performance	Validation (n)
Risk factor screening
Yan et al[70]	SVM	EV	Identify esophageal varices with high bleeding risk	CT image (796), China	AUC = 0.74	External validation (405)
Hong et al[77]	EfficientNet/Grad-CAM	EV	Assess the risk of EV bleeding within 12 months	Endoscopy (675), China	ACC = 91%	External validation (400)
Chen et al[76]	DCNNs	EV	Classify the EV grade to identify patients at high risk of bleeding	Endoscopy (10655), China	ACC = 97%	Independent validation (200)
Gao et al[44]	LightGBM	EC	Screen for carcinoma at the esophagogastric junction	Questionnaire, cytology and endoscopy (17000), China	AUC = 0.96	External validation (2901)
Iyer et al[41]	Transformer	EC	Screen for patients with BE and EC among healthy individuals	Clinical data (260000), United States	AUC = 0.84	Independent validation (10%)
Rosenfeld et al[25]	Logistic regression	BE	Screen for BE and alert patients who may need endoscopy	Questionnaire (1299), United Kingdom	AUC = 0.86	Independent validation (523)
Fockens et al[54]	EfficientNet/MobileNetV2/DeepLabV3 +	BE, EC	Detect and mark lesion sites of BE and EC	Endoscopy (13000), Netherlands	Sensitivity = 88%	Independent validation (200)
Prognosis and survival analysis
Wang et al[75]	DCNNs	EV	Determine size, shape, color and bleeding signs of EV and GV	Endoscopy (17000), China	ACC = 93%	External validation (11000)
Nopour[51]	Random forest	EC	Predict 5-year survival of patients with EC	Clinical data (1656), Iran	AUC = 0.76	External validation (100)
Lu et al[53]	Random forest	EC	Predict 3-year and 5-year survival of patients with EC	Clinical data (2521), China	AUC = 0.77	Independent validation (30%)
Knabe et al[56]	VGG16	EC	Identify BE, early EC, and advanced tumors (T1b sm2, T3, T4)	Endoscopy (1020), Germany	ACC = 73%	Independent validation (199)
Personalized medication
Sasagawa et al[66]	Decision tree	EC	Predict chemotherapy response by immunogenomic features	Genomic and transcriptomic data (121), Japan	ACC = 84%	Independent validation (30%)
Chuwdhury et al[65]	Random forest	EC	Predict neoantigens by multi-omics features	Genomic and transcriptomic data (805), China	AUC = 0.87	Independent validation (211 peptides)
Zhu et al[49]	XGBoost	EC	Analyses differences in G4RIL risk between proton and photon therapies	Clinical data (746), United States	AUC = 0.78	Independent validation (247)
Chu et al[50]	XGBoost and logistic regression	EC	Personalized prediction of G4RIL based on CDS and four clinical risk factors	Clinical data (860), United States	AUC = 0.78	Independent validation (20%)
Clinical phenotype classification
Li et al[26]	ResNeXt50	GERD	Classify the Los Angeles grade of reflux esophagitis	Endoscopy (3498), China	ACC = 90.2%	Independent validation (396)
Jia et al[59]	Consensus clustering	EC	Classify EC into four clusters with BRCA1, PD-1, vascular invasion, and tumor stages	CT image (546), China	P = 0.035	External validation (546)
Chempak Kumar and Mubarak[57]	Artificial bee colony/CNN/SVM	BE, ESCC, EAC	Detect and analyze BE, EAC, and ESCC	Endoscopy (1028), India	ACC = 97%	3-fold cross-validation
Takahashi et al[69]	Hierarchical clustering	Achalasia	Classify achalasia into three clusters, according to Chicago classification	Demography, clinical data and endoscopy (1824), Japan	AUC: 0.61-0.7	External validation (1824)
Carlson et al[68]	Decision tree	Achalasia	Classify achalasia and distinguish spastic from non-spastic types	HRM data (180), United States	ACC = 78%	Independent validation (40)
Faghani et al[40]	YOLOv5	BE	Classify BE into NDBE, LGD, and HGD	Histology (542), United States	ACC = 81.3%	Independent validation (70)
Clinical decision support system
Rogers et al[28]	Decision tree	GERD	Identify MNBI, and diagnose GERD based on MNBI	pH-impedance data (325), United States and Italy	ACC = 88.5%	Independent validation (2049)
Wong et al[30]	ResNet18	GERD	Identify reflux events and calculate the PSPW index	pH-impedance data (106), China	ACC = 87%	Independent validation (10%)
Yen et al[36]	VGG16/RF	GERD	Assess esophageal mucosal damage	Endoscopy (496), China	ACC = 92.5%	Independent validation (32)
Zhou et al[31]	S4 model	GERD	Identify reflux events	pH-impedance data (45), United States	AUC = 0.87	Independent validation (20)
Meng et al[46]	XGBoost	EC	Diagnose ESCC according to the relative abundance of salivary flora	Microbiome (8000), Multisource	ACC = 89.9%	5-fold cross-validation
Rubenstein et al[45]	Logistic regression/decision tree/XGBoost	EC	Distinguish early- and late-stage EC	Clinical data (11400000), United States	AUC: 0.75-0.85	Independent validation (2600000)
Yuan et al[43]	YOLACT	EC	Detect and segment lesions of patients with EC	Endoscopic image (10000) and video (140), China	ACC = 87%	External validation (1141)
Thavanesan et al[47]	Logistic regression, random forests, extreme gradient boosting and decision tree	EC	MDT treatment decisions predict and curative EC	Clinical data (399), United Kingdom	AUC: 0.71-0.79	10-fold cross-validation
Huang et al[48]	Extremely randomized trees	EC	Individualized treatment decisions for elderly patients with inoperable ESCC	Clinical data, CT and endoscopy (189), China	AUC = 0.84	Independent validation (20%)
Li et al[55]	eUNet	EC	Segment lesions in EC endoscopic images	Endoscopy (2848), Multisource	Dice = 0.89	Independent validation (300)

AUC: Area under the curve; ACC: Accuracy; BE: Barrett’s esophagus; CDS: Composite dosimetric score; CNN: Convolutional neural network; DCNNs: Diffusion-convolutional neural networks; Dice: Sørensen-Dice coefficient; EAC: Esophageal adenocarcinoma; EC: Esophageal carcinoma; ESCC: Esophageal squamous cell carcinoma; eUNet: edge-enhanced UNet; EV: Esophageal varices; G4RIL: Grade 4 radiotherapy-induced lymphopenia; GERD: Gastroesophageal reflux disease; HGD: High-grade dysplasia; LGD: Low-grade dysplasia; LightGBM: Light gradient boosting machine; MNBI: Mean nocturnal baseline impedance; NDBE: Nondysplastic Barrett’s esophagus; PD-1: Programmed death-1; PSPW: Post-reflux swallow-induced peristaltic wave; ResNet: Residual neural network; ResNeXt: Residual networks with next; RF: Random forest; S4 model: Structured state space sequence model; SVM: Support vector machine; VGG: Visual geometry group; XGBoot: EXtreme gradient boosting; YOLACT: You only look at coefficients; YOLO: You Only Look Once; MDT: Multidisciplinary team; CT: Computed tomography; pH: Potential of hydrogen.

Model selection poses another significant challenge. Current research often relies on empirical testing of multiple models to identify optimal performers[33,44,45,51,53,64,74,81-84], lacking a principled approach. Traditional ML models, based on statistical principles, possess specific features that can guide selection based on data attributes, supported by strict mathematical derivations. Conversely, deep learning models often behave as “black boxes”, with selection heavily relying on experience and transfer learning from other domains. Although the structure of deep learning models is known, the explanation for how millions of parameters specifically impact data processing and modeling remains unclear. As such, explainable AI (XAI) methods are being explored to address this, with gradient-based methods like gradient-weighted class activation mapping (Grad-CAM) and perturbation methods observing output changes by masking or modifying input features. When van der Velden et al[85] conducted a statistical analysis of XAI in medicine, they found Grad-CAM to prevail. Nonetheless, this challenge remains, and is compounded by limited model interpretability which can hinder clinical adoption despite strong performance metrics.

Model deployment also faces practical challenges. From a cost perspective, obtaining raw data directly from numerous medical devices with various models is impractical. Consequently, the deployment of ML models still necessitates integration with the hospital’s electronic medical record systems and is reliant on support of system developers. Even after deployment, the model may still face resistance and hesitancy from the clinical staff, particularly when considering the lack of transparency of deep learning models. Therefore, it is essential to provide thorough training to clinical doctors on how to effectively use ML models with emphasis on the supportive nature of ML tools that will benefit both their practice and their patients.

Furthermore, ethical and security concerns restrict access to medical data. Medical data are typically compartmentalized within isolated hospital intranets, preventing the aggregation of large, centralized, high-quality medical datasets. The application of ML models must also work within the framework of complicated regulatory hurdles and in respect to ethical issues related to real-world patient data. Addressing these challenges necessitates international collaboration among medical organizations to establish ethical consensus, coordinate cross-border data flows, and ensure accountability. Several strategies have been proposed to address these limitations. Data augmentation may benefit the low-data scenario, but the inherent risks of overfitting and reduced generalizability must be considered. Transfer learning has demonstrated promising value in such scenarios[86], however it depends on similar data distributions between the source and target domains. To overcome the challenge posed by fragmented (“small and scattered”) medical data, federated learning (FL)[87], a distributed learning approach, has emerged. FL allows multiple institutions to train models locally using their own data, exchanging only model parameters and weights. This approach enhances data security while expanding the effective dataset size and improving model robustness and generalizability. Consequently, FL has paved the way for multicenter studies and broader application of ML in esophageal diseases. Looking ahead, several key developments could significantly advance the field. The continued improvement of data sharing mechanisms and wider adoption of FL could facilitate larger, more diverse datasets. By integrating multimodal data, ML can facilitate the development of more precise risk prediction models for EC and other esophageal diseases, enabling earlier warnings and interventions. It can also assist clinicians in analyzing pathological images and interpreting endoscopic findings, thereby improving diagnostic efficiency and accuracy, and ultimately enabling personalized treatment strategies. Moreover, analyzing patient information can enable the prediction of recurrence risk and survival time, informing treatment strategies and predicting treatment response, thereby advancing precision medicine.

CONCLUSION

This review highlights the substantial potential of ML across broad applications of esophageal disorder management, including risk factor screening, prognostic modeling, phenotype classification, clinical decision support, and personalized treatment planning (summarized conceptually in Figure 3). By leveraging diverse data modalities, AI facilities risk stratification, promotes the development of individualized survival predictions, and enables objective characterization of disease phenotypes while enhancing clinical decision-making. However, realizing the full potential of these technologies requires overcoming persistent challenges, particularly regarding data standardization, model interpretation, and workflow integration into existing clinical workflows. Looking ahead, the current trajectory suggests continued expansion of ML in esophageal disease management. Opportunities include improved early warning systems for disease progression, enhanced analysis of pathological and endoscopic findings, and more accurate prediction of treatment responses. Eventually, enhancing model interpretability is essential to foster clinician trust and accelerate the application of ML tools in clinical practice, ultimately translating AI innovations into tangible improvements in esophageal diseases management.

Figure 3 Conceptual overview of artificial intelligence-driven applications in esophageal disorder management focusing on risk factor screening, prognostic modeling, phenotype classification, clinical decision support, and personalized treatment planning. AI: Artificial intelligence; CT: Computed tomography; EV: Esophageal varices; PSPW: Post-reflux swallow-induced peristaltic wave.