Can artificial intelligence improve the diagnosis and management of patients with eosinophilic esophagitis?

Abstract

Eosinophilic esophagitis (EoE) is a chronic, immune-mediated condition leading to esophageal inflammation and a range of symptomatic complications if inadequately managed. Recent epidemiological trends indicate a significant increase in EoE prevalence, complicating patient care amid diagnostic challenges associated with conventional methods such as endoscopy and histopathological analysis. This review explores the promise of artificial intelligence (AI) and deep learning models in enhancing the diagnosis and management of EoE, addressing the limitations of traditional approaches including inter-observer variability, invasiveness, and delays in diagnosis. By synthesizing findings from peer-reviewed studies, we demonstrate that AI algorithms exhibit high diagnostic accuracy in recognizing subtle endoscopic features and quantifying eosinophilic tissue infiltration. Moreover, these technologies can streamline workflows, reduce dependency on manual assessments, and enhance personalized care strategies. Despite the potential benefits, challenges regarding the integration of AI into clinical practice remain, including issues of algorithmic bias, data privacy, and the need for robust validation across diverse healthcare settings. Future research should focus on multicenter studies to confirm AI’s effectiveness, explore non-invasive diagnostic alternatives, and promote the ethical application of AI to optimize patient outcomes in EoE. This review highlights AI’s transformative capacity to reshape the diagnostic landscape of EoE, underscoring the requirement for ongoing evaluation and collaboration among clinicians, researchers, and technology developers to realize its full potential within the healthcare framework.

Key Words: Artificial intelligence; Deep learning; Eosinophilic esophagitis; Esophagus; Inflammation

Core Tip: The integration of artificial intelligence (AI) and deep learning models into the diagnosis and management of eosinophilic esophagitis shows great promise in enhancing diagnostic accuracy, reducing human variability in interpretation, and facilitating timely interventions. By addressing the limitations of traditional diagnostic methods, AI technologies can streamline the clinical workflow, improve patient outcomes, and potentially transform the approach to managing this complex, chronic condition. However, ongoing research is necessary to validate these tools across diverse clinical settings and ensure their ethical application.

INTRODUCTION

Eosinophilic esophagitis (EoE) is a chronic, immune-mediated disease characterized by inflammation of the esophagus that can lead to symptoms such as dysphagia, food impaction, and long-term tissue remodeling if not properly managed[1,2]. The rising prevalence and clinical complexity of EoE have important implications for both patient care and healthcare systems, as the disease often requires multifaceted diagnostic and therapeutic approaches[3,4]. EoE is primarily caused by an abnormal immune response to environmental or food allergens in genetically predisposed individuals. Exposure to these allergens triggers a T helper type 2 (Th2)-mediated inflammatory reaction in the esophagus, leading to the recruitment and accumulation of eosinophils. These release inflammatory proteins and cytokines, such as interleukin-5 and interleukin-13, which further amplify the immune response and contribute to tissue injury, remodeling, and fibrosis. Consequently, the inflammation can cause symptoms such as difficulty swallowing and food impaction. Although the precise mechanisms remain unclear, both genetic susceptibility and environmental factors, including early-life exposures and atopic conditions, play substantial roles in the development of EoE[1]. Current diagnostic standards, including endoscopic assessment and histopathological analysis, while well-established, are limited by factors such as delayed recognition, reliance on invasive procedures, and variable inter-observer agreement[5,6]. Furthermore, EoE frequently presents in conjunction with atopic conditions such as asthma and food allergies and can exhibit subtle or absent endoscopic findings, complicating its detection and increasing the risk of progressive complications such as esophageal strictures[1,2]. Despite standardized criteria, the diagnostic process remains labor-intensive and subject to human variability, often resulting in delayed or missed diagnoses that can adversely affect long-term outcomes[4,5]. Currently, only two medications are Food and Drug Administration approved for EoE treatment, outlining the need for expanding research about diagnosis and management[7,8].

Recent advances in artificial intelligence (AI) and deep learning (DL) models have begun to influence various domains of medicine, offering new possibilities for enhancing diagnostic accuracy, automating image analysis, and supporting individualized clinical decision-making[9,10]. Within the field of gastroenterology, the application of AI to EoE is a developing area of research, aiming to address some of the inherent limitations of current practices[11]. AI-driven tools have shown high concordance with manual histopathological assessment and have demonstrated diagnostic performance in endoscopic image analysis that compares favorably to traditional clinical evaluation[12,13]. These approaches have the potential to reduce the burden on medical personnel, increase consistency in diagnostic criteria application, and facilitate earlier detection, all of which are critical for optimizing patient care[14]. However, challenges remain in terms of integrating these technologies into clinical workflows, addressing data heterogeneity, ensuring robust validation across diverse populations, and managing ethical concerns such as algorithmic transparency and patient data privacy[15,16]. Furthermore, deployment of AI enabled products or schemes requires a strict regulatory framework to ensure benefit while protecting from harm.

The primary objective of this research paper is to provide a comprehensive review of the current and emerging roles of AI and DL in the diagnosis and management of EoE. The central research question addressed is how AI-based technologies can improve diagnostic precision, facilitate timely identification, and support personalized management strategies for EoE, as well as what barriers and limitations impede their implementation. This work is based on a critical synthesis of peer-reviewed studies, systematic reviews, and clinical trials, with a focus on interpreting and contextualizing existing evidence rather than generating new experimental data[3,17]. Key methodological approaches include evaluation of AI algorithms for histological and endoscopic diagnosis, analysis of the reliability and generalizability of these findings across different healthcare settings, and source criticism with attention to potential biases and gaps in the literature[18,19].

EOE AND CURRENT DIAGNOSTICS

To comprehensively grasp the current state of EoE, it is imperative to gain a holistic understanding of its clinical manifestations, epidemiological trends, and the diagnostic obstacles encountered by healthcare professionals. This foundational knowledge is crucial for acknowledging the limitations of conventional diagnostic modalities and investigating emerging advancements, such as the incorporation of AI, to augment detection and management strategies. This section serves as a prelude to assessing existing approaches and identifying avenues for innovative solutions in the realm of EoE diagnosis.

Disease characteristics and burden

EoE has experienced a dramatic rise in its epidemiological burden over the past few decades, with recent studies acknowledging an 800% increase in prevalence compared to data from the late 20^th century. This surge can be attributed to a dual phenomenon of enhanced recognition due to advancements in diagnostic tools and physician awareness, and an actual rise in disease incidence[20,21]. Researchers have also documented EoE incidence rates ranging from 3.7 to 10 new cases per 100000 annually, with prevalence rates spanning from 22.7 per 100000 to over 100 per 100000 in high-prevalence nations such as Spain and the United States[22,23]. Furthermore, estimates suggest that up to 1 in 2000 individuals in the United States may be affected, emphasizing the substantial burden of this condition[23]. The possibility of regional disparities or underdiagnosis is evidenced by the wide geographic variability in reported cases, necessitating robust epidemiological surveillance systems to better capture the true scope of the disease[20]. Beyond these figures, contributing factors such as environmental changes, increased allergic sensitization across populations, and heightened endoscopic applications suggest a multifactorial etiology behind the observed uptrend, further justifying expanded research into the underlying mechanisms[22].

The burden of EoE is substantial, with the condition responsible for 2%-7% of all endoscopies performed for esophageal symptoms, positioning it prominently among frequent esophageal disorders encountered in clinical practice[22,24]. In adults, EoE significantly contributes to dysphagia, accounting for 12%-22% of cases[22]. Moreover, it plays a critical role in esophageal food impaction, because it underlies up to 50% of such cases in this demographic[23]. These statistics not only underline the clinical significance of EoE but also emphasize its impact on healthcare systems, which must address the growing number of affected individuals. EoE’s prominence in acute care settings further reinforces the importance of timely and accurate diagnosis, since delays could exacerbate complications and increase resource utilization. The increasing patient burden calls for optimized diagnostic protocols and the exploration of emerging tools such as AI to complement traditional methods.

The strong association between EoE and atopic disorders is a hallmark of its type 2 immune-mediated pathogenesis. Up to 75% of individuals diagnosed with EoE have coexisting atopic conditions such as asthma, eczema, allergic rhinitis, or food allergies[22]. This overlap highlights shared immunologic pathways, that could open avenues for therapeutic strategies targeting both EoE and broader allergic profiles[23]. The emphasis on multidisciplinary care becomes evident, since managing concurrent conditions can improve long-term outcomes for individuals with EoE. Furthermore, the high prevalence of comorbid allergic conditions supports the case for systematic screening for EoE in populations with complex allergic profiles. Such targeted approaches may identify at-risk individuals earlier, providing opportunities for timely interventions that could mitigate disease progression.

Delayed diagnosis remains a major concern, because the risk of developing esophageal strictures rises each year EoE is left undiagnosed. Studies have consistently shown that this risk increases by 9% annually[22,23]. The link between diagnostic delay and disease progression illustrates the shift from reversible inflammatory stages to irreversible esophageal remodeling. Chronic untreated EoE leads to subepithelial fibrosis, esophageal narrowing, and an elevated risk of food impaction, adding to the condition’s significant morbidity[22]. In addition to degrading quality of life, these complications place a strain on healthcare resources since patients require recurrent medical interventions[25]. Early diagnosis and intervention, supported by epidemiological modeling, could prevent much of this stricture-related morbidity. This finding lends further credence to the importance of initiatives aimed at raising awareness among primary care and gastroenterology professionals, as well as the integration of advanced diagnostic tools such as AI to reduce diagnostic delays and improve patient outcomes.

The heterogeneous presentation of EoE further complicates its detection, with up to 12% of patients exhibiting normal findings on endoscopy despite confirmed disease[20,25]. This phenomenon is particularly evident in early or less active cases, where the absence of hallmark endoscopic abnormalities may hinder accurate identification. Pediatric studies have shown only moderate correlation (r = 0.61) between endoscopic severity scores and histologic activity, highlighting the insufficiency of endoscopy alone for diagnosing EoE[25]. Combining endoscopic assessment with histopathology and other advanced modalities is thus critical to ensure diagnostic accuracy. Missing subtle disease features during endoscopy can lead to under-diagnosis or misclassification, delaying the initiation of treatment.

Certain endoscopic signs, such as trachealization, exudates, and longitudinal furrows, are recognized as being more common in EoE than in non-EoE populations. However, their absence in some individuals complicates diagnosis. Liacouras et al[24] report that trachealization was observed in 25% of pediatric EoE patients but was absent in non-EoE controls. Additionally, the concurrent presence of exudates, trachealization, and longitudinal furrows was documented in 16% of EoE cases but in none of the control group. These findings support the necessity of developing disease-specific scoring systems and emphasize the potential of computer-aided diagnostic platforms in improving the recognition of such features. AI applications in endoscopic pattern recognition can not only enhance diagnostic accuracy but also reduce inter-observer variability, paving the way for more consistent outcomes.

Optimal diagnostic protocols for EoE are increasingly seen as a synthesis of clinical history, comprehensive atopic assessments, and detailed analyses of endoscopic and histologic findings. Advanced technologies, including AI and DL models, hold promise for integrating these diverse data points to facilitate real-time diagnostic accuracy. Recent innovations have demonstrated tangible benefits in this area; for instance, the application of AI to advanced imaging modalities like optical coherence tomography has achieved 97% diagnostic accuracy in detecting EoE-related basal layer thickening[26,27]. Such approaches can overcome some of the limitations of traditional diagnostic methods and provide a foundation for more effective disease management strategies. As the discussion progresses toward standard diagnostic approaches, the need for collaboration across technological and clinical domains to refine EoE detection and monitoring becomes evident. Table 1 summarizes the current diagnostic and endoscopic criteria for eosinophilic esophagitis.

Table 1 Common clinical features and diagnostic criteria of eosinophilic esophagitis[1,65].

Feature	Description	Relevance
Definitive diagnosis	Histological identification of 15 or more eosinophils per HPF[65]	Gold standard for diagnosis
Endoscopic findings	Rings, linear furrows, exudates, edema[65]	Key visual indicators
Associated atopic conditions	Asthma, eczema, allergic rhinitis, food allergies[1]	Common co-morbidities, important for management
Symptoms	Dysphagia, food impaction, chest pain, reflux-like symptoms[1]	Impact on quality of life
Risk of complications	Increased risk of esophageal strictures and remodeling[1]	Importance of early diagnosis

HPF: High power frame.

Standard diagnostic approaches

The endoscopic evaluation is a crucial initial step in the diagnosis of eosinophilic EoE, providing key visual hints of prevalent mucosal alterations, including rings (trachealization), linear furrows, white exudates, and edema. In spite of its diagnostic utility in contexts, the endoscopic process on its own is taxed by very real limitations. As discussed previously, in a minority of instances, up to 12% of EoE patients can have normal-appearing endoscopy, particularly in early or mild cases, and thus visual inspection alone is a dubious diagnostic criterion[28,29]. Moreover, infiltrates in EoE are often patchy in distribution; thus, focused and large biopsy sampling is needed in order to optimize sensitivity for diagnosis. Current recommendations are to obtain a minimum of six biopsies from various esophageal segments in an attempt to minimize sampling errors and maximize yield[30]. This requirement underlines the need for procedural rigor and identifies a significant limitation: Patient discomfort and procedural burden, particularly in high-prevalence areas where the demand for endoscopy is intensive. These limitations identify the reason why endoscopy, although being an absolute necessity, would preferably be complemented by adjunctive diagnostic methods in an effort to alleviate its limitations.

Histopathologic examination of esophageal biopsies, specifically identifying at least 15 eosinophils per high-power field (HPF), remains the gold standard for diagnosing EoE. This criterion is rooted in consensus guidelines and supported by extensive pathologic studies, offering an objective measure for diagnosis[24,31]. However, this threshold is not without challenges. Interobserver variability in counting eosinophils can lead to inconsistencies in diagnostic decisions, particularly in borderline cases. The specificity of this cutoff is further questioned by its overlap with conditions like proton pump inhibitor-responsive esophageal eosinophilia, where histological eosinophilia mimics EoE but resolves with acid suppression therapy[24]. Furthermore, the heterogeneous and patchy nature of eosinophilic infiltration introduces the risk of sampling error, especially when biopsies are insufficient or not adequately obtained from affected areas. Thus, reliance on eosinophil counts alone may oversimplify disease characterization, neglecting the dynamic interplay of tissue inflammation and remodeling that defines disease progression. This highlights the necessity for revised or supplemental histologic criteria that encompass broader pathologic changes beyond peak eosinophil counts.

While endoscopy and histopathological analysis represent the cornerstone of EoE diagnosis, these traditional approaches are invasive, posing procedural risks and discomfort for patients. The need for repeated endoscopies, often required for both diagnosis confirmation and monitoring of treatment response especially in pediatric and medically complex patients exacerbates these limitations[32]. Furthermore, delays in diagnosis, whether due to under-recognition of subtle endoscopic findings or unrepresentative biopsies, extend the disease course and contribute to complications like esophageal strictures. The risk of stricture formation increases by approximately 9% annually if diagnosis is delayed[22]. Beyond the impact on patient outcomes, the reliance on tissue sampling imposes a considerable burden on histopathology services. Pathologists must manually count eosinophils, a labor-intensive task subject to subjective variability[24]. These limitations call for innovative strategies and adjunctive tools to reduce the invasiveness, variability, and resource demands of the current diagnostic paradigm while improving diagnostic efficiency.

In response to the shortcomings of standard diagnostic practices, several ancillary tools and emerging modalities have been explored for their potential to enhance EoE detection. Among these, magnifying endoscopy with narrow-band imaging and endoscopic ultrasonography have been investigated for their ability to provide better visualization of subtle mucosal abnormalities, which may be overlooked during standard endoscopy[31]. Another promising tool is the cytosponge-a minimally invasive device that collects a significant number of cells from the esophageal mucosa and has demonstrated an 83% accuracy rate in identifying active EoE in early studies[32,33]. This device offers unique advantages in sampling large surface areas, aiding in detecting active inflammation and potentially identifying molecular biomarkers. Although its use in EoE is not yet established to the same degree as in Barrett’s esophagus, the cytosponge shows potential as a diagnostic adjunct, particularly in settings where endoscopy is less accessible[33]. Another emerging technology, endoscopic impedance probes, has demonstrated high sensitivity (90%) and specificity (91%) in distinguishing active inflammatory disease by assessing mucosal integrity, providing a functional biomarker correlating closely with histological activity[32]. While these modalities are still in the developmental and validation stages, they represent promising alternatives or supplements to traditional methods, especially for cases with early or patchy disease manifestations. However, further studies in diverse and larger cohorts are necessary to establish their clinical utility and integration into routine diagnostic pathways.

In short, while traditional diagnostic approaches endoscopy and histopathology remain foundational to EoE diagnosis, their limitations stress the need for supplemental and innovative strategies. Emerging tools and AI-driven solutions hold great promise for addressing these gaps, potentially transforming the diagnostic landscape of this challenging condition.

AI IN EOE MANAGEMENT

AI is revolutionizing the diagnosis and management of EoE by enhancing accuracy, efficiency, and consistency across various diagnostic modalities. Its integration into endoscopic and histopathological evaluation holds promise for overcoming traditional limitations and enabling more personalized patient care.

AI definitions and methods

The term AI usually alludes to computer generated algorithms that are capable of learning and have problem solving abilities much like the human mind[34]. Machine learning (ML) is a field that can be taught to analyze data samples, discriminate their characteristics and then generate conclusions based on previously learned experiences[35]. Supervised ML with support vector machines involves researchers’ crafted algorithms in which they direct the system to recognize distinctive variables of concern to assess them and provide output[34]. DL is a subset of ML based on artificial neural networks capable of autonomously extracting data and generating analysis. They are often crafted as convolutional neural networks (CNNs)[34,36]. Algorithms trained in DL models can be further refined longitudinally using prospective feedback methods[37]. However, DL models are similar to black boxes where the output is reached and clear but the mechanism by which it is generated is not[38]. Much effort and research are currently invested to understand and elucidate the mystery behind these processes[39,40]. Another field progressing at a fast pace is that of “omics” including genomics, transcriptomics, epigenetic and others, where large sets of biological data offer new information about pathogenesis of diseases[41]. ML models can be quite useful in these instances where large sets of data need to be integrated, analyzed and studied[42]. Table 2 summarizes the various technical AI terms and their functions[34,35].

Table 2 Summary of artificial intelligence methodologies applicable to eosinophilic esophagitis diagnosis.

AI methodology	Description	Applications in EoE
Supervised ML	Algorithms that rely on labeled training data to learn and predict outcomes	Classifying endoscopic images, predicting eosinophil counts
DL	Subset of ML utilizing neural networks for pattern recognition	Analyzing endoscopic and histopathological data
Convolutional neural networks	Special type of DL particularly adept at image recognition	Enhancing diagnostic accuracy in endoscopic images
Random decision forest	ML algorithm combining multiple data sources for improved decision making	Integrating clinical and endoscopic data for diagnosis
Natural language processing	AI field focused on interaction between computers and human language	Analyzing electronic health records for diagnostic insights

AI: Artificial intelligence; EoE: Eosinophilic esophagitis; ML: Machine learning; DL: Deep learning.

AI models have already been validated in multiple medical disciplines. For example, in radiology modules have been created to detect diseases such as breast cancer or acute neurological events with high accuracy[43]. Furthermore, data suggests that AI significantly increases detection of colorectal neoplasia during lower endoscopy[44,45]. Similarly in upper endoscopy AI models have been shown beneficial in detecting subtle lesions and reducing variability related to an endoscopist’s experience[38]. For instance, one recent meta-analysis showed that the sensitivity and specificity in diagnosing Barrett’s neoplasia is 89% and 86% respectively[46]. In general, there is clear evidence that AI generated models perform well in gastrointestinal pathologies.

AI-enhanced diagnostic methods

AI in improving endoscopy: The use of AI in diagnosing EoE has garnered considerable attention for its potential clinical impact, particularly through enhanced diagnostic methodologies. DL algorithms offer particular value by addressing the variability often seen in EoE endoscopic presentations. Subtle features, such as esophageal rings, linear furrows, white exudates, and edema, can sometimes be overlooked by endoscopists, particularly less experienced clinicians or in atypical cases. AI models trained to recognize these features consistently enhance diagnostic accuracy and reduce reliance on individual expertise[35]. While promising, these algorithms must undergo rigorous external validation to ensure robustness across varied clinical scenarios, particularly in resource-constrained environments where diagnostic expertise may vary. Integral to the success of AI-assisted diagnostic methods is the incorporation of established scoring systems, such as the EoE endoscopic reference score (EREFS), into AI algorithms. This approach improves diagnostic capacity and facilitates objective assessment of disease activity, which is crucial for monitoring treatment responses and disease progression[44]. The integration of EREFS reflects the potential of AI to standardize assessments, although questions remain regarding its applicability in pediatric populations and individuals with overlapping esophageal pathologies. AI has demonstrated high diagnostic accuracy in analyzing endoscopic images of individuals with EoE. For instance, DL models such as the AI-EoE-EREFS algorithm have achieved sensitivity, specificity, and accuracy levels of 0.96, 0.94, and 0.95, respectively, with an area under the curve (AUC) of 0.992[44]. These figures highlight the significant reliability and potential of AI tools as adjuncts in endoscopic settings. However, the validity of these outcomes necessitates critical consideration of data sources, diversity within training datasets, and the reproducibility of results across different populations and healthcare facilities.

Okimoto et al[47] performed a very elegant study where they developed a CNN model and trained it to recognize EoE endoscopically using a set of images. The model performed very well with an accuracy of 95% when tested against patients and controls. Similarly, another trial compared EoE endoscopic appearance to normal controls and patients with esophageal candidiasis using another CNN model. The algorithm was trained on 484 endoscopic images from 134 patients. The results were compared and validated against 3 endoscopists, showing an accuracy of 91% vs 83% for the physicians[48]. Additionally, it seems that adding objective criteria to the model such as the EREFS score infers more benefit. German researchers concluded that the prediction of a CNN model improved when the endoscopic score was added to the algorithm increasing sensitivity to 85% and specificity to 95%[49]. The performance consistency of AI-EoE models, as demonstrated by high sensitivity, specificity, accuracy, and harmonic mean across different cohorts and image qualities, further supports their clinical utility. This robustness highlights the scalability of such systems; however, their long-term performance in real-world, multi-institutional settings remains a crucial area for further exploration. The extent to which these technologies can be seamlessly integrated into routine workflows, particularly in lower-resource environments, also warrants deeper investigation.

AI has also facilitated the development of random decision forest models that integrate both clinical and endoscopic data to enhance diagnostic precision. These models have achieved impressive AUCs of 0.90 with clinical data alone and 0.94 when combining clinical and endoscopic data, demonstrating their utility in improving patient selection and biopsy guidance in real-world settings[50]. However, the complexity of these multi-source systems may pose challenges to their seamless integration into routine practice. One major advantage is that the predictive accuracy of ML models improves with the inclusion of readily accessible clinical variables alongside endoscopic findings, enabling better prioritization for invasive procedures[50]. Such advancements could streamline patient management, however, disparities in healthcare infrastructure and diagnostic capacity must be acknowledged, as these factors could affect the performance of AI algorithms in underrepresented settings. Table 3 summarizes the main AI models applied to endoscopic diagnosis of EoE, their methodology and key limitations.

Table 3 Comparative table of the main artificial intelligence models applied to endoscopic diagnosis of esophagitis diagnosis.

Model	Training dataset	Population size	Validation status	Key limitations	Methodological rigor
AI-EoE-EREFS (DL/CNN)[48]	Endoscopic images, EREFS-labeled, multi-center	484 images from 134 patients	Retrospective, some external validation	Retrospective data, limited diversity, mainly binary classification, limited real-world scenarios	High accuracy (AUC = 0.992), but generalizability and real-world robustness uncertain
CNN[47]	Endoscopic images (EoE vs controls)	1192 characteristic endoscopic images of 108 patients	Tested against patients and controls	Dataset size not reported, unclear diversity, retrospective design	High accuracy (95%), but limited transparency and unclear reproducibility
CNN with EREFS integration[49]	Endoscopic images annotated with EREFS	200 WLIs, including 100 WLIs from EoE patients and 100 WLIs of normal esophagus	Compared to human experts	Training/test data overlap possible, unclear external validation	Sensitivity improved to 85%, specificity 95% with EREFS
Random decision forest[50]	Clinical + endoscopic data	Not specified	Real-world data, AUC up to 0.94	Complexity, integration challenges, population diversity not detailed	Robust with multi-source inputs, but needs more transparent validation
General endoscopic AI (DL/CNN)	Mixed datasets (WLI, NBI, CT images)	Varies; often small	Mostly retrospective, single/multi-center	Data heterogeneity, lack of multi-center, prospective validation	High reported accuracy, but shallow validation, risk of bias due to dataset curation

AI: Artificial intelligence; EoE: Eosinophilic esophagitis; DL: Deep learning; EREFS: Eosinophilic esophagitis endoscopic reference score; CNN: Convolutional neural network; WLI: White light image; NBI: Narrow band imaging; CT: Computed tomography; AUC: Area under the curve.

AI in improving histology: AI has also shown remarkable promise in the automated histological evaluation of tissue biopsies. A DL approach achieved near-perfect diagnostic performance in one study, demonstrating 99.0% accuracy, 100% sensitivity, and 98.2% specificity in identifying ≥ 15 eosinophils per HPF in esophageal biopsies[24]. These results highlight the advanced precision of AI in meeting diagnostic thresholds.

AI-driven eosinophil counting has demonstrated excellent concordance with pathologist counts, achieving a Pearson correlation coefficient of 85% in gastrointestinal biopsy analyses[51]. This accuracy supports the broader adoption of AI as a quality control instrument in histopathology. Automated counting methods significantly reduce errors and alleviate fatigue-related inconsistencies in human reporting, fostering efficiency in high-volume pathology environments[51]. The faster turnaround times enabled by automated methods could lead to more prompt clinical decisions and interventions, further supporting the utility of AI tools for improving patient outcomes in EoE[51]. Nevertheless, ensuring consistent performance across diverse healthcare systems remains an ongoing challenge. Automated eosinophil quantification further enables comprehensive mapping of disease heterogeneity, an innovation that holds promise for distinguishing between phenotypic variants of EoE and predicting treatment responses[24]. This capability could help clinicians refine personalized management strategies. In general experts agree that CNN models can reliably take the place of manual eosinophils count.

Recent analytical advancements indicate that focusing solely on peak eosinophil counts may overlook critical diagnostic features of EoE. ML and DL analyses have offered innovative insights into global histopathologic changes, such as basal cell hyperplasia, spongiosis, and lamina propria fibrosis, which may help differentiate active EoE from non-EoE states[31]. This suggests a critical opportunity to expand histologic assessment frameworks by integrating both quantitative (eosinophil counts) and qualitative (global tissue features) analysis. Leveraging AI-driven models may enhance diagnostic consistency, decrease inter-observer variability, and allow for a more comprehensive assessment of disease activity. Importantly, adopting AI platforms for biopsy-based diagnostics may enable automated and reproducible analyses, addressing issues related to subjectivity. This advancement represents a significant step toward a more nuanced and technologically advanced approach to EoE evaluation. Looking further than eosinophilic count, pathologists have elaborated an EoE histologic scoring systems to include other important criteria such as fibroblasts in the laminated propria[52]. Additional characteristics beside eosinophil counts reliably predicted EoE in CNN models with an accuracy of 85%[53].

Dynamic CNNs have successfully classified esophageal tissue biopsies, achieving accuracy, sensitivity, and specificity rates of 85%, 82.5%, and 87%, respectively[54]. Notably, these systems incorporate global histopathologic features such as tissue architecture rather than solely relying on eosinophil counts, thus strengthening their diagnostic capacity. Downscaling images and utilizing image patches in dynamic CNNs has enabled these models to function effectively with real-world, heterogeneous data, maintaining true positive rates of up to 82.5%[55]. This adaptability is crucial for integrating AI tools into everyday pathology workflows, but challenges such as ensuring hardware compatibility and optimizing user interfaces for pathologists demand attention. ML models excel at capturing subtle histological changes beyond peak eosinophil counts, which traditional diagnostic thresholds might miss. Such nuanced analyses are particularly advantageous in cases where standard features fail to fully characterize disease activity[55].

Larey et al[56] developed an ML model incorporating not only eosinophilic counts of > 15/HPF but also basal cell fraction. They used 1066 biopsies to train their model. This algorithm predicted severity more accurately than the standard method (accuracy of 87% vs 79%). This carries additional importance in addressing treatment response and remission[56]. Similarly, Ricaurte Archila et al[57] presented a CNN model to evaluate histologic features of EoE and compared the results to five experienced pathologists. The AI model analyzed several pathologic features and faired well in comparison to the pathologists scores (94.5% vs 94.8%). Another important study researched an image classification model looking at multiple features other than the eosinophils count. It was able to predict the diagnosis of EoE in 99% of slide patches from treatment naive patients, the model was also useful in prediction of remission after follow up[58]. Even more impressive is the DL digital pathology model that incorporated novel features of EoE looking at the whole biopsy histology. It examined features such as tissue subregions, eosinophilic abscesses, cell types and even the spatial relationships between various cells[59]. These models offer the possibility of diagnosing difficult and unconventional patients such as those without elevated eosinophils or those maintained on chronic proton pump inhibitors.

Dunn et al[60] applied an AI clustering analysis on Th2 cytokine expression resulting in major breakthrough in EoE patients subtypes and classifications. They were able to delineate 5 groups of endotypes of patients. To note that all of these patients did not differ in their eosinophil count further confirming the theory peripheral eosinophil count may not be enough for diagnosing EoE patients. Group 1 showed the lowest interleukin-5 expression vs group 5 which showed the highest (suggesting an elevated inflammatory load). Groups 2 to 4 showed elevated levels of interleukin-13 and tendency for more fibrostenotic disease[60]. It is still unclear whether these subtypes result from a continuous progression in disease severity or distinct patients phenotypes.

The application of DL for histopathological analysis effectively mitigates longstanding issues of inter-observer subjectivity in manual eosinophil counting. Pathologist variability, particularly in borderline cases, has historically hindered consistent diagnoses. Automated systems not only provide reproducibility but also offer scalability in high-volume pathology labs, addressing critical workflow bottlenecks. The high specificity and sensitivity achieved with AI-based histopathological tools highlight their potential in supporting overburdened histopathologists, who face increasing demand for esophageal biopsies[24]. Nevertheless, implementing such technology requires significant investment, and its adaptability to small-scale or underfunded pathology centers is yet to be fully assessed. Employing sufficiently large and diverse training datasets has significantly improved the adaptability of AI models to various clinical conditions and institutions[38]. Models that are generalizable across populations are essential for equitable healthcare delivery. While the reported AI performance metrics such as 99% accuracy appear impressive, it is important to interpret these results with caution. Such high figures may inadvertently signal potential overfitting, especially if the model has primarily been evaluated on limited or non-representative datasets. Overfitting occurs when a model performs exceedingly well on training data but fails to generalize effectively to new, real-world scenarios. Additionally, models trained and validated on small sample sizes or narrowly defined datasets may embed underlying biases, leading to performance gaps when exposed to broader populations or diverse input conditions. This underscores the need for thorough testing across varied, real-world datasets and the inclusion of robustness checks to detect and mitigate bias. Figure 1 illustrates the some of the available AI driven innovations and their performance in EoE diagnosis.

Figure 1 Performance of emerging technologies and tools in eosinophilic esophagitis diagnosis[24,44,54]. ML: Machine learning; DCNN: Digital convolutional neural network.

AI in improving biomarkers: Point-of-care decision support enabled by ML models could drastically reduce unnecessary biopsies while ensuring that high-risk individuals are not overlooked[36]. ML applications that utilize peripheral blood markers, such as eosinophil levels, have demonstrated strong diagnostic associations in EoE and other gastrointestinal conditions[53]. Expanding these capabilities could aid in streamlining the diagnostic process; although, their accuracy and reliability when applied to complex, multi-disease scenarios require further study. The inclusion of auxiliary markers, such as platelet-lymphocyte ratios and white blood cell counts, enhances the predictive power of diagnostic models[53]. By considering these multifactorial inputs, AI could offer better decision-making insights for clinicians managing complex cases. Additionally, non-invasive, AI-driven analysis of peripheral blood data could serve as an effective triage tool, reducing reliance on invasive procedures. This approach has significant implications for populations with limited access to endoscopic services; nonetheless, further validation is crucial to ensure that it does not compromise diagnostic accuracy.

Multiple studies have explored non-invasive AI facilitated diagnostic tools ranging from swallowed strings to sponges. They capture eosinophils-derived proteins, but they still seem to offer lower accuracy in diagnosis that the EREFS score[61,62]. A more innovative approach is the new swallowable capsule that can identify eosinophils in the esophagus through reflectance microscopy. It is swallowed and then retrieved using a string. A CNN model was created to analyze the generated images and resulted in an 86% accuracy in diagnosis[63]. Wechsler et al[64] used an AI random forest model analyzing data from a combined six eosinophils associated blood markers and two urine markers. It accurately predicted the diagnosis 84% of the times.

CONCLUSION

This review aimed to assess the transformative potential of AI and DL models in the diagnosis and management of EoE, a chronic, immune-mediated esophageal disease characterized by its increasing prevalence and presenting significant diagnostic challenges. The primary objective was to determine whether AI-based methodologies can provide substantial improvements over conventional diagnostic paradigms and to evaluate the extent to which technological advancements overcome existing barriers in clinical practice. By synthesizing current literature and emerging evidence, the work systematically evaluated the limitations of conventional approaches, the efficacy of advanced computational tools, and the broader implications for disease detection and patient care. Standard diagnostic modalities, such as endoscopy and histopathology, are fundamental but are limited by significant shortcomings. Inter-observer variability, the invasiveness of procedures, and the potential for diagnostic delays substantially increase morbidity. Data presented demonstrate that AI-driven algorithms now attain high sensitivity and specificity in both endoscopic pattern recognition and histopathologic image analysis. Advanced models reliably identify subtle mucosal abnormalities and precisely quantify eosinophilic infiltration, resulting in outcomes that are comparable to or surpass those of expert human evaluators. Furthermore, these systems introduce unprecedented scalability, facilitate more uniform reporting across institutions, and offer the potential to streamline patient pathways by integrating multifactorial clinical data. Notably, evidence supports that the integration of AI into diagnostic workflows can diminish reliance on resource-intensive manual assessments and may facilitate more individualized disease monitoring. Within the ongoing research landscape, the findings presented in this review align with and complement current trends in the field of digital medicine. This underscores the exceptional suitability of AI solutions for intricate, multifaceted diseases such as EoE. The empirical evidence reviewed supports the notion that DL models demonstrate remarkable proficiency in integrating diverse diagnostic indicators that might otherwise be imperceptible to conventional approaches. Additionally, they also raise pertinent questions pertaining to standardization, equity, and clinical translation. It is important to acknowledge several limitations that define the current state of knowledge. A notable constraint is the reliance on secondary literature, which, while comprehensive, limits the capacity for direct experimental validation. Furthermore, variation in patient populations, training datasets, and healthcare resources may impact the generalizability of AI algorithms across different clinical environments. One such examples would be applicability in low-income countries where resources are scarce and funding for healthcare is limited. There remain unresolved questions related to algorithmic bias, data privacy, interpretability, and the seamless integration of sophisticated computational tools into everyday practice. Future research should prioritize large-scale, multi-center studies to validate the robustness and clinical effectiveness of AI-powered diagnostic solutions, foster cross-disciplinary collaborations, and advance the development of transparent AI models that are both explainable and adaptable. Additionally, expanding the application of AI to non-invasive diagnostic pathways and individualized treatment planning represents a crucial frontier for optimizing outcomes in EoE and similar chronic diseases. Figure 2 suggests an algorithm detailing where AI may play a major role. The review not only confirms the capacity of AI to revolutionize the management of EoE but also stresses the importance of ongoing critical appraisal, ethical vigilance, and continued methodological refinement. As research in this field continues to evolve, fostering a collaborative environment among clinicians, researchers, and technologists will be essential to harness the capabilities of AI effectively. In doing so, we can optimize diagnostic pathways, enhance patient outcomes, and ultimately redefine the landscape of care for individuals suffering from EoE. The convergence of traditional practices with cutting-edge technology heralds a new era in the management of this chronic condition, paving the way for both increased efficacy and improved quality of life for patients.

Figure 2 Schematic representation of the integration points for artificial intelligence algorithms in the diagnostic pathway for eosinophilic esophagitis. Artificial intelligence (AI) augments the traditional workflow at multiple stages: (1) During endoscopic evaluation, AI models assist in real-time image analysis, enhancing detection of subtle mucosal features and standardizing scoring (e.g., eosinophilic esophagitis endoscopic reference score), and can guide targeted biopsy sampling; (2) In histopathology, AI automates eosinophil counts and evaluates global tissue features, reducing inter-observer variability and expediting diagnosis; and (3) AI-driven biomarker analysis and triage tools (using blood, urine, or non-invasive devices) further support risk stratification and may reduce the need for invasive procedures. These integrations aim to increase diagnostic accuracy, efficiency, and consistency, ultimately supporting more timely and individualized patient management. AI: Artificial intelligence; EREFS: Eosinophilic esophagitis endoscopic reference score; HSS: Histologic scoring system.