INTRODUCTION
The use of artificial intelligence (AI) in gastroenterology is possibly one of the most significant recent technological innovations in modern medicine. AI refers to systems that can execute tasks that typically require human-like reasoning. In medicine, AI technology includes machine learning (ML), deep learning (DL), natural language processing, and computer vision and enables the analysis of extensive and complex clinical information. Thus, AI technology is a health management tool that can be applied to diagnosis, treatment prediction, and the delivery of medical services. It is an efficient tool for revealing useful new data patterns that can assist physicians[1,2].
In recent years, the utilization of AI technology in gastroenterology has proven to be very promising. Digestive diseases produce vast amounts of diverse data, including endoscopic images, histology results, radiologic studies, laboratory tests, and clinical data. Given its ability to process complex data sets, AI, particularly DL technology, is well suited to this branch of medicine. There is a growing body of evidence highlighting the important role of AI technologies in detecting and staging gastrointestinal lesions, such as colon polyps, early gastric cancer, and tumors of the biliary tract. In endoscopic procedures, computer-assisted detection and diagnosis systems can provide instant feedback during the procedure, leading to improved lesion detection rates and diagnostic precision, as well as decreased dependence of the diagnosis on the operator[3,4].
AI usage extends far beyond gastrointestinal endoscopy, as it can be applied to various aspects of gastroenterology, such as assessing the risk of colorectal cancer, performing non-invasive evaluation of liver fibrosis, predicting the prognosis and treatment of inflammatory bowel disease (IBD), and improving patient care through predictive analytics. AI models can thus contribute to innovative treatment strategies by utilizing data from different modalities, such as imaging and clinical features coupled with molecular data[5].
Nevertheless, despite encouraging advancements in this area, the implementation of AI technology in clinical practice is still associated with multiple challenges. Problems related to the quality of data used for training AI systems, external validation of AI systems, algorithm transparency, regulatory approval, and ethical implications still prevent AI from being widely used in practice. In addition, many existing AI technologies are based on retrospective designs and need to be validated prospectively before incorporation into clinical practice, resulting in delays in their clinical application[6].
Taking into account the rapid expansion of knowledge regarding the application of AI technology, it is important to synthesize the current knowledge in this field. Therefore, the purpose of the current review is to provide updated information on the application of AI technologies in digestive tract diseases. The review aims to summarize the basic principles of the application of AI in clinical practice and research, analyze current data, and discuss perspectives, challenges, and limitations regarding the use of AI technologies for diagnosing, treating, and prognosticating gastrointestinal diseases.
METHODOLOGY
A comprehensive search of the literature was conducted through PubMed/MEDLINE, Scopus, EMBASE, and Web of Science to identify pertinent studies on the use of AI in gastroenterology. The search was performed using a combination of controlled vocabulary (MeSH terms) and free-text keywords referring to “AI in Healthcare,” “AI in Gastrointestinal Endoscopy,” “AI in Gastrointestinal Oncology,” “AI in Liver Diseases,” “AI in Pancreatic Diseases,” “AI in Inflammatory Bowel Disease,” “AI in Functional and Motility Disorders,” and “AI in Microbiota.”
Additional searches were conducted by reviewing the reference lists of relevant original papers, meta-analyses, and reviews and by tracking citations of key papers to identify additional relevant studies. The focus was placed on randomized controlled trials, prospective and retrospective observational studies, systematic reviews, and meta-analyses, with an emphasis on the most recent and reliable evidence. However, pivotal studies and prior publications were also included to provide background and mechanistic information.
Eligible studies were chosen based on their relevance to the scope and aim of the review. We did not define a systematic review protocol. There was no assessment of study-level risk of bias or data synthesis because the main goal of the work was to provide a qualitative and narrative overview of the existing evidence. The findings from the studies were interpreted in accordance with the applied study design and level of evidence, with particular consideration given to the difference between mechanistic and clinical (experimental) evidence. The literature search was conducted immediately before submission of the article (May-June, 2026) to consider all recently published evidence.
FUNDAMENTALS OF AI IN HEALTHCARE
The healthcare sector has found AI to be a valuable tool. With the rise of AI, it has become possible to study vast amounts of complex biomedical data that can assist in medical decision-making, disease detection, and prediction of treatment outcomes. AI systems are able to accomplish tasks that usually require human intelligence, such as recognizing patterns, reasoning, and making predictions. In medical applications, AI uses medical data to make predictions and classifications that facilitate the work of healthcare professionals who care for patients. Developments in digital health data, computing power, and algorithms have accelerated the use of this technology in different areas of medicine, such as gastroenterology[7].
AI is a broad field that includes various computational methods that carry out cognitive tasks, including learning, problem-solving, and decision-making. In this context, ML is one of the most important subdivisions of AI and involves the development of algorithms that learn from data patterns automatically without being programmed for every specific computing task. ML algorithms employ different types of statistical methods to identify correlations in data and enhance their predictive capabilities through iterative training[7,8]. A more advanced type of ML is DL. It is based on artificial neural networks that consist of multiple processing levels. These deep neural networks are used for learning and extracting various types of information from raw data. While DL has proven effective in solving many tasks, such as image recognition, signal processing, and natural language processing, in medicine it is applied specifically to image analysis and disease detection based on hidden patterns in images[9,10].
From a conceptual viewpoint, AI, ML, and DL can be considered related branches. AI is the broadest area, with ML being a subset of AI and DL a subset of ML[9]. There are many types of AI systems that can be used in medical research and practice. One of the most commonly employed methods in medicine is the supervised learning framework. In this system, algorithms are trained on labelled data, in which input information (for instance, patient or imaging data) corresponds to known results[9,11].
Unsupervised learning is a method that uses data sets with no pre-established labels and seeks to identify the underlying structure of the data. This approach enables the identification of various disease subtypes, the discovery of new markers, and the study of complicated biological interactions[11]. Another method is reinforcement learning. Reinforcement learning algorithms flexibly adapt their strategies to enhance positive outcomes via continuous environmental feedback. Clinical applications of this procedure are still in their infancy; however, its potential to augment therapeutic strategies and support adaptive decision-making is considerable[11]. In addition to these approaches, several specialized DL models have potential clinical applications. For instance, convolutional neural networks (CNNs) excel in medical imaging tasks.
The availability and diversity of data provide a rationale for implementing AI technology in healthcare. Gastroenterology generates various types of data, making this discipline particularly suitable for the application of AI in medicine. Much of these data are related to medical imaging. Different endoscopic techniques, such as colonoscopy, upper gastrointestinal endoscopy, capsule endoscopy, and endoscopic ultrasound, generate large amounts of visual data. Gastroenterology clinics can apply advanced ML technologies to analyze images and videos to identify and classify lesions. For example, a study by Kader et al[12] highlighted the application of CNNs to characterize polyps and demonstrated that the system was capable of distinguishing adenomas from sessile serrated lesions and hyperplastic polyps[12]. Data obtained from histopathology are also important. The development of digital pathology marked a new era by converting histopathological slides into usable digital images. These systems can assist pathologists in identifying dysplastic tissue and grading tumors. Moreover, omics data have increasingly been incorporated into clinical and imaging studies in AI research to identify molecular signatures of disease susceptibility, disease course, and response to therapy. Overall, the growth of AI methods, together with the wider availability of different biomedical data sets, has created many possibilities for improving diagnosis and patient care.
AI IN GASTROINTESTINAL ENDOSCOPY
The use of AI in gastrointestinal endoscopy is a promising field. There is a vast volume of visual data to interpret from endoscopic procedures, making them suitable for DL and computer vision algorithms. AI systems are increasingly being used to assist endoscopists with lesion diagnosis and clinical decision-making. The main current applications include computer-aided detection (CADe) systems used to detect lesions and computer-aided diagnosis (CADx) systems used for lesion diagnosis. In addition, AI technologies are progressing quickly in upper gastrointestinal endoscopy and capsule endoscopy, increasing the effectiveness and accuracy of these procedures.
Colonoscopy is the gold standard for colorectal cancer screening; however, a substantial number of colorectal polyps may be missed during the procedure because of endoscopist fatigue, differences in skill levels among specialists, and lesion morphology. CADe technology has been developed to address these issues. CADe systems usually consist of deep neural networks, mostly CNNs, that analyze endoscopic videos in real time and indicate polyps with visual signals. To illustrate, a systematic review and meta-analysis showed that AI-assisted colonoscopy markedly improved both the polyp detection rate and adenoma detection rate (ADR) compared with traditional colonoscopy[13]. More up-to-date evidence from randomized controlled trials that included more than 11000 patients confirmed that CADe systems enhance the ADR. In a meta-analysis by Shiha et al[14], the ADR was greater with AI-assisted colonoscopy than with conventional colonoscopy (41.4% vs 33.0%), with improvements seen across different polyp sizes, types, and anatomical locations. The adenoma miss rate is considered an important quality indicator in colorectal cancer screening. Literature data show that AI-aided colonoscopy has been useful in decreasing this rate. A recent meta-analysis indicated that CADe systems may help reduce missed adenomas and increase the detection of small or inconspicuous lesions[15]. Furthermore, Yamada et al[16] showed that AI technologies may have 97% sensitivity and 99% specificity in polyp detection. Based on this research, it may be inferred that CADe systems have the potential to improve colonoscopy quality and possibly contribute to a decrease in colorectal cancer occurrence through early detection of disease precursors.
In addition to being used for detection, AI techniques are actively used for characterization of polyps, which involves determining whether lesions detected during colonoscopy are neoplastic or non-neoplastic (CADx). CADx uses ML algorithms applied to a number of endoscopic features, including blood vessels, mucosal texture, and lesion shape. The aims of CADx systems include assisting with real-time optical biopsy, which enables endoscopists to determine the histological diagnosis of lesions without taking samples. Accurate identification of small polyps may facilitate endoscopic management with or without biopsy, saving time and reducing costs. Several studies on the accuracy of CADx systems have indicated that this approach allows lesions to be properly distinguished. AI models have shown high sensitivity and specificity in distinguishing adenomatous from hyperplastic polyps, approaching the performance of professional endoscopists[13]. According to a study from Japan, AI was successful in detecting malignancy in colorectal lesions (97% sensitivity and 100% specificity)[17]. However, there are still certain issues related to the implementation of CADx systems.
In upper gastrointestinal endoscopy, AI has been considered useful, especially in the identification of dysplasia due to Barrett’s esophagus and early gastric cancer. The diagnosis of these conditions may be difficult because the lesions can be subtle and easily missed during routine examinations. DL algorithms developed using large data sets derived from endoscopic images can identify mucosal lesions related to early neoplasia. Trials have shown that AI systems may detect esophageal and gastric cancers with high accuracy; in some instances, these systems have outperformed non-specialist endoscopists. For instance, AI models have shown diagnostic accuracy close to 99% for detecting specific upper gastrointestinal cancers[18]. AI algorithms can locate dysplastic sites that should be biopsied during Barrett’s esophagus screening[19-21]. AI technologies have also been developed to identify early-stage gastric cancer, gastric atrophy, and intestinal metaplasia, which are precancerous conditions[22-24]. One study conducted in Japan demonstrated the ability of a CNN to detect esophageal cancer with 98% sensitivity[25]. In the same year, another study conducted in China on deep neural networks for diagnosing esophageal cancer reported 98% sensitivity and 85% specificity[26]. Subsequently, Liu et al[27] reported that the sensitivity of a CNN was 94% and the specificity was 95%. According to the study by Tokai et al[28], AI accuracy in detecting esophageal squamous cell carcinoma was 96%. Studies on CADe in gastric cancer have demonstrated the potential of AI. According to Hirasawa et al[29], the CNN examined 2296 images in 47 seconds, correctly identifying 71 of 77 gastric cancer lesions and achieving a sensitivity of 92%. Wu et al[30] studied deep neural networks for gastric cancer detection, achieving 92.5% accuracy, 94.0% sensitivity, and 91.0% specificity in differentiating early gastric cancer from benign lesions. Moreover, a multicenter case-control study reported that AI demonstrated diagnostic accuracy similar to that of experienced endoscopists in upper gastrointestinal cancer detection, with 96% accuracy, surpassing the results of skilled and junior endoscopists[31]. Li et al[32] also reported sensitivity, specificity, and accuracy of the CNN for early gastric cancer diagnosis of 92%, 91%, and 91%, respectively. These studies show that AI can play an important role in the diagnosis of upper gastrointestinal diseases.
Capsule endoscopy is now used for the evaluation of small intestinal diseases such as obscure gastrointestinal bleeding, Crohn’s disease, and small intestinal tumors. However, the interpretation of capsule endoscopy procedures takes considerable time because a single procedure generates a large number of images[33,34]. Spada et al[35] studied the usefulness of AI in capsule endoscopy in a multicenter clinical trial in 2024 and reported that AI-assisted evaluation was more effective than conventional interpretation, especially for identifying lesions in the small intestine. AI algorithms have been introduced for the evaluation of capsule endoscopy images and recognition of pathologies such as bleeding, ulcers, growths, and tumors. For example, research by Ribeiro Andrade et al[36] found that conventional and AI-based readings in capsule endoscopy demonstrated different sensitivities for identifying ulcers and erosions in the small intestine (90% vs 70%). DL methods can process many images in a short time, thus allowing medical professionals to save considerable time while improving diagnostic accuracy[33,34,37]. Recent research shows that AI systems can successfully identify gastrointestinal lesions on capsule endoscopy[38]. These systems can also support both lesion localization and identification[39]. In a study by Klang et al[40], AI was tested for the detection of small-bowel ulcers in patients with Crohn’s disease, and the findings revealed that AI achieved 95%–97% accuracy in detecting such lesions in the study cohort. These studies support the use of AI in gastrointestinal endoscopy.
AI is increasingly being applied to the diagnosis of gastrointestinal bleeding. DL algorithms are being used to detect luminal blood, vascular lesions, ulcers, and other small-bowel abnormalities, potentially reducing missed lesions and shortening reading time. A CNN trained on 22095 capsule endoscopy images detected blood with 98.6% sensitivity and 98.9% specificity[41]. Another DL model analyzing 53555 images achieved 88% sensitivity and 99% specificity for identifying small-bowel lesions and their hemorrhagic potential[42]. More recent models have demonstrated high accuracy for detecting active bleeding and vascular lesions[43]. However, prospective multicenter validation and integration into routine clinical workflows remain necessary.
AI IN GASTROINTESTINAL ONCOLOGY
AI has become a powerful tool in gastrointestinal oncology, improving cancer detection, risk assessment, histopathological evaluation, and prognostic prediction. Gastrointestinal cancers such as colorectal cancer, gastric cancer, esophageal cancer, and pancreatic cancer are still important causes of cancer morbidity and mortality worldwide. The rapid growth of digital health data, advanced computational algorithms, and large clinical datasets has made AI more useful in cancer research. AI systems can use various types of data, such as imaging, histopathology, genomic data, and electronic health records, for more accurate diagnosis and personalized treatment strategies[44,45].
A critical factor in decreasing CRC mortality is early detection of cancer. Colonoscopy is the most effective tool for CRC screening. However, it has some pitfalls, such as operator inconsistency, missed lesions, and subtle mucosal changes. AI-based automated image analysis and CADe can help physicians detect colorectal cancer at early stages. According to the results of a randomized controlled trial, an AI detection method led to the detection of 11% more adenomas than in patients screened without AI support[46]. Abnormal mucosal features suggest adenomas or cancer at early stages, and DL algorithms (especially CNN systems) can analyze colonoscopic images for the presence of such abnormalities. A systematic review and meta-analysis compared the detection of adenomas using AI techniques with traditional approaches, and the results showed that AI increased the ADR and decreased the percentage of missed colorectal lesions[47]. Moreover, other studies showed that different DL systems can be used in diagnostic procedures because they were able to accurately identify colorectal polyps and early-stage lesions[48,49]. Other non-invasive imaging studies may also use AI in diagnostic procedures[50].
Another application of AI in gastrointestinal oncology is risk evaluation. AI is capable of analyzing large amounts of information collected in clinical practice and identifying relationships between risk factors and cancer. Specialized algorithms analyze electronic health records and evaluate the risk of colon cancer. These algorithms take into account factors such as age, heredity, lifestyle, metabolism, and laboratory values. Recent studies showed that AI improved targeted screening of risk groups because it could determine the risk of colon cancer significantly before diagnosis[51]. Assessing the risk of various types of gastrointestinal cancer, including gastric cancer and pancreatic cancer, may be possible with the same technology. These algorithms can identify patients who have a higher risk of being diagnosed with cancer by analyzing clinical, imaging, and genomic data[52].
Histopathological examination is still considered the benchmark for diagnosing gastrointestinal cancers. Nonetheless, analyzing histological slides can be laborious and subject to interobserver variation. Modern digital pathology has permitted the digitization of histological slides and their analysis using AI algorithms, especially DL systems. AI-based histopathological systems are capable of determining tumor regions, identifying tumor types, and detecting relevant morphological features. For colorectal cancer, DL systems have shown excellent capabilities in discriminating between malignant and benign tissue and detecting histological features associated with tumor development[53]. Furthermore, AI algorithms are able to measure subtle morphological changes in the tumor microenvironment that may not be obvious to human observers[50].
The utilization of AI systems in prognostic evaluation and treatment allocation in gastrointestinal oncology is on the rise. Models built using ML techniques can assimilate various types of data, including clinical data, imaging data, pathological information, and genomic data, to predict patient prognosis in terms of survival, risk of recurrence, and treatment response, among others. Numerous investigations have shown that AI prognostic models can perform better than traditional tumor staging methods for predicting the prognosis of patients with colorectal cancer. For example, ML approaches combining molecular data and data from histopathological analysis have improved the accuracy of predicting overall survival and recurrence risk[53]. Furthermore, AI technologies are used in precision oncology and can determine molecular tumor types and predict drug response. By combining multi-omic datasets with clinical information, AI can separate patients into various groups based on their biological properties and identify appropriate therapies based on these characteristics[52].
AI is increasingly being used in gastrointestinal oncology for improved early cancer identification, risk assessment, histopathological examination, and prognostic modeling. AI-based technologies have shown good progress in colorectal cancer screening, computational pathology, and personalized treatment planning.
AI IN LIVER DISEASES
In recent years, hepatology has witnessed a growing interest in AI, which has proved effective in processing complex clinical, imaging, and molecular data. Chronic liver diseases such as metabolic dysfunction-associated steatotic liver disease (MASLD), severe fibrosis, cirrhosis, and liver cancer represent serious global health problems. Early diagnosis and risk stratification are challenging because many liver diseases are asymptomatic until they reach an advanced stage. The use of AI in medical diagnostics has become more common because the technology is capable of improving non-invasive diagnostic techniques and approaches to disease classification, monitoring, and prognosis.
Proper classification of liver fibrosis is crucial for treatment, enabling physicians to define the appropriate treatment course and estimate the prognosis in patients with chronic liver disease. The conventional method for assessing liver fibrosis is liver biopsy. However, despite being the gold standard, it is invasive, expensive, and subject to a high degree of variability stemming from sampling techniques. Non-invasive methods of assessing liver fibrosis are becoming more common in practice[54,55]. ML algorithms are being developed for the analysis of imaging methods such as ultrasound, computed tomography (CT), magnetic resonance imaging (MRI), and elastography, with a view to predicting the level of fibrosis. Certain studies indicate that ML-based models utilizing imaging data together with clinical indicators are highly effective in detecting severe fibrosis and cirrhosis. For example, studies indicate that ML methods may greatly improve the diagnostic accuracy of transient elastography and other non-invasive tests of fibrosis when used in combination with clinical and laboratory data[56,57]. In addition, AI radiomic approaches can extract quantitative information from imaging data that cannot be seen with the naked eye. This information can be used to create models capable of correctly differentiating various levels of fibrosis. It appears that the use of AI technologies in image analysis will significantly increase the accuracy of non-invasive identification of fibrosis levels[56,57].
MASLD has become the most common chronic liver disease globally. Some patients develop metabolic dysfunction-associated steatohepatitis, which may progress to fibrosis, cirrhosis, and hepatocellular carcinoma (HCC). Nevertheless, identifying patients with progressive forms of the disease continues to be difficult[58,59]. Recent literature has examined the potential use of AI to detect and stage MASLD/metabolic dysfunction-associated steatohepatitis. ML models can combine demographic data, metabolic markers, imaging studies, and laboratory data to identify patients who are prone to serious disease. For example, recent reports have shown that AI-based algorithms demonstrated better results in predicting steatohepatitis and fibrosis than traditional scoring systems[60].
Ultrasound and MRI have been combined with DL approaches and have produced promising results for the detection of liver steatosis. Even at the initial stages of the disease, these models are capable of detecting hidden patterns that indicate the development of MASLD[61]. In addition, liver biopsy evaluations in metabolic dysfunction-associated steatohepatitis have been enhanced with the assistance of AI in histopathological analysis. These algorithms assess steatosis, inflammation, ballooning, and fibrosis, thereby reducing interobserver variability in histopathological evaluation[62].
HCC is the most common malignancy of the liver. It develops mostly in patients with chronic liver disease and cirrhosis. Timely diagnosis is crucial because treatment methods resulting in the best outcomes (surgery, transplantation, or ablation) are most effective at early stages of cancer[63,64]. AI is broadly applied in medical care to facilitate HCC diagnosis and monitoring. Early tumor lesions can be identified by trained DL models that evaluate ultrasound, CT, and MRI scans. According to recent literature, AI-based algorithms can evaluate the development of hepatic tumors and differentiate HCC from benign liver lesions[65,66]. Moreover, tumor shapes and textural properties can be investigated using AI-assisted radiomics, allowing tumor aggressiveness to be predicted[67]. To improve the detection of HCC in high-risk groups, recent research has incorporated imaging, laboratory indicators such as alpha-fetoprotein, and clinical information into ML models[64,68].
The final phase of liver disease is cirrhosis. It involves a high rate of complications such as portal hypertension, liver failure, and liver-related death. For effective treatment and prioritization for liver transplantation, accurate prediction of disease development and clinical outcomes is necessary. Commonly used traditional approaches include the Model for End-stage Liver Disease scoring system, which may miss some details about the disease[69,70]. AI-based scoring systems can help predict outcomes in patients diagnosed with advanced cirrhosis using laboratory, imaging, and clinical information, and such algorithms can identify complex nonlinear relationships among multiple variables that cannot be analyzed by the Model for End-stage Liver Disease and Child–Pugh scoring systems[71]. Numerous studies have revealed that AI models have the ability to improve predictions of mortality, risk of hospitalization, and risk of hepatic decompensation compared with traditional prognostic systems. For instance, the use of ML algorithms that operate with electronic health records provided improved results in predicting immediate mortality and clinical deterioration in patients with advanced liver disease[71]. AI models have also been studied for predicting complications such as variceal bleeding, hepatic encephalopathy, and ascites, potentially allowing faster responses and customized monitoring of patients with these conditions[72].
Using multimodal data, including clinical and imaging data, it is possible to develop models that offer improved diagnostic accuracy in the management of patients with liver diseases. However, issues concerning algorithm validation and data standardization remain.
AI IN PANCREATIC DISEASES
The utilization of AI has grown in the area of pancreatic diseases, particularly because of its potential to aid in the early detection, diagnosis, and treatment planning of pancreatic conditions such as pancreatic ductal adenocarcinoma, pancreatic cystic lesions, and pancreatitis. Pancreatic cancer generally has a poor prognosis, partly because of late presentation of the disease. Therefore, novel medical technologies driven by AI are under investigation to improve the diagnosis of pancreatic diseases and facilitate early intervention.
Pancreatic cancer has one of the highest mortality rates, owing mainly to factors such as late diagnosis and the absence of effective screening tools. For this purpose, AI technologies have been developed. They have the potential to identify subtle signs of the disease based on electronic medical records and imaging studies. These technologies can process complex statistical records, laboratory data, and medical imaging data to assess the risk of pancreatic cancer development. ML-based prediction models yielded pooled diagnostic results with an area under the curve (AUC) of around 0.77 for identifying pancreatic cancer risk in patients with newly diagnosed diabetes, according to a recent meta-analysis that included more than 2.2 million participants[73]. Furthermore, AI-based disease trajectory models have demonstrated the capability to detect pancreatic cancer years before clinical diagnosis by revealing relationships among different symptoms and comorbidities in patient records[74,75].
Medical imaging is an integral part of evaluating pancreatic diseases. AI has shown great potential for enhancing the efficiency of imaging interpretation. Automatic detection and classification of pancreatic lesions are enabled by DL methods using imaging technologies such as CT, MRI, ultrasound, and positron emission tomography. High-dimensional image features can be identified by radiomics and CNNs. A meta-analysis studied AI-based radiomics models using CT, MRI, positron emission tomography, or ultrasound for pancreatic cancer detection and reported 88% sensitivity and 93% specificity[76]. Moreover, 92% sensitivity and 92% specificity were reported for DL models in the accurate diagnosis of pancreatic cancer[77]. AI can support different methods of tumor and lesion localization and segmentation and may also be useful for differentiating between malignant and benign pancreatic lesions[78].
Cystic lesions of the pancreas, particularly intraductal papillary mucinous neoplasms, present clinical challenges because of their varying potential to become malignant. Current standards may lead to unnecessary surgeries or missed periods of malignant changes in these lesions. However, AI technologies that utilize multimodal data, such as MRI radiomics, cyst morphology, and clinical data, may allow more accurate prediction of the risk of malignancy[78].
AI solutions are being developed to predict the outcomes of pancreatic conditions. By combining clinical data, imaging characteristics, and genomic information, ML algorithms can predict survival, treatment outcomes, and the possibility of relapse in patients with pancreatic tumors. Prognostic models based on radiomics can analyze tumor heterogeneity and tumor microenvironment characteristics associated with tumor aggressiveness and aid in clinical decision-making regarding treatment options, such as surgery, chemotherapy, or targeted therapy[79].
Moreover, AI is emerging as a promising adjunct to endoscopic ultrasound, particularly for pancreatic disease. ML and DL algorithms have been used to analyze endoscopic ultrasound images to detect pancreatic masses, differentiate pancreatic ductal adenocarcinoma from chronic or autoimmune pancreatitis, and characterize cystic lesions such as intraductal papillary mucinous neoplasms. Early studies have reported high diagnostic performance, with sensitivity and specificity generally exceeding 80%[80]. More recent work suggests potential applications in automated lesion detection, segmentation, anatomical recognition, and integration of imaging with clinical or cytological data[81]. However, limited external validation, heterogeneous datasets, operator variability, and a lack of prospective multicenter studies remain important barriers.
AI studies predominantly address pancreatic cancer; nevertheless, AI can also be applied to other pancreatic diseases, such as acute and chronic pancreatitis. ML algorithms can be used to analyze clinical parameters, medical images, and laboratory tests to assess disease severity, complications, and associated mortality risk[82,83]. AI is increasingly being used for image analysis to detect structural changes in the pancreas, such as fibrosis, necrosis, or atrophy, that occur with chronic pancreatitis and are difficult to detect visually[84]. These models have the potential to aid radiologists in identifying subtle parenchymal abnormalities while improving patient management and risk assessment in patients with chronic pancreatic disease. However, there are numerous obstacles that prevent physicians from using AI routinely. A large proportion of the research is based on retrospective databases, which decreases the generalizability of this technology. Model interpretability, concerns regarding data privacy, and the need for extensive, multifaceted databases remain challenges[79]. In addition, these tools still require further prospective validation and incorporation into hospital and clinical practice. Despite these challenges, AI can use big data and modern computational technologies to identify hidden patterns of disease.
AI IN IBD
AI has been used in IBD in recent years. AI applications in IBD improve disease monitoring, treatment strategies, and prediction of clinical outcomes. Relevant areas of development include automated endoscopic disease activity scoring, prediction of patient response to therapy, and prediction of flares using clinical data and various biomarkers.
Endoscopic examination remains the gold standard for evaluating mucosal inflammation in IBD, using different scoring systems such as the Mayo Endoscopic Subscore and Ulcerative Colitis Endoscopic Index of Severity. Nevertheless, endoscopic evaluations have several drawbacks related to interobserver variability and the need for expert evaluation. AI-based computer vision approaches, based particularly on CNN and transformer-based technologies, have been implemented for automated assessment of disease activity via colonoscopy images and recordings[85,86]. Several studies have shown that developed models have outstanding diagnostic capabilities. In particular, DL models applied to endoscopic images showed over 90% accuracy in predicting Mayo endoscopic scores, surpassing experienced endoscopists in some multicenter validation datasets[87]. A study involving over 12000 endoscopic images obtained from over 1100 patients demonstrated that CNN-based models achieved 97% accuracy in detecting endoscopic remission and over 90% accuracy in disease grading[88].
AI is capable of processing complete colonoscopy videos, making it possible to classify active and inactive disease with an AUC of approximately 0.89–0.92[89]. Thus, this technology may standardize the disease scoring process, make video review faster, and facilitate timely decision-making in endoscopy. In addition, AI technology could eliminate the need for histologic remission assessment by predicting it, thereby requiring fewer biopsies in practice[88]. However, one of the biggest challenges in IBD management is the heterogeneous response of individual patients to biologics, including anti-tumor necrosis factor medications, integrin blockers, and interleukin blockers. AI models that combine clinical parameters with laboratory data, images, and genetic information are currently being developed to predict treatment efficacy[90]. ML systems that employ multicenter datasets of patients with Crohn’s disease treated with infliximab have reached an AUC of up to 0.91 for identifying long-term responders and non-responders[91]. In addition to clinical information, molecular and transcriptomic markers are also currently being investigated. Various methods based on gene expression have been successfully used to identify biomarker signatures associated with treatment responsiveness[92]. This indicates that AI is becoming more prominent in IBD therapy, as it assists physicians in choosing the best possible therapy for each patient.
IBD can fluctuate between periods of active disease and remission, which is why accurate prediction of impending disease activity is important for appropriate patient management. AI can incorporate different types of information, such as relevant clinical data, past records, and laboratory tests, to predict possible disease activity[93]. In this study, the authors were able to increase the AUC for predicting endoscopic activity from 0.68 (calprotectin only) to 0.93 when calprotectin was combined with several other clinical parameters[93]. Combining information from blood, urine, and feces produces highly accurate predictive AI models. ML classifiers outperform traditional single-marker approaches, such as C-reactive protein alone[94]. This is important because inflammatory markers are elevated earlier than clinical signs appear, allowing for earlier therapeutic intervention and better disease management.
Nonetheless, the implementation of AI in IBD management can lead to more objective endoscopic evaluations, selection of suitable therapies for patients, and prediction of flare-ups in advance. With growing datasets and improved models, AI is becoming more relevant in precision medicine in gastroenterology.
AI IN FUNCTIONAL AND MOTILITY DISORDERS
AI has also been utilized in functional gastrointestinal disorders and motility disorders such as irritable bowel syndrome, functional dyspepsia, and esophageal motility disorders. Because these disorders lack evident structural abnormalities, diagnosis and treatment can be more difficult. However, AI can be applied to the analysis of physiological tests, detection of patterns in the microbiome, and prediction of gut–brain axis phenomena.
High-resolution manometry (HRM) has become the cornerstone of assessment in esophageal motility disorders such as achalasia and esophagogastric junction outflow obstruction. The interpretation of HRM data generally requires an expert and may differ among observers. AI technologies, especially ML and DL approaches, are being developed to help automate the analysis of manometric recordings. Research has shown that ML methods can effectively classify patterns of esophageal motility based on the Chicago Classification, which is the standard for data interpretation in HRM. A multicenter study in Portugal, Spain, and the United States concluded that ML approaches would be useful for identifying esophageal motility disorders using HRM, implying that the use of advanced technologies in HRM may help standardize interpretations and reduce interobserver variability[95]. The use of CNNs to analyze HRM pressure topography data has been shown to yield an accuracy exceeding 90% in diagnosing conditions such as achalasia and distal esophageal spasm. Researchers demonstrated that AI achieved an accuracy of 97% in the HRM diagnosis of such esophageal disorders[96]. Such models have the ability to reveal intricate pressure patterns over time and automate classifications. AI can also be used to detect motility abnormalities that might otherwise not be identified. Recent research conducted in China has shown that applying AI to analyze HRM data yielded accuracies exceeding 90% in diagnosing esophageal motility disorders[97]. ML algorithms can also be applied to HRM data analysis to identify and investigate pressure patterns and correlate them with patient symptoms to better understand functional esophageal disorders.
AI IN MICROBIOTA
AI, along with ML and DL, has become one of the most powerful approaches to studies of gut microbiota. The gut microbiome provides high-dimensional datasets that are processed via 16S rRNA sequencing, metagenomics, metabolomics, and transcriptomics[98]. AI approaches are used to identify microbial signatures of diseases, predict clinical outcomes, stratify patients, and implement precision medicine[99]. The human gut is home to trillions of microorganisms whose composition varies among individuals and changes over time. Standard statistical methods may be unable to overcome challenges related to high dimensionality (thousands of microbial features), sparsity of datasets, nonlinear biological connections, and multi-omic integration. AI methods such as random forests, support vector machines, gradient boosting, neural networks, graph neural networks, and transformer architectures are more effective in solving complex problems than traditional approaches[100]. The main applications of AI in studies of gut microbiota include disease diagnosis and prediction, precision medicine, and multi-omics integration. Another study applied a random forest algorithm to establish distinct microbiota colonization as a characteristic of endoscopic recurrence in Crohn’s disease and revealed that microbiota recolonization after ileocecal resection differed between patients with and without recurrence[101].
Recent literature provides evidence regarding the use of AI in gut microbiota, including multi-omics. The gut microbiota consists of trillions of microorganisms. They contribute to metabolic regulation, immune homeostasis, nutrient processing, and protection against pathogens. Dysbiosis is defined as alterations in the microbial composition or function of the gut. IBD, obesity, diabetes mellitus, colorectal cancer, cardiovascular disease, and neurodegenerative disorders have been reported to be associated with dysbiosis. High-throughput sequencing technologies, such as 16S rRNA sequencing, shotgun metagenomics, metatranscriptomics, metabolomics, and proteomics, have enabled comprehensive characterization of microbial ecosystems[102]. However, traditional biostatistical methods are challenged by the high-dimensional, sparse, and nonlinear datasets generated by these technologies. AI and ML algorithms can reveal complex relationships within microbiome data and translate biological findings into clinically actionable information. Recent literature indicates that random forests, support vector machines, XGBoost, neural networks, and DL architectures are among the most frequently applied methods in gut microbiome studies[103,104]. Microbial features are used by supervised learning methods to predict disease states or clinical outcomes in most microbiome studies. Different algorithms have been studied in recent research. Random forest is the most frequently utilized method because of its robustness in handling sparse microbiome datasets. Figure 1 summarizes the use of AI in gastroenterology.
Figure 1 Summary of artificial intelligence use in gastroenterology.
AI: Artificial intelligence; CADe: Computer-aided detection; CADx: Computer-aided diagnosis; CNN: Convolutional neural network; ML: Machine learning; XGBoost: Extreme gradient boosting.
LIMITATIONS
Although AI has demonstrated considerable potential across gastroenterology, several limitations should be considered when interpreting the current evidence. First, many of the available studies are retrospective, single-center, or based on relatively small and highly selected patient populations. Consequently, the reported performance of AI models may not be representative of real-world clinical practice, and their generalizability to different healthcare systems, patient populations, and disease prevalences remains uncertain. Larger prospective and multicenter studies are required to validate these findings. Another important limitation is the heterogeneity and quality of the data used to develop AI algorithms. Differences in endoscopic equipment, image quality, acquisition protocols, radiological techniques, laboratory measurements, and data labeling can significantly affect model performance. In addition, many datasets are derived from expert centers, which may introduce selection bias and limit the applicability of models in settings with different levels of clinical expertise. Standardization of data collection, annotation, and reporting is therefore essential. Data privacy and security represent additional challenges because AI systems often require large amounts of sensitive clinical, imaging, genomic, and electronic health record data. Appropriate data protection, secure storage, informed consent, and compliance with relevant regulations are necessary for responsible AI development. Furthermore, differences in regulatory requirements and the lack of universally accepted standards for evaluating and approving medical AI systems may delay their widespread clinical implementation. Another limitation is that high diagnostic accuracy does not necessarily translate into improved patient outcomes. Many studies primarily evaluate technical performance measures such as sensitivity, specificity, accuracy, or AUC, while relatively few demonstrate improvements in long-term clinical outcomes, patient safety, quality of life, or healthcare costs. The potential for automation to increase unnecessary biopsies, procedures, or false-positive diagnoses should also be carefully evaluated. These limitations highlight the need for prospective multicenter validation, standardized datasets and evaluation criteria, transparent and explainable algorithms, robust data protection, and evidence demonstrating meaningful clinical benefits. Addressing these challenges will be essential for the safe, effective, and equitable integration of AI into routine gastroenterological practice.
CONCLUSION
AI is rapidly becoming an important tool in gastroenterology, with promising applications in endoscopy, gastrointestinal oncology, liver and pancreatic diseases, IBD, motility disorders, and microbiome research. AI can improve lesion detection and characterization, support early cancer diagnosis, assess disease severity and prognosis, and assist in personalized treatment decisions. Despite these advances, several limitations remain, including limited external validation, retrospective study designs, data quality and privacy concerns, algorithmic bias, lack of interpretability, and uncertainty regarding integration into routine clinical practice. Further large, prospective, multicenter studies are needed to confirm the clinical and cost-effectiveness of AI systems. Overall, AI should be viewed as a tool that complements rather than replaces clinicians. With appropriate validation, standardization, and ethical implementation, AI has the potential to improve diagnostic accuracy, reduce variability, enable earlier disease detection, and enhance personalized care in gastroenterology.
