Applications of Artificial Intelligence in Gastroenterology and Hepatology

Abstract

Gastroenterology (GI) and hepatology are in the early stages of incorporation of artificial intelligence (AI) into clinical practice. The two major areas of AI deep learning technology which can be considered applicable to GI and hepatology are image recognition analysis and clinical data analysis. Additional areas of AI such as generative AI also may have roles in clinical practice. Continued development, validation, and real-world modeling of AI systems will be needed prior to wider integration. Based on the trajectory and rapid developments within AI, it is likely that in the coming years new areas of AI applications in GI and hepatology will be proposed and current AI applications will be enhanced and become standard of care.

1. Introduction

The concept of artificial intelligence (AI) has evolved significantly from its inception in the 1950s to today and new advances in the field have the potential to assist or disrupt multiple industries including health care [1]. The fields of gastroenterology (GI) and hepatology will be vastly transformed in the coming years by incorporating AI and are currently at early stages of integration.

AI broadly is defined as any program that has the capacity to learn and problem solve in a way similar to a human being. One particular subset is machine learning (ML). ML algorithms are statistical tools that allow a system to imitate human intelligence and behavior in data analysis by generating mathematical algorithms to address a new problem without explicit programmed instructions on how to do so [2]. These algorithms are potentially transformative for health care as they can identify patterns, predict outcomes, and assist in diagnosing and treating various conditions. The type of patient data fed into these systems can range from medical histories to diagnostic images and laboratory results. For clinicians, this means the potential for more accurate and personalized patient care, as well as improved efficiency in managing large data sets [3]. One prominent subset of ML is deep learning, a technique inspired by the human brain’s architecture of biological neural networks. Deep learning algorithms, specifically convolutional neural networks (CNNs), are built to process visual information, making them particularly pertinent in analyzing medical imaging [4]. CNNs specifically are modeled off of the way neurons process visual data and how the human visual cortex processes multiple inputs to extract image patterns [5]. Physicians can think of CNNs as virtual detectives that autonomously learn to recognize patterns and features within medical images, both radiographic and endoscopic. This capability allows CNNs to identify abnormalities that may be challenging for the human eye to detect [6].

The application of deep learning and CNNs in medicine aligns with a growing emphasis in the field of precision medicine, which underscores the importance of providing accurate, proactive, and personalized diagnoses and treatments for patients by incorporating all available tools. The integration of ML, deep learning, and CNNs in GI hold the promise of advancing diagnostic capabilities, optimizing treatment approaches, and improving overall care in a field where early and accurate detection is crucial for sustained health [7]. Deep learning networks additionally allow for processing large volumes of patient data to assist in disease risk factor prediction and treatment outcomes [8].

Overall, the two major areas of AI deep learning technology which can be considered applicable to GI and hepatology are image recognition analysis and clinical data analysis (see Table 1). Additional areas of AI such as generative AI also may have roles in clinical practice.

Anatomical region	Pathologies
Esophagus	Barrett’s esophagus
	Esophageal adenocarcinoma
	Esophageal squamous cell carcinoma
	Eosinophilic esophagitis
	Esophageal varices
Stomach	Gastric cancer
	Gastrointestinal stromal tumor
	Ulcers
	Helicobacter pylori
	Arteriovenous malformations
Small intestine	Ulcers
	Celiac disease
	Small intestinal malignancy
	Arteriovenous malformations
Colon	Adenoma detection
	Colorectal cancer
	Inflammatory bowel disease
Pancreas	Pancreatic cysts/cancers
	Pancreatitis
Gallbladder/biliary tree	Choledocholithiasis
	Biliary malignancy
Liver	Liver transplant
All regions	GI bleeding

Table 1.

List of pathologies with currently studied artificial intelligence applications.

2. Methodology

A literature search was performed on several databases, primarily PubMed, for original and review articles pertaining to the interface of GI/hepatology and AI. Initial search yielded 2524 results. To narrow findings, more recent results from within the past 5 years were given priority, and additional search terms were used for specific topics. For instance, addition of “endoscopy,” “colonoscopy,” or “endoscopic ultrasound” yielded 956, 368, and 90 results, respectively. Additional studies of interest were identified by adding search terms related to pathologies, AI modalities (i.e. computer aided detection or diagnosis, generative AI), and review of primary references cited in review articles. Ultimately, 112 articles were utilized in compiling this chapter.

3. Image recognition analysis

3.1 Endoscopy and colonoscopy

Image recognition AI softwares have the capacity to compensate for human error and improve the efficiency and quality of upper endoscopies and colonoscopies. Image recognition can fall into a category of computer aided detection (CADe) or computer aided diagnosis (CADx). CADe can help with segmentation (identification and localization of an abnormality) and CADx can help distinguish between diagnoses. The role of CADe and CADx varies depending on the pathology being assessed endoscopically [9].

In the esophagus, identifying premalignant conditions and subtle changes of malignancy can be assisted by CADe and CADx. Identifying possible areas of Barrett’s esophagus is of utmost importance for endoscopists as early detection is associated with decreased mortality from esophageal adenocarcinoma. The miss rate of early Barrett’s neoplasia is also reported to be up to 40% [10]. CADe of Barrett’s mucosa has been assessed my multiple systems starting in 2016 [11]. Initial algorithms were fed data from volumetric laser endomicroscopy images, which provided scans of esophageal wall layers up to 3 mm deep. The detection is based on texture patterns and mucosal colors from this data set [12, 13]. A subsequent study suggested that use of a trained model was associated with higher sensitivity and specificity for identifying Barrett’s mucosa than by non-expert endoscopist identification alone [14]. Additionally, models have started to be developed to help with CADx between Barrett’s mucosa and early esophageal adenocarcinoma [15, 16]. Studies have shown that various AI models outperform experienced endoscopists in CADx between different degrees of tumor depth [10]. One pilot study created a model based on still images to distinguish between T1a and T1b esophageal adenocarcinoma. The model was able to do so but was not superior to expert identification [17]. Multiple studies, primarily from Asian countries due to the higher prevalence there, have investigated AI for the detection of esophageal squamous cell cancer (SCC) as well [18, 19]. Models also show promise in distinguishing invasion depth of esophageal SCC [20].

Eosinophilic esophagitis (EoE) is another area in which CADe has been used to identify and monitor disease progress. CNNs have been shown to have higher accuracy, sensitivity, and specificity than non-expert endoscopists on test sets of images [21]. Additional models have been developed that incorporate the EoE Endoscopic Reference Score (EREFS) and have algorithms to identify scores for all individual EREFS components [22]. Models are currently in early stages of training, and all available studies at this time are based on still image data sets rather than real world analysis [21, 22, 23].

A meta-analysis from 2021 showed that AI systems have a high accuracy, of up to 90%, in detection of all upper GI neoplasias including gastric cancers [10]. Some studies have also focused on predicting the depth of tumor invasion in early gastric cancer, revealing that AI predicted invasion depth with high sensitivity and specificity, except in cases of undifferentiated histology where the accuracy of the AI model was significantly lower [24, 25]. A CNN model has been shown to have similar accuracy to magnification endoscopy with narrow band imaging, suggesting a role for CNNs in CADx between gastritis and early gastric cancer [26]. CADx of distinguishing early gastric cancer from gastritis CADe models have been shown to have similar accuracy to expert endoscopists and superior accuracy to non-expert endoscopists for upper GI cancers [27]. In addition to the diagnostic utility of the technology, this also suggests a role in endoscopist training for AI systems.

Endoscopic images of Helicobacter pylori (H. pylori) infected gastric mucosa have also been used to feed into training models. The gastric mucosa in H. pylori infected patients exhibits varying degrees of inflammation, intestinal metaplasia, and atrophy. Studies have shown that AI could outperform an endoscopist in detecting these changes. A meta-analysis of all studies employing AI models found that using CNN models for CADe was superior to a traditional ML support vector model in detecting H. pylori infection [28]. Another meta-analysis noted AI algorithms had a pooled accuracy of up to 82% for H. pylori infection on visual detection alone [29]. These may indicate when these models are more validated and mature, it could be possible to identify H. pylori infection endoscopically alone, without the need for biopsy.

One of the most well studied areas in AI endoscopy is the use of CADe in adenoma detection during colonoscopies. Adenoma detection rate (ADR) is one of the most validated indicators of colonoscopy quality, and increased ADR is associated with decreased interval colorectal cancers [30]. A meta-analysis of 43 publications that included 15,000 colonoscopies noted an adenoma miss rate (AMR) of 26%, a miss rate of 9% for advanced adenomas, and 27% for serrated polyps [31]. A meta-analysis of five studies compared ADR in traditional colonoscopy to ADR in colonoscopies using CADe and noted a higher pooled ADR in the CADe group of 36.6% compared to 25.2% in the control. This finding was statistically significant and suggests a strong benefit for incorporating CADe into clinical practice [32].

Based on the strong results of a plethora of RCTs suggesting benefit in colonoscopy, commercial systems have recently become approved by the Food and Drug administration (FDA) [30, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44]. The first FDA approved CADe system is GI Genius (Medtronic), approved in 2021 [45]. This system utilizes visual identification markers of green boxes to highlight polyps in real time during endoscopic evaluation (Figure 1). Another system was approved in 2022 called EndoScreener (Wision AI) [33]. Other commercially available systems include EndoBrain (Cybernet Systems Co.), Discovery (Pentax Medical), REiLI CAD EYE (Fujifilm), and ai4gi (Imagia). As CNNs and AI technology in general continue to develop, it is likely that new commercial systems will come to market and apply for regulatory approval [46]. However, as real-world data has begun to be published on the use of commercially available CADe systems in colonoscopy, results have not consistently indicated statistically significant differences between AI assisted and traditional colonoscopy ADRs. A multicenter randomized clinical trial by Wei et al. assessed the EndoVigilant system (Endovigilant), an investigational real time platform for adenoma detection, and noted no difference in detection of adenomas with CADe [47]. Two additional studies looking at the real-world performance of the GI Genius system noted no difference in ADR between CADe and traditional colonoscopy as well [48, 49]. It is possible that as these technologies mature and are more widely deployed, and as data sets used to train the models increase in size, future real world application studies may correspond with results from the pre-approval RCTs.

Figure 1.

Comparison of regular colonoscopy (left) with AI aided polyp detection (right) via the GI Genius system (images reproduced with permission from Medtronic).

CADx in colonoscopy incorporates real time analysis to determine the type of polyp identified via optical diagnosis. The goal of training and deploying CADx systems in colonoscopy is to determine which polyps need to be removed (i.e. adenomatous/neoplastic polyps) and which can safely be kept in situ or resected and then discarded without histologic analysis (i.e. hyperplastic polyps). One large scale study of over 1000 patients in the UK, Norway, and Japan indicated no clinically significant difference between CADx and traditional optical diagnosis in polyp type identification [50]. Another study of 320 patients suggested higher accuracy in endoscopists compared to currently available CADx systems [51]. A study of the GI Genius CADx module noted a high negative predictive value for adenomatous histology specifically for polyps less than 5 mm. These results would support a resect and discard or leave-in-situ strategy, but further studies and validation must be performed as the technologies mature given the other discordant results [52]. CADx may also have a role in determining staging of neoplastic lesions in the colon, similar to its use in esophageal and gastric malignancies [53].

In addition to CADe and CADx, colonoscopy may incorporate AI for monitoring withdrawal time, another important indicator of colonoscopy quality. An average withdrawal time after reaching the cecum of at least 6 minutes is associated with higher ADRs [54]. One recently developed CNN utilized a data set of 100 colonoscopy videos to accurately predict and report withdrawal times from the cecum [55]. A study of an AI-aided speedometer which utilized an optical flow estimation to predict endoscope speed showed no difference in withdrawal time or ADR between speedometer-assisted colonoscopies and traditional colonoscopies [56]. However, a preliminary study of ENDOANGEL, a multiple use CNN model that includes withdrawal speed monitoring, indicated a 50% higher ADR in the speedometer-assisted colonoscopy compared to traditional [57]. This will likely be an area of future study and commercial development. Size estimation of polyps can also be assisted by AI. One study used both a computer vision technique previously applied to topographic studies to estimate polyp size in an animal model, and this system had a statistically and numerically higher accuracy in polyp size estimation than endoscopist estimation alone (85.2–59.5%). This same group developed a CNN model using videos of human colonoscopies and noted a high accuracy of 80% [58]. Another group used a CNN trained on ophthalmologic and colonoscopic data sets and showed increased accuracy of AI measurement compared to non-expert endoscopists’ polyp size estimations [59].

Another area of potential benefit in endoscopic management of GI pathologies is in assessment of GI bleeds (GIB). ML models may be used to predict rebleeding risk, success of potential interventions, and mortality in GIB with a greater accuracy than existing clinical risk stratification tools [60]. Specifically for esophageal varices (EV) and gastric varices (GV), the multiuse ENDOANGEL CNN model (previously studied in withdrawal speeds for colonoscopies, as above) has been shown to be accurate. ENDOANGEL was also able to identify high risk stigmata such as the red wale sign [61]. An additional study of this system showed it outperformed expert endoscopists in identifying high risk rebleeding signs in EV and GV, which suggests that AI models may be useful adjuncts especially for this group of patients in determining the correct intervention [62]. Specifically, most GV locations currently do not have an endoscopically feasible intervention available and must be treated via an interventional radiology approach. If AI can assist in risk stratifying these patients with a high degree of accuracy, this will likely lead to improved morbidity and mortality in this group.

3.2 Capsule endoscopy

Wireless video capsule endoscopy (VCE) allows for less invasive intraluminal image capture than traditional endoscopy. However, image analysis in VCE is time consuming and operator dependent, which can result in missed lesions or pathologies. CNNs have been studied for VCE image analysis in order to improve the analysis of these results [11, 63].

CADe can be used to assess for the presence of small intestinal bleeding, which is the most common indication for capsule studies. Multiple models have been proposed in the literature, though currently none are approved in the US. A model by Aoki et al. was able to identify mucosal breaks, angioectasias, protruding lesions, and blood with high accuracy [64]. Subsequently, this system was compared in a real world setting to another CNN system and was noted to have similar abnormality detection rates for high-risk bleeding lesions [65]. Reading times for capsules likewise improved with the use of CNNs in this study. Additionally, models for colon capsule endoscopy to detect blood have been developed with high sensitivity and specificity [66]. VCE CADe also has a role in detecting gastric and duodenal ulcers, with larger ulcers being more easily detected [64]. Deep learning models are being developed for assessment of abnormal mucosa in celiac disease as well [67, 68]. These models may be used for diagnosis and classification of celiac phenotype [69].

The abundance of CNN models for VCE analysis suggests that AI will be very useful in the future for analysis. The fragmentation of the VCE market amongst multiple companies likely means that any CNN model will also become proprietary, so it is unlikely that only one model will emerge as the dominant capsule analysis program.

3.3 Endoscopic ultrasound and endoscopic retrograde cholangiopancreatography

The role of AI in advanced endoscopy, specifically endoscopic ultrasound (EUS) and endoscopic retrograde cholangiopancreatography (ERCP) is linked to a need to diagnose malignant and premalignant pancreaticobiliary lesions. Pancreatic cancer has been difficult to diagnose at an early stage, and for this reason, mortality has increased by 53% over the past 25 years [70]. Sometimes pancreatic cancer arises from a pancreatic cyst. Pancreatic cysts are often initially identified by CT or MRI. Once a high-risk pancreatic cyst is identified, usually defined by above 2 cm, growth of 5 mm in 1 year, dilated pancreatic duct, mural nodules, etc., the patient is usually referred to gastroenterology for an endoscopic ultrasound for further characterization and potential sampling through fine needle aspiration or biopsy [71]. Detection of high-risk pancreatic cysts is critical in preventing the progression to pancreatic cancer.

AI assisted EUS in premalignant pancreatic cysts has been evaluated as a potential answer to the limitations of CT, MRI, and traditional EUS in differentiating between benign lesions, high risk intraductal papillary mucinous neoplasms (IPMNs) and malignancy. A small retrospective study showed that image processing with AI increased sensitivity to detect malignant IPMNs from 56% with traditional EUS to 95% [72]. Another retrospective study in China saw that AI image interpretations of EUS images from patients with and without pancreatic cancer had a sensitivity, specificity, positive predictive value and negative predictive value all above 93% [73]. Based on these studies AI can be a helpful tool in the diagnosis of both high-risk pancreatic cysts and pancreatic cancer, which could potentially improve outcomes.

Limitations in AI in EUS are largely due to operator dependency and lack of data with high power. EUS is considered an advanced modality and so endoscopists with more experience are more likely to detect lesions at a higher rate. Most studies looking at AI do not contain more than 1000 patients likely because of the rarity of pancreatic cancer. There are also multiple confounding factors that could depreciate the utility of AI, including age. A study was conducted in Turkey used a computer aided diagnosis system with image processing to retrospective analysis of EUS images taken from patients who were subsequently diagnosed with pancreatic cancer or benign lesions based on biopsies and stratified into age groups (<40, 40–60 and >60). The study found that the sensitivity was higher in the older age group (87 vs. 93%) [74]. While there are limitations to AI in EUS, it can be deduced that with time these weaknesses can be overcome with the natural progression of technology over time.

Other important premalignant lesions that could be identified with more accuracy include gastrointestinal stromal tumors (GISTs). These are tumors arising from the muscularis propria that have malignant potential. For this reason, all GISTs >2 cm and most GISTs <2 cm should be resected in good surgical candidates [75]. A meta-analysis including 7 studies which utilized AI assisted image recognition of EUS images in 2431 patients found that the program had a sensitivity of 92% and a positive likelihood ratio of 4.55 (95% CI 2.64–7.84) [76]. These studies suggest that AI assistance could help in the early diagnosis of GISTs, which could lead to earlier resections and better outcomes.

Finally, another core procedure in the advanced endoscopy repertoire is ERCP. This procedure is useful in relieving biliary obstruction with or without infection in a variety of diagnoses including choledocholithiasis and malignancies. One way for AI to assist endoscopists performing ERCPs is by helping to identify the ampulla in patients with difficult anatomy. A study utilizing a convolutional neural network that works during ERCP procedures to identify the ampulla was found to be 76% precise [77]. This may lead to easier cannulation in patients with anatomical variations of the ampulla, including diverticula. Another way AI can be helpful with ERCP is by recognizing patients who are at increased risk of developing post-ERCP pancreatitis. One study developed a ML algorithm that used 10 variables in 544 patients and found that the ML algorithm was more accurate compared to traditional logistic regression in predicting which patient will develop post-ERCP pancreatitis. This can help proceduralists have risk/benefit discussions with their patients and optimize them for their procedures [78].

3.4 Pathology

Image recognition models may be useful in gastrointestinal pathology for multiple disease as well. For example, in EoE, models are starting to be created to use CADe segmentation to identify eosinophils on biopsy samples [79]. This would be clinically relevant for definitive diagnosis of the disease, and also for monitoring treatment response. CNNs may also assist with pathologic identification of gastric and colonic malignancies, however most studies thus far are observational or retrospective [80]. The nascent field of computational pathology utilizes AI based tools and ML to enhance accuracy and speed of diagnosis for GI pathology, but likewise large scale studies are still needed to prove a benefit over traditional pathologic methods [81]. A specific area in which AI may be particularly useful in GI pathology is automating time consuming processes. One example in the literature is identifying coccoid forms of H. pylori, which are associated with refractory infection [82].

4. Clinical data analysis

4.1 Risk factor prediction

Over the last 5 years, a series of studies have explored the potential of AI in predicting disease development, outcomes, and treatment responses [83]. The specific pathologies most studied for risk factor identification include inflammatory bowel disease (IBD), gastric cancer, and GIB. IBD in particular has benefited from large volume data set analysis for identification of genetic risk factors. Isakov et al. assessed 16,390 genes in patients with IBD, identified pathways associated with disease expression, and were able to use deep learning to identify 347 target genes for IBD. Of these, 67 were novel [84]. Romagnoni et al. employed artificial neural networks on gene expression profiles from nearly 30,000 Crohn’s disease patients. They created a model from this that was able to utilize epigenetic data of gene expression to diagnose Crohn’s disease with 80% accuracy [85].

Identifying risk factors for gastric cancer recurrence and staging is another area that deep learning may be useful in. One AI model was noted to have a higher accuracy than the current tumor, nodes, metastasis staging system (TNM). This model utilized data inputs of biomarkers, histology, and computed tomography (CT) images to predict the likelihood of metastasis to the liver or lymph nodes, with a sensitivity of 66.7% and specificity of 97.1% [86, 87].

Risk stratification for GI bleeding is also an area of interest in AI-assisted data analysis. Ayaru et al. developed a predictive model for non-variceal upper GIB that was able to accurately predict rebleeding risk in 88% of cases [88]. Likewise, Das et al. contributed to this advancement by creating a predictive ML model specifically tailored for lower GI bleeding [89]. Notably, their model exhibited superior accuracies, sensitivities, and specificities when compared to the current standard BLEED score. This suggests a notable improvement in the ability to predict outcomes related to lower GI bleeding, underlining the potential for more effective clinical decision support in this critical aspect of gastroenterological care.

4.2 Treatment outcome prediction

Several studies have utilized ML models to attempt to predict IBD flares by analyzing surrogate IBD-related outcomes such as hospitalizations, steroid use, and initiation of biologics. Waljee et al. looked at insurance claims data of over 20,000 patients to obtain information on steroid use or hospitalizations, and used this data to create a predictive algorithm for IBD flares which was superior to fecal calprotectin [90, 91]. These studies emphasize the potential cost advantage of the use of ML in health care and the potential of utilizing existing EMR data sets to improve outcomes rather than individual biomarkers.

Another area of IBD outcome assessment using AI is prediction of response to biologics prior to initiation, and subsequently predicting continued response. One ML study used data on vedolizumab in Crohn’s disease patients to develop a system for prediction of steroid free remission on this medication [92]. Data fed into the model included demographics, medication dosing intervals, concurrent medication use prior to biologic initiation (e.g. immunomodulators or steroids), serologic markers, and prior biologic exposure. At week 6, this model had a sensitivity of 76% and specificity of 71% for predicting steroid free remission. Another similar study using ustekinumab instead of vedlizumab was able to predict disease remission with a high degree of accuracy by week 8 of therapy [93]. Additional studies have used similar approaches for assessing risk of adverse events on therapy [94].

Within hepatology, large data sets and ML may be utilized in predicting outcomes of liver transplantation [95, 96]. Currently, statistical models such as the model for end stage liver disease (MELD) are employed to assist with organ allocation. However, MELD has limited utility in predicting post-transplant survival. One retrospective study by Yu et al. compared traditional statistical models (cox regression, MELD, donor MELD, balance of risk score) against ML methods including artificial neural networks and random forest [97]. The random forest model and cox regression models showed the highest area under the receiver operating characteristic curve (0.81 at 12 months for random forest and 0.77 for cox regression). These results suggest that ML methods are comparable to traditional methods for post-transplant survivial. For hepatocellular carcinoma, one model of an artificial neural network called Metroticket had greater accuracy in predicting 5-year post transplant survival than the Milan criteria or alternate HCC scoring systems such as the AFP score [98]. Another study by Nam et al. yielded a statistically significant improvement in determining post-transplant HCC recurrence by using a deep neural network called MoRAL-AI. This was found to be superior to the Milan criteria and other comparable outcome systems (MoRAL, UCSF, Kyoto criteria, and up-to-seven) [99].

Despite the promise of these technologies, it is important to acknowledge the limits of ML models. They can establish correlations between variables but not causation, and their efficacy is limited by the availability and quality of the data sets. In the clinical setting, predictive outcomes play a crucial role. They enable the identification of patients at risk for negative outcomes, facilitating early detection and intervention. Predictive models also contribute to personalized patient recommendations, guiding clinicians in tailoring interventions and care strategies to reduce the probability of poor outcomes.

Looking ahead, the future of predictive modeling in gastroenterology involves a shift from traditional statistical analyses to ML, emphasizing accuracy in outcome prediction over significance or interpretability. Transferability of models across different institutions and the contextual interpretation of outcomes are crucial considerations. Understanding the context of a model’s output is essential, as predictions may vary based on interventions and priorities in different clinical settings. Prospective models that address reliability, outcomes, and cost considerations are integral for the continued advancement and integration of AI in gastroenterological practice.

5. Additional AI areas of interest

5.1 Generative AI

Generative AI refers to an algorithm capable of creating a new form of content by learning from a large body of data in the form of text, audio, images, or video. In clinical practice, generative AI models such as ChatGPT, a conversational large language model (LLM) have shown promise in areas such as development of differential diagnoses and valid treatment plans [100]. However, these same studies also identify multiple flaws with language based generative AI. One study challenging ChatGPT to answer questions of management of cirrhosis showed that it was proficient at describing the algorithm for initial workup but provided incorrect answers or outdated information for treatment [101]. At this time in its infancy, the accuracy of generative AI is limited greatly by the data available to the database which includes unsubstantiated medical fallacies, conflicting medical data, and outdated recommendations. Furthermore, current LLMs tend to produce “hallucinations,” defined as fake data generated to fill the gaps in a model’s knowledge base to falsely substantiate its claims [102]. Another limitation to generative AI in medical decision making is its inherent bias. It has been well established that LLMs in general have inherent racial, gender, and religious biases based on their datasets [102]. These biases are also compounded by artificial intelligence being presented with the task of making decisions based on data reflecting disparities created by human bias. Although generative AI diagnosis and treatment is promising, at its current capabilities LLMs are limited by their databases and unable to provide reliable medical decisions.

Generative AI also shows potential as a powerful tool for patient education. Currently, several studies in various fields have shown that Chatbots produce accurate, reproducible answers to patients’ questions and in some cases, are non-inferior to physician answers [103, 104, 105, 106]. Despite its success, the AI was not able to understand emotion and express empathy which is a limitation AI will likely not be able to overcome despite advances [104, 106]. One study assessing a ChatGPT chatbot’s ability to answer gastrointestinal health related questions found that it is a cost effective and readily available tool although, like discussed earlier, tends to be inaccurate and not be comprehensive [107]. Similarly, another study using ChatGPT to generate patient specific education material about transjugular intrahepatic portocaval shunt (TIPS) showed it to be accurate but again, failed to be comprehensive not including hepatic encephalopathy as an adverse effect. Ideally, we expect generative AI in the future to be able to provide patient centered education taking into consideration their comorbidities, level of education, and language barriers.

One study calculated that physicians spend about 34–55% of their days reviewing charts and documenting which is time limiting their ability to see more patients and provide quality care [108]. Generative AI offers a solution to this problem by allowing for a more efficient method of documentation to hopefully reduce physician burnout and increase the amount of time physicians can allocate to providing a higher quality of care. Currently, LLMs excel at generating well written notes such as discharge summaries with brief description input by the physician [109]. Although efficient and effective, some providers found that LLMs are unable to capture the nuances and details in procedure notes that make each case unique [110]. While AI generated documentation may not completely replace human documentation, it will likely increase efficiency by generating much of the content of the reports. In fact, one study designed an AI algorithm to generate a note capable of calculating withdrawal times and automated photo documentation capable of identifying landmarks and interventions, reducing the burden of routine documentation [55]. Not only can AI assist in clinical documentation, but it can also be applied to chart reviewing and data extraction. One study used a LLM called “Versa Chat” to analyze CT and MRI reports in patients with hepatocellular carcinoma, to assess characteristics such as LI_RADS score, number of lesions, macrovascular invasion, extrahepatic metastasis, and maximum tumor diameter [111]. The study found AI to be, in general, more accurate than manual chart review at performing simple tasks such as identifying the presence of extrahepatic metastasis and determining LI_RADS scores, it was less accurate for more complicated tasks such as determining maximum tumor diameter. Based on current LLM’s inherent ability at analyzing large data input and producing a coherent output, we will likely see generative AI becoming proficient and widely used for documentation in medicine sooner than later.

5.2 Phonoenterography

An additional novel area of CNN use in GI is in analyzing recordings of bowel sounds (phonoenterography). Advances in recording device technology since the technique was first coined in 1967 have allowed for the advent of digital phonoenterography. One of the challenges of the technique is distinguishing bowel sounds from other acoustic noises recorded by the receivers. A potential application of AI is in using deep learning to process the audio information to identify trends that could distinguish bowel sounds and therefore create a processing technique to filter all other noises out. This could have applications in multiple GI diseases, including ileus, obstruction, Hypertrophic pyloric stenosis, and IBS. This field is in its nascency but standardization of bowel auscultation through AI technologies has the potential to increase the clinical utility of the technique and decrease health care costs [112].

6. Future directions

GI and hepatology are in the early stages of incorporation of these novel technologies into clinical practice. Continued development, validation, and real-world modeling of AI systems will be needed prior to wider integration. Based on the trajectory and rapid developments within AI, it is likely that in the coming years new areas of AI applications in GI and hepatology will be proposed and current AI applications will be enhanced and become standard of care.