Artificial intelligence in the management of inflammatory bowel disease: What’s next?

Abstract

Inflammatory bowel disease (IBD) is a chronic, relapsing-remitting autoimmune disorder of the gastrointestinal tract. The management of IBD is complex and requires accurate assessment of disease extent and severity which guide therapeutic decisions. Endoscopic evaluation with biopsy remains the standard for diagnosing and assessing disease activity. Additionally, other modalities such as computed tomography enterography are used for suspected small bowel involvement. However, these processes are costly, time consuming, and often rely on subjective interpretation which is influenced by clinician experience. Artificial intelligence (AI) has been used to standardize and improve efficiency in many facets of healthcare. Similarly, in the past decade, there has been growing interest in the applications of AI in the management of IBD. The applications of AI in IBD to date include automated endoscopic and histologic assessment, analysis of non-invasive imaging, discovery of novel biomarkers for the development of disease prediction models and the use of chatbots. In this article, we will discuss recent advancements in the use of AI in IBD as well as some of the practical and ethical concerns with large scale implementation of AI into clinical practice.

Key Words: Inflammatory bowel disease; Ulcerative colitis; Crohn’s disease; Artificial intelligence; Machine learning; Deep learning

Core Tip: Inflammatory bowel disease (IBD) is a chronic relapsing-remitting autoimmune disorder of the gastrointestinal tract. The management of IBD is complex and relies on subjective interpretation of invasive and non-invasive assessments to guide therapeutic decisions. Artificial intelligence (AI) is emerging as a potential tool to standardize the assessment of IBD and personalize therapeutic selection. This article highlights recent advances in AI in IBD with a focus on endoscopy, non-invasive imaging, histology, AI chatbots, therapeutic prediction models and current barriers to widespread implementation of AI in clinical practice.

INTRODUCTION

Inflammatory bowel disease (IBD) is an autoimmune disorder of the gastrointestinal tract that affects over 5 million people worldwide with increasing incidence and prevalence in recent decades[1-3]. IBD consists of two distinct diseases, ulcerative colitis (UC) and Crohn’s disease (CD). The pathogenesis of IBD remains unknown but is likely multifactorial involving genetics and environmental factors[2,4,5].

The management of IBD is complex and requires frequent invasive and non-invasive monitoring. Current Selecting Therapeutic Targets in IBD-II guidelines suggest a treat-to-target approach with endoscopic healing achieves the best short, intermediate, and long-term outcomes[6-8]. However, endoscopic and histologic assessment of IBD relies on semi-quantitative and observer-dependent scoring systems[9-12]. Analysis of non-invasive imaging modalities such as computed tomography enterography (CTE), magnetic resonance enterography (MRE), and intestinal ultrasound (IU) are also subjective and dependent upon user experience[13-15]. Additionally, while therapeutic options for IBD have increased over the past decade, treatment selection remains largely based upon clinician experience.

Artificial intelligence (AI) has shown promise in overcoming many of the obstacles in the management of IBD. Traditional machine learning (ML) methods such as support vector machines (SVM), random forest (RF), and least absolute shrinkage and selection operator, have demonstrated the capability to improve analysis of CTE images in patients with CD while extreme gradient boosting has been used to develop prediction models to help guide clinician decisions on biological therapies in UC[16-18]. Deep learning (DL) models such as convolutional neural networks (CNN) have shown promise in automating the analysis of imaging and histopathology, improving accuracy and efficiency while also reducing bias[19-21]. Additionally, large language models (LLMs), such as ChatGPT, can respond to patients’ queries about their IBD and may one day see widespread use in clinical practice[22,23]. However, despite these advancements, significant barriers remain which prevent widespread implementation of AI into clinical practice[24]. In this article, we will highlight recent developments of AI in IBD in the areas of endoscopy, non-invasive imaging, histology, LLMs and chatbots, and therapeutic prediction models while also examining practical challenges, ethical concerns, and emerging trends with the use of AI in IBD (Figure 1).

Figure 1 Limitations of traditional management of inflammatory bowel disease and potential improvements with the use of artificial intelligence. Figure created with BioRender.com (https://BioRender.com/29up61d). The traditional management of inflammatory bowel disease relies on endoscopy, non-invasive imaging, and histology, interpreted by clinicians with variable experience in inflammatory bowel disease, who then make treatment decisions based upon these results and clinical symptoms. This often leads to a one size fits all approach. In an artificial intelligence-assisted model, clinicians are assisted by artificial intelligence, which may improve analysis in endoscopy, non-invasive imaging, and histology. With these results, clinicians may then use prediction models to determine the best therapeutic regiment tailored to the patient’s unique disease phenotype. AI: Artificial intelligence; CNN: Convolutional neural networks; ML: Machine learning.

OVERVIEW OF AI

AI encompasses any application or technology that allows machines to perform tasks that mimic human intelligence. A subset of AI, ML, involves methods which allow machines to learn from data without being programed with specific rules. Traditional ML methods such as SVM and RF have been widely utilized in IBD. For example, SVM utilizes kernel functions to transform data into higher dimensional feature spaces allowing for analysis of complex inputs such as endoscopic imaging[25]. RF, which predicts outcomes by averaging the decisions of its underlying trees, can use categorical and continuous variables as inputs making them useful in the development of therapeutic prediction models[26].

Though useful in many applications, one of the drawbacks of traditional ML methods is that they rely on user defined features as inputs. However, imperceptible patterns or features may exist which may improve model performance. This automated hierarchal feature extraction is one of the major strengths of DL models[27]. DL, a subset of ML, relies on artificial neural networks (ANN). An ANN is comprised of an input layer, 1 or more hidden layers, and an output layer. DL models are ANNs with multiple hidden layers which allow for hierarchical feature extraction. One of the most common types of DL models are CNNs, which specialize in image-based tasks, where they often outperform traditional ML models due to this feature extraction. However, many other DL models exist including recurrent neural networks, which specialize in time-series and sequential data processing. Finally, natural language processing (NLP), a subfield of AI, are used for understanding text and human language. Within NLP are LLMs such as ChatGPT[27].

Despite the improvements of many DL models compared to traditional ML models, these improvements come at a cost. For example, CNNs require large datasets for training and high computational demands, some of which can be reduced by transfer learning and data augmentation using methods such as generative adversarial networks (GANs)[27,28]. Nonetheless, these models have shown promise in improving the efficiency and accuracy of image analysis in IBD.

AI IN ENDOSCOPY

Current endoscopic scoring systems for IBD are limited by subjectivity and interobserver variability[10,11,29,30]. AI has shown promise in enhancing endoscopic assessment through reducing subjectivity and enabling more detailed analysis of the mucosal surface.

White light endoscopy

Many studies using AI in endoscopy apply to white light endoscopy (WLE). Early retrospective studies showed that DL models, such as CNN-based computer-aided diagnosis (CAD) models, could accurately predict Mayo Endoscopic Subscore (MES) and discriminate between MES 0-1 and MES 2-3 using still images extracted from WLE with an area under the receiver operating characteristic curve (AUC) of about 0.97[31,32]. One of the CNN-based CADs also correlated well with histologic inflammation grade in patients who were not using topical therapies[32]. Though instrumental in showing the feasibility of DL in UC, both studies highlighted implementation challenges, such as the need for high-quality images, the lack of a ground truth in expert image annotation, and its effects on model performance when delineating between neighboring MES values (Table 1).

Table 1 Performance of deep learning models in endoscopic assessment of mucosal healing.

Ref.	Input	Model	Key performance metrics
Stidham et al[31]	Images	CNN	MES 0-1 vs MES 2-3: AUC = 0.97
Ozawa et al[32]	Images	CNN	MES 0 vs MES 1-3: AUC = 0.86; MES 0-1 vs MES 2-3: AUC = 0.98
Gottlieb et al[35]	Videos	RNN	EH by eMS: Accuracy = 95.52%; EH by UCEIS: Accuracy = 97.04%
Takenaka et al[34]	Images	DNUC	UCEIS ≤ 1: Accuracy = 90.1%
Higuchi et al[68]	Images1	CNN	MES 0-3: Accuracy = 91.3%-99.4%
Yao et al[33]	Videos	CNN	MES 0-1 vs MES 2-3: Accuracy = 83.7%
Iacucci et al[47]	Videos	CNN	PICaSSO ≤ 3 (VCE): AUC = 0.94; UCEIS ≤ 1: AUC = 0.85

¹Derived from capsule endoscopy.

Select studies showing the performance of common deep learning models in assessing endoscopic healing by Mayo endoscopic score, endoscopic Mayo score, ulcerative colitis endoscopic index of severity, and Paddington international virtual chromoendoscopy score. VCE: Virtual chromoendoscopy; CNN: Convolutional neural network; RNN: Recurrent neural network; DNUC: Deep neural network for the evaluation of ulcerative colitis; MES: Mayo endoscopic score; AUC: Area under the receiver operating characteristic curve; EH: Endoscopic healing; eMS: Endoscopic Mayo score; UCEIS: Ulcerative colitis endoscopic index of severity; PICaSSO: Paddington international virtual chromoendoscopy score.

Manual removal of low-quality images is a time-consuming but necessary step in the AI workflow as low-quality data can affect both training and implementation of AI models. However, this is impractical in clinical practice. One method to overcome this barrier is by applying an image classifier that can detect motion blur, focus, debris and lighting to exclude low-quality images prior to analysis[33-36]. Using this method, Yao et al[33] developed an autonomous CNN-based CAD to predict MES with excellent correlation to human reviewers, with the image classifier obtaining an AUC of 0.961 for deciphering gradable vs ungradable images.

Endoscopic analysis occurs in real time in clinical practice and clinical trials. While analysis of still images is an important step in demonstrating feasibility of AI models in endoscopy, full video analysis is required for real-world implementation. Though the feasibility of this approach was first shown by Yao et al[33], Gottlieb et al[35] demonstrated that it was possible to scale this to the level of a clinical trial. In this study, their recurrent neural networks was found to correlate well with central readers, with a quadratic weighted kappa of 0.844 for endoscopic Mayo Score (eMS) and 0.855 for the UC Endoscopic Index of Severity (UCEIS). Additionally, their model was able to accurately predict eMS and UCEIS. However, despite the model having excellent overall performance, individual level predictions for eMS and UCEIS, especially mid-range values, were less accurate. This suggested potential difficulties with annotation of videos due to the quality of images or variability in reviewer assessment. Nonetheless, these results are promising and suggest that DL models may one day allow for more objective assessment in clinical trials.

Histological remission has become an important therapeutic target in IBD due to its association with a lower risk of adverse events such as future corticosteroid use and hospitalization[7,37-39]. Visual analysis of WLE cannot accurately assess histological remission[40,41]. However, the deep neural network for evaluation of UC in both retrospective and prospective studies, has shown promising results in predicting endoscopic and histological remission with accuracy > 90% from both still images and full-length videos. It also demonstrated excellent correlation in predicting UCEIS compared with expert reviewers with an intraclass correlation coefficient of 0.927[34,42,43]. Additionally, in a prospective study, Takenaka et al[42] showed the ability of deep neural network for evaluation of UC to prognosticate, with no significant differences in its ability to predict relapse risk, corticosteroid use, colectomy or hospitalization as compared to experts. Thus, the studies from Takenaka et al[34,42,43] highlight many of the potential benefits of AI incorporation into routine endoscopic evaluation. For example, decreasing biopsy frequency and improving endoscopic efficiency may reduce resource utilization and decrease cost to healthcare systems. Additionally, the ability to prognosticate may alert clinicians to those who need more intensive management, potentially reducing the need for higher levels of care such as inpatient admission. Thus, these innovations could help to improve clinical workflows, decrease healthcare resource utilization, and improve patient outcomes.

To date, many DL models have demonstrated their ability to predict traditional endoscopic scores and their correlation with clinical outcomes and prognosis[42]. However, IBD is heterogeneous and scoring models should reflect that heterogeneity. In a retrospective study using data derived from A Study to Evaluate the Safety and Efficacy of Ustekinumab Induction and Maintenance Therapy in Patients with Moderately to Severely Active UC and the Janus kinase inhibitors for UC trials, Stidham et al[36] developed the Cumulative Disease Score (CDS), which measured the burden of disease by squaring MES values from 50 evenly spaced left-sided colonic segments to generate a spatial map of disease activity, with a maximum disease score of 450. By generating a topographic map of disease burden and an expanded scoring range, the system was more sensitive to endoscopic changes than MES alone. The sensitivity of such a system could have major impacts in several areas of IBD including improving power in clinical trials and characterization of partial responders who would typically require escalation of biological therapies[36]. However, further work is needed for prospective validation and to examine if CDS is applicable to right-sided disease.

Advanced endoscopy

Unlike traditional WLE, advanced endoscopy such as virtual chromoendoscopy (VCE), confocal laser endoscopy (CLE), and endocytoscopy allow for closer examination of mucosal surfaces[44]. However, access to these technologies is limited and interpretation is often difficult. AI may allow for further adoption of these technologies by reducing user dependency and improving efficiency of their analysis.

VCE allows for detailed assessment of mucosal surfaces and vessels through contrast enhancement and can be used to assess disease activity in UC using the Paddington International VCE Score[45,46]. Initial results from a prospective, multi-center study using a CNN-based CAD applied to VCE have been promising and appear to outperform WLE in assessing endoscopic activity with AUC of 0.94 and 0.85, respectively[47]. However, the CNN was trained only on rectum and sigmoid images and were obtained by an endoscopist with experience in optical imaging which may bias the model’s performance. Furthermore, when the same CNN-based CAD was applied to WLE and VCE for histological assessment, no significant differences were observed[47]. This suggests that hierarchical feature extraction of the CNN-based CAD systems from WLE may be sufficient for histological analysis, potentially limiting the utility of CNN-assisted VCE clinically.

Endocytoscopy allows for up to 1400-fold magnification for detailed assessment of the intestinal mucosa[48]. Similarly, CLE allows for high-resolution mucosal imaging and can distinguish between lesions in UC vs CD[48,49]. However, due to the need for highly trained endoscopists, few studies have evaluated the application of AI to endocytoscopy and CLE. One retrospective study based on endocytoscopy showed the utility of a SVM-based CAD. The CAD achieved an accuracy of 91% for detecting histologically active inflammation in all test segments as well as in samples labeled with a MES 0-1[25]. Additionally, in a small retrospective study, Quénéhervé et al[49] used logistic regression (LR) to extract mucosal features from CLE images to develop the IBDiag score, which was able to distinguish patients with IBD and without IBD with 100% sensitivity and specificity and delineated between UC and CD with 92.3% sensitivity and 91.3% specificity. Although few facilities have the capability to perform endocytoscopy or CLE, these advanced endoscopic techniques may have a niche in helping to further classify patients in whom the diagnosis of IBD is uncertain or in patients with indeterminate colitis.

Lastly, although technically based on WLE, Bossuyt et al[50] developed a red density (RD) algorithm derived from red pixel intensity in the red-green-blue channels in images from 29 patients with UC and 6 healthy controls. The RD score correlated well with MES, UCEIS, and Robarts Histopathology Index, with an RD score cutoff of 60 having 96% sensitivity and 80% specificity for detecting histologic remission in their prospective cohort. One benefit of this technique over other advanced imaging modalities, is that the red channel can be extracted from any traditional endoscope. However, further investigation is needed as the study used a prototype Pentax endoscope and processor which means that the RD score may not be applicable to other endoscopes currently on the market.

Capsule endoscopy

Capsule endoscopy (CE) enables non-invasive imaging of the small intestine and colon and is frequently used in the diagnosis and monitoring of CD[51-53]. Each CE session captures tens of thousands of images, making manual review time-consuming. Additionally, high interobserver variability in reading CE is common due to the length of CE, reader experience, and fatigue[54,55]. Early studies in CE focused on traditional ML models such as SVM to detect common lesions in CD such as ulceration[56,57]. However, more recent studies use DL models for the detection of common mucosal pathologies in CE (Table 2).

Table 2 Performance of deep learning models in capsule endoscopy for the assessment of common pathologies in inflammatory bowel disease.

Ref.	Model	Key performance metrics
Klang et al[58]	CNN	Ulcers vs normal mucosa: AUC = 0.94-0.99
Barash et al[61]	CNN	Ulcer grade 1 vs 3: AUC = 0.958; ulcer grade 1 vs 2: AUC = 0.565; ulcer grade 2 vs 3: AUC = 0.939
Ribeiro et al[67]	CNN	Ulcers and erosions: AUC = 1.00
Kratter et al[69]	CNN	Ulcer detection (cross-domain model): AUC = 0.921-0.948; ulcer detection (combined model): AUC = 0.984-0.998
Klang et al[59]	CNN	Strictures vs ulcers: AUC = 0.942; strictures vs mild ulcers: AUC = 0.992; strictures vs moderate ulcers: AUC = 0.975; strictures vs severe ulcers: AUC = 0.889
Majtner et al[65]	CNN	SB ulcers: Accuracy = 98.5%; LB ulcers: Accuracy = 98.1%
Ferreira et al[66]	CNN	SB and LB ulcers: AUC = 1.00

Select studies showing the performance of common deep learning models in capsule endoscopy in the evaluation of ulcers, erosions, and strictures. SB: Small bowel; LB: Large bowel; CNN: Convolutional neural networks; AUC: Area under the receiver operating characteristic curve.

Ulcers and strictures are classic findings in CD. Early retrospective studies of CNN-based CAD models demonstrated accuracy in identifying at least 95% of ulcers as compared to healthy mucosa[58]. These results were further validated in a follow-up study, which showed a CNN-based CAD had the ability to accurately differentiate between ulcers, strictures, and normal mucosa with an AUC of 0.889 to 0.992[59]. Though promising, both studies relied on data sets derived from small cohorts, with relatively limited pathology, which may affect each model’s ability to generalize to larger populations.

Depth and severity of ulceration are important to determine as it is associated with risk for surgery in CD[60]. In this context, a retrospective study by Barash et al[61] showed that an ordinal CNN, which encodes label order as compared to conventional CNNs, can accurately differentiate between grade 1 and 3 ulcers with an accuracy of 91%, but similar to humans, the model had less ability to differentiate between grade 1 and grade 2 ulcers with an accuracy of 62.4%. This limitation may be due to a lack of ground truth in image annotation as well as the small number of images, 1242, used for training, which may have subsequently affected the performance of the model. Nonetheless, initial studies demonstrated the potential capabilities of DL models to efficiently and accurately quantify disease burden in the small intestine.

Traditionally, CE has been used for the assessment of small bowel pathologies. However, with the PillCam Crohn’s Capsule, PillCam Colon Capsule, and PillCam COLON2 it is also possible to quantify disease burden in the large intestine[62-64]. Several CNN-based CAD models have been applied to analysis of PillCam Crohn’s Capsule and the PillCam Colon Capsule and have shown the ability to reduce analysis time and accurately detect intestinal ulceration and erosions[65-67]. Additionally, similar to applications of AI to WLE, a prospective study using a ResNet50-based CNN applied to PillCam COLON2 images was able to predict MES 0-3 with accuracies of 91.3% to 99.4% in the validation cohort while also creating a topographic map of disease activity in patients with UC[68]. Though promising, the model’s performance may have been affected due to more than 50% of images being of inadequate quality for scoring. Nonetheless, the prospect of measuring total disease burden presents exciting opportunities for further understanding disease phenotypes.

Though pan-enteric CE grading presents a potentially novel way to evaluate patients with IBD, the widespread implementation of several AI models to assess CE remains challenging. To address this, recent efforts toward developing a universal multi-domain CNN framework to assess CE images from two different capsules has shown promising results, achieving an AUC of 0.92 to 0.95 for differentiating healthy and ulcerated mucosa in cross domain testing, though accuracies ranged from 56.9% to 88%[69]. This was further improved in a combined model, which achieved a mean AUC of 0.99 and accuracy of 97.4% when a single CNN was trained on images from two separate CEs[69]. This underscores the importance of diverse training sets from multiple sources to improve model performance and prevent overfitting, which will allow for generalization of these models to large populations. Though further validation is needed, AI models such as this are one step toward the development of a larger universal model for image analysis in CE.

To date, many DL models have demonstrated the ability to accurately detect common IBD pathologies. However, rather than use DL models to predict disease activity in IBD, DL may also be able to reduce the burden of CE analysis by filtering images for review. This could allow for CE analysis to be performed in a fraction of the time as demonstrated by the AXARO^® framework which decreased average review time of pan-enteric CE to less than 4 minutes with a 97% negative predictive value for IBD[70]. Systems such as this may allow for efficient exclusion of patients without IBD, negate the need for further invasive testing and ultimately lead to lower resource utilization and cost.

AI IN NON-INVASIVE IMAGING

CTE, MRE, and IU are widely available and validated for disease assessment in IBD where they are important in assessing and monitoring transmural disease and complications including strictures, internal fistulas, and perianal disease[71-74]. However, several challenges exist which increase the risk of interobserver variability including differences in image acquisition protocols between institutions, need for specialized training, and subjective interpretation by radiologists. Given these limitations, AI may have the potential to standardize and streamline imaging in IBD.

CTE

CTE is a validated tool for the assessment of IBD with many studies exploring the applications of AI in small intestinal CD[73]. In CTE, specific features such as mural attenuation and bowel wall thickness correlate with active inflammation in CD[75,76]. When used as inputs into an image fusion model using an SVM classifier, features such as mural hyperenhancement, mural stratification, mesenteric hypervascularity, mesenteric fibrofatty proliferation, mesenteric fat density, intestinal fistula and bowel strictures combined with radiomics features accurately predicted moderate to severe Simple Endoscopic Score for CD achieving an AUC of 0.896 in the validation cohort[18]. However, there are drawbacks to using Simple Endoscopic Score for CD as it may not capture the heterogeneity of inflammation seen in CD. To address this, Stidham et al[77] developed the simplified cumulative intestinal disease severity score. This was accomplished through image segmentation to create mini-segments followed by radiomic feature extraction and RF training to replicate expert grading on these segments. Each segment was then used to build a spatial map of disease severity and a CDS which correlated well with radiologist scoring, with a correlation coefficient of 0.971. Additionally, the RF model had good agreement with radiologists in segment grading, Cohen’s kappa 0.80, and future risk of surgical intervention within 3 years with an AUC of 0.76. Similar to the CDS applied to WLE, the ability for ML to develop topographical maps of disease severity with CTE may allow for more accurate quantification of disease activity and burden[36].

Fibrosis is often seen in cases of CD and is quantifiable on CTE. Identifying fibrosis in CD is important due to its associations with increased morbidity[78]. Compared to radiologists, a radiomics classifier using LR trained on 1454 extracted radiographic features on CTE showed superior performance in distinguishing between none-to-mild vs moderate-to-severe fibrosis[17]. Similarly, a 3-dimensional (3D) CNN model outperformed radiologists in grading the severity of fibrosis. Though the performance was comparable between the 3D CNN and radiomics model using LR, processing time was significantly faster with the 3D CNN, highlighting the efficiency of DL models in imaging-based tasks[79].

Mucosal healing is associated with improved outcomes and radiomics models and ML classifiers have shown promising results in predicting mucosal response following therapy initiation in CD[6,80]. Notably, ML models may have further utility in differentiating disease where clinical symptoms and gross endoscopic features may overlap. For example, Yang et al[81] developed a LR model using retrospective CTE data such as mucosal enhancement and fibrofatty proliferation to assist in differentiating CD from intestinal Behçet’s disease. Their study demonstrated that using CTE features in addition to both endoscopic evaluation and assessment of clinical symptoms could successfully differentiate the two diseases with an accuracy of 84.15% as compared to either endoscopy alone or endoscopy combined with clinical symptoms.

MRE

Compared to CTE, MRE is more time-consuming and expensive, but spares patients from ionizing radiation, and certain protocols can spare patients from contrast dye. Similar to CTE, MRE can evaluate wall thickness, edema, relative contrast enhancement (RCE), and ulceration which correlate with histopathology in patients with IBD[14,71]. Many prior studies in MRE have focused on the use of texture analysis, semi-automated quantification of bowel wall thickness, and motility to characterize inflammation[82,83]. These features now serve as a foundation for AI models in MRE.

Beyond semi-automated analysis, several studies have employed DL models to improve image quality to enhance detection of IBD. Lian et al[84] used an optimized CNN to reconstruct MR images with improved signal-to-noise ratios and diagnostic sensitivity in IBD. Similar results were obtained in a more recent study using DL reconstruction leading to improved image quality and interobserver agreement[85]. An additional study using a CNN-based model demonstrated the potential of 3D reconstruction of pelvic MR images in patients with CD to aid in surgical planning. Though this technology requires further validation prior to implementation, it holds the potential to one day improve outcomes by allowing for enhanced visualization of complex perianal disease such as fistulas[86].

The tortuous nature of the small intestines makes accurate measurement of structural abnormalities difficult on MRE. Curve planar reformatting is one method for visualizing the small intestine in two-dimensions but manual segmentation is time intensive. However, curve planar reformatting paired with a CNN-based segmentation model applied to contrast-enhanced T1-weighted MR images has demonstrated promising results for improving visualization and quantitative assessment of disease burden in CD as it applies to assessing stricture length and bowel wall thickness[87]. Rather than assess anatomic changes, van Harten et al[88] used stochastic tracking with a CNN-based orientation classifier to assess small bowel motility in 3D cine MR. Advancement such as this may improve the ability to assess the functional impact of strictures and inflammation in IBD, which may provide new insights into the pathophysiology of IBD and one day guide therapeutic decisions.

IU

IU is becoming commonplace in clinics worldwide likely due to its low cost, lack of radiation or need for contrast, and its ability to impact patient care in real time. IU may be especially helpful for certain populations such as young patients, pregnant patients, or those with a large body habitus that may not accommodate computed tomography or magnetic resonance imaging scanners[74]. In this context, one recent development by Carter et al[89] was the creation of an AI model that could aid potentially inexperienced IU operators in accurately assessing inflammation in CD. They used a CNN model to recognize bowel wall thickening > 3 mm on IU as a surrogate for active inflammation with an overall accuracy of 90.1%. More recently, a small feasibility study by Gu et al[90] demonstrated that a radiomics-based model using extreme gradient boosting could classify normal from abnormal intestinal images with bowel wall thickening > 3 mm with 93.7% accuracy compared to a CNN-based classification model with an AUC of 0.754. Though these two studies showed the capability of AI in automating bowel wall assessment, it is important to note that Carter et al[89] used a larger dataset for training, which may partly explain the improvement in performance of their model.

As with CTE and MRE, investigation is actively ongoing into the utilization of AI in IU, and its application will likely increase significantly in the future as ultrasonography machines become more widespread and affordable. Here, AI may have the potential to aid in standardization of data retrieval and interpretation given that operators in clinics throughout the world likely have widely variable training backgrounds.

AI IN HISTOPATHOLOGY

Like endoscopy, AI may improve both the efficiency and standardization of histological scoring[91]. Several studies have demonstrated the abilities for CNNs to extract specific cellular features from whole slide biopsies. For example, Vande Casteele et al[92] developed a CNN to assess eosinophil counts on whole slide colonic biopsies from patients with UC with almost perfect agreement with pathologist, achieving an intraclass correlation coefficient of 0.805 to 0.917. However, they also found that eosinophil density was not associated with histological activity, even during subgroup analysis when corticosteroid use was excluded. However, the limited sample size of 88 patients may have contributed toward this finding and further work is needed to validate these results.

In regard to neutrophil quantification, Gui et al[93] established the Paddington International VCE Score Histologic Remission Index (PHRI) based upon the presence of neutrophils. Using a CNN-based CAD, they demonstrated the ability for their CNN to correctly classify neutrophils with 88% accuracy and predict histological activity with 86% accuracy. Similarly, Ohara et al[94] used semantic segmentation to separate the epithelium from the lamina propria, and then assessed for the presence of neutrophils using an object-detection based DL model, YOLOv5. This model showed strong agreement with three expert pathologists when predicting the Nancy Histological Index (NHI) and PHRI but also provided exact localization of neutrophils allowing insight into the model’s quantification. However, the authors note that the model did not incorporate mucosal ulcers and other features of active UC which would allow it to determine higher grades of PHRI and NHI, potentially limiting its extrapolation to patients with severe disease.

The intestinal mucous layer may play a role in susceptibility in IBD. In this context, a single-center retrospective study performed by Ohara et al[95] demonstrated the ability for a CNN to classify goblet cell mucus area, which correlated with both disease activity and risk of clinical relapse over 12 months in patients with UC. However, only a small portion of patients in the cohort experienced relapse, and thus further studies are needed for validation. However, a model such as this may provide valuable insight to clinicians during routine IBD monitoring and flag which patients may require earlier intervention.

The development of ML and DL models to automate histological indices would be beneficial for clinical practice and trials. To this end, DL models such as CNNs have demonstrated the ability to predict Robarts Histopathology Index, PHRI and NHI scores, in addition to endoscopic disease activity and risk for flare occurrence[47,96,97]. In another recent study, Rymarczyk et al[98] developed a model which used a CNN to extract features from whole slide biopsies for use as inputs into an attention-based multi-instance learning model. The model was able to accurately predict subgrades of the Geboes and Global Histology Activity Score with moderate agreement and an accuracy of 65% to 89%. However, the study was limited in that it relied on a single central reader which may impart bias on ground truth labeling during model development, limited data for training on ileal tissue, and lacked external validation, possibly affecting the model’s accuracy.

Explainable AI will be important for clinical integration and quality control monitoring by both experts and regulatory agencies. To this end, Najdawi et al[99] used a CNN which output human interpretable features such as cell classification which were then used as inputs into an RF model to predict the NHI with a weighted kappa of 0.91 and spearman correlation of 0.89 when compared to pathologists grading. However, the novelty of this model was that it allowed for a look into the opaque-decision making of the CNN model by taking pixel level predictions from the CNN and overlaying them onto each slide. This type of explainable AI is promising and may one day be used to train junior pathologists, allow for real-time quality control, and build trust in AI systems.

Though AI has shown promising results in standardizing histological assessment in IBD, new frontiers are emerging through the development of Endo-Histo-Omics, which combines histologic, endoscopic, and omics data to further understand IBD phenotypes[100,101]. In this context, a small prospective trial by Iacucci et al[101] involving 29 patients with IBD on anti-tumor necrosis factor inhibitors or vedolizumab, combined CAD-analysis of probe-based confocal laser endomicroscopy, ex-vivo fluorescein-labeled biologic binding, and gene expression analysis to predict response to treatment. In their study, they found that markers such as crypt morphology and vessel tortuosity and tissue level-biologic binding were predictive of response to therapy. Additionally, seven differentially expressed genes (DEGs) including chemokine (C-X-C motif) ligand 6 were predictive of response to anti-tumor necrosis factor inhibitors therapy with an AUC of 0.862 in the validation cohort. With this technology, Endo-Histo-Omics offers a novel approach to further understanding disease phenotypes in IBD and may one day allow for personalized therapy.

LLMS AND CHATBOTS

Patients with chronic conditions such as IBD often face challenges with symptom management, pain control, and frequent healthcare visits. With the rise in demand for healthcare services, LLM-based chatbots may offer patients and providers new and convenient avenues for accessing medical information.

Patients with IBD frequently inquire about symptoms, medications, and finances and NLPs can categorize these requests, demonstrating their potential utility in the management of IBD[102]. For example, in one study, the LLM ChatGPT-4.0 performed well when asked 88 questions related to nutrition and IBD, providing accurate information 83% of the time, demonstrating its potential as a source for nutritional advice[103]. When it comes to medical management of IBD, ChatGPT-3.5 showed promising results for common questions around topics such as pregnancy and IBD but provided inaccurate information on questions regarding oral therapies[104]. The inaccuracies in both studies are likely multifactorial in the setting of ChatGPT using outdated information due to its inability to extract real-time updates from medical resources such as PubMed. However, it also underscores a fundamental difference between general LLM-based chatbots such as ChatGPT and domain-specific LLM-based chatbots. Whereas LLMs like ChatGPT are trained on large data sets, the training data is not purely domain specific which allows it to generalize, and interact in a more natural manner, but risks erroneous results. This contrasts with domain-specific LLMs, which are often trained on focused data sets for specific tasks, allowing for more accurate and detailed responses, though their scopes are often limited[103]. For these reasons, ChatGPT and general LLMs may be a more valuable source of information for healthcare professionals who are able to verify outputs of these models[22]. Nonetheless, the potential for domain-specific LLM-chatbots to improve both patient and clinician education and improve clinical workflows through reducing administrative tasks such as inbox management and note-writing are promising, with early studies receiving positive feedback from clinicians (Figure 2)[105,106].

Figure 2 Limitations of traditional management of administrative duties and potential improvements with applications of artificial intelligence. Figure created with BioRender.com (https://BioRender.com/ygwivmn). In traditional inflammatory bowel disease management, patients may send clinicians messages, that are normally screened by ancillary staff, prior to a response from a clinician. During clinic visits, clinicians are often multi-tasking to improve efficiency. This often leads to increased staff burnout, patient dissatisfaction, and high administrative burden. In an artificial intelligence (AI)-assisted model, patients may be able to use chatbots to answer basic questions about their disease. If they require more specific advice, patients may send clinicians messages, with responses being AI-generated with clinician oversight for accuracy. During AI-assisted appointments, notes are written by AI-assisted technology, decreasing administrative burden. This may lead to improvements in clinician and patient satisfaction. AI: Artificial intelligence; LLM: Large language model.

BIOLOGICAL PREDICTION MODELS

Biological medications have revolutionized the treatment of IBD. However, biologics are expensive, and many patients may be primary non-responders or partial responders to therapy[107]. Studies suggest that patients who are biologic-naive tend to have a better response to therapy than patients who are biologic-exposed[108,109]. Thus, predicting who may respond to one treatment over another in early stages of IBD is critical. Currently there is no established method for predicting who will respond well to a biological therapy. Consequently, there is growing interest in using AI to develop disease prediction models which may allow for the implementation of individualized treatments in IBD.

Many studies have focused on identifying DEGs in response to biologic therapy using AI models including K nearest neighbor for golimumab[110], ANN, SVM, RF, and ensemble ML for infliximab[111-114] and least absolute shrinkage and selection operator and ensemble ML methods for ustekinumab[115,116]. Though many of these studies showed some degree of correlation between their unique gene signatures and mucosal healing, remission, or biologic response, many of these studies are retrospective, derive their gene signatures from the Gene Expression Omnibus repository, and lack prospective validation (Table 3). Additionally, while DEGs may be indicative of underlying response to therapy, it is known that DEGs may differ between populations, which may limit the external validity of these gene signatures to larger populations.

Table 3 Performance of therapeutic prediction models in assessing response to common biologic therapies in inflammatory bowel disease.

Ref.	Model	Drug name	Outcome	Key performance metrics
Telesco et al[110]	LR	Golimumab	Mucosal healing at week 6	AUC = 0.688
Chen et al[111]	ANN	Infliximab	PNR at weeks 6 and 14	RNA model: AUC = 0.850
Feng et al[112]	RF, ANN	Infliximab	PNR at week 14	AUC = 0.81
Ghiassian et al[118]	PNN	Infliximab	Inadequate response at weeks 4-6 and 8	AUC = 0.83
Mishra et al[120]	RF	Infliximab	Clinical remission at week 14	DNAm-DEGs: AUC = 0.88
Derakhshan Nazari et al[113]	LASSO, ensemble ML	Infliximab	Improvement in symptoms, endoscopic activity, histologic scores	Accuracy = 73%-95%
Ye et al[114]	LASSO, RF, SVM-RFE	Infliximab	Endoscopic or histological activity at weeks 6 and 14	AUC = 0.912-0.920
Zhang et al[121]	LASSO	Infliximab	PNR at week 14 vs NDR at weeks 22-52	AUC = 0.812-0.888
Konikoff et al[129]	CatBoost	Infliximab and vedolizumab	Drug sustainability at week 54	AUC = 0.86
Joustra et al[119]	Stability selected gradient boosting	Tofacitinib	Disease response after week 8	AUC = 0.74
Lees et al[126]	RF, LR	Tofacitinib	Partial Mayo score response at weeks 4 and 8	AUC = 0.74-0.85
Chaparro et al[124]	GAM	Ustekinumab	Clinical remission	AUC = 0.796
Doherty et al[122]	RF	Ustekinumab	Therapeutic response at week 6	AUC = 0.733-0.844
He et al[115]	LASSO	Ustekinumab	Disease response by CDAI	AUC = 0.734
Koustenis et al[116]	XGBoost, XRT, DRF GBM, GLM	Ustekinumab	Clinical remission at 6 months	AUC = 0.925-1.0
Morikubo et al[16]	SVM, RF, LR, KNN	Ustekinumab	Steroid-free clinical remission at week 22	Accuracy = 56.5%-69.6%
Chen et al[125]	ENRR, RF, SMOTE	Vedolizumab	Clinical remission at week 52	AUC = 0.811
Waljee et al[26]	RF	Vedolizumab	Corticosteroid-free biologic remission at week 52	AUC = 0.65-0.75
Waljee et al[123]	RF	Vedolizumab	Corticosteroid-free endoscopic remission at week 52	AUC = 0.62-0.73
Dulai et al[127]	LR (VDZ-CDST)	Vedolizumab	Clinical and corticosteroid-free remission, mucosal healing, deep remission at week 26	AUC = 0.66-0.75
Alric et al[128]	LR (VDZ-CDST)	Vedolizumab and ustekinumab	Steroid-free clinical remission at week 48	Vedolizumab: AUC = 0.69; ustekinumab: AUC = 0.52

Select studies showing the performance of machine learning models in predicting various forms of response to therapy such as primary non-response, non-durable response, clinical, endoscopic or histologic remission. LR: Logistic regression; ANN: Artificial neural network; RF: Random forest; PNN: Probabilistic neural network; LASSO: Least absolute shrinkage and selection operator; SVM: Support vector machine; RFE: Recursive feature elimination; GAM: Generalized additive model; XGBoost: Extreme gradient boosting; XRT: Extreme random trees; DRF: Distributed random forest; GBM: Gradient boosting machine; GLM: Generalized linear model; KNN: K-nearest neighbors; ENRR: Elastic-net-regularization regression; SMOTE: Synthetic minority over-sampling technique; VDZ-CDST: Vedolizumab clinical decision support tool; CDAI: Crohn’s disease activity index; DNAm: DNA methylation; DEGs: Differentially expressed genes; ML: Machine learning; AUC: Area under the receiver operating characteristic curve.

One approach to addressing the limitations of DEGs is to map them to the human interactome to allow for the identification of connected protein networks which may correlate with treatment response[117]. For example, Ghiassian et al[118] was able to identify the largest connected component of proteins indicative of response to infliximab and used the genes in this network as inputs into a probabilistic neural network. Overall, the system performed well with a 100% positive predictive value and 64% sensitivity for identifying inadequate response to infliximab. Though this cross-cohort study was small, the identification of a largest connected component demonstrates that there may be shared pathways which are indicative of a response to infliximab therapy and help to overcome some of the barriers we currently face in biomarker discovery.

Aside from DEGs, DNA methylation is important in gene regulation, and alterations in DNA methylation may help predict response to therapies. For example, gradient boosting applied to peripheral DNA methylation profiles at baseline can predict tofacitinib response in UC at week 8 with an AUC of 0.74[119]. Similarly, Mishra et al[120] applied a RF model to baseline and week 2 methylation and gene expression profiles to predict response to infliximab in CD at 14 weeks with an AUC of 0.88 in their validation cohort. Finally, polymorphisms may play an important role in genetic susceptibility in IBD and may also be useful in predicting non-response to infliximab[121].

The microbiome has been a source of great interest in recent years for its role in the pathogenesis of IBD and may play a role in response to therapies. For example, using baseline fecal microbiota and clinical data, a RF model was able to accurately predict remission with ustekinumab as early as 6 weeks in CD with an AUC of 0.844[122].

Serum drug levels for some biologics are routinely measured in clinical practice and may be used to predict early response to therapy. For example, an RF model using vedolizumab serum levels, demographics, clinical, and laboratory data derived from GEMINI I and GEMINI II through week 6 can moderately predict corticosteroid-free biologic remission at 52 weeks in CD and endoscopic remission in UC, with AUC of 0.75 and 0.73, respectively[26,123].

Other ML models have relied on clinical and laboratory data without the need for serum drug levels to predict outcomes in UC and CD[16,124]. For example, including additional baseline inputs such as past medical history into an elastic net regularized regression may be able to further improve the ability to predict treatment response in patients on vedolizumab highlighting the unique role underlying comorbidities may have in predicting therapeutic response[125]. Another study used clinical data such as C-reactive protein at baseline and week 2 from OCTAVE Induction 1 and 2 to train their RF and LR models. Using these variables, Lees et al[126] was able to predict tofacitinib responders with 84% to 87% accuracy at week 4 and 74% to 79% accuracy at week 8 during induction, further demonstrating that serum drug levels may not be necessary to accurately predict therapeutic response to some biological therapies.

A clinical decision tool that mines data from the health record to help inform decisions on biologic therapy is one step toward personalized medicine in IBD. Using LR, Dulai et al[127] identified factors such as albumin were associated with achieving various stages of remission. When weighted, these factors were used to create the vedolizumab clinical decision support tool (VDZ-CDST) which stratified patients into high, intermediate, or low probability of responding to vedolizumab. Interestingly, at a cutoff of 13 points, the model had 92% to 100% sensitivity, but 25% to 31% specificity for prediction of all forms of remission at 26 weeks in patients with CD. Though promising, the model had poor calibration during external validation in Vedolizumab for Health Outcomes in IBD, which the authors suggest may be due to differences in case severity and outcome definitions. Despite the poor calibration, a follow up study achieved similar results at 48 weeks for patients on vedolizumab when using the VDZ-CDST, which underscores that poor calibration does not necessarily affect the sensitivity of the model[128]. Though when applied to CD patients on ustekinumab, VDZ-CDST was not correlated with clinical or steroid free clinical remission, indicating it may not be universally applicable[128]. However, other AI-based tools such as the gradient boosted decision tree model by Konikoff et al[129] showed that a single AI model to predict response to multiple biologic therapies in UC may be possible. In their study, they showed that a single model may be able to predict drug sustainability at 54 weeks for patients on vedolizumab or infliximab with an AUC of 0.86. Additionally, using SHapley Additive exPlanations, which assigns a Shapley value to each feature in a model, they were able to demonstrate which features were most influential to the model’s predictions. This type of explainable AI will be important for personalized therapeutic selection in clinical practice, while allowing for physician oversight of model performance.

While there have been many studies over the past several years that demonstrate the potential of prediction models to usher in an era of precision medicine, early results must be interpreted with caution. In the near term, clinically-derived predictive models such as VDZ-CDST are likely closer to clinical implementation due to the ease of access to common markers such as C-reactive protein and their established guidelines for clinical use. Although biomarker-based models which incorporate DEGs may one day offer an additional level of accuracy to these models, the lack of standardized assays, cost, and lack of reliable gene signatures will likely impede their near term clinical integration. However, regardless of the model, given that many of these models are trained on small retrospective cohorts, they are at risk of overfitting, which may affect their ability to generalize to larger populations. Thus, rigorous external validation will be required to refine these models prior to clinical integration.

CURRENT BARRIERS TO CLINICAL IMPLEMENTATION

Implementation of AI systems into clinical practice and trials has the potential to change the way we manage IBD. However, many practical, technical, and ethical concerns remain that preclude its widespread adoption.

Cost

Cost is a major factor in the implementation of AI into clinical practice for healthcare systems. To offset this, fee-for-service (FFS) payments may partially cover reimbursement for AI through the New Technology Add-on Payment. However, questions remain, should we reimburse for AI at all? Several concerns surround the use of FFS, and there are numerous examples of how FFS has led to overutilization of healthcare resources such as medical imaging[130]. This is particularly true with AI, where the clinical benefit of implementation into daily practice remains unclear, and if not carefully incorporated into daily practice, could lead to a rapid rise in billing for new technology. One way to circumvent this could be by providing incentives for efficiency[131]. If AI improves efficiency through aiding diagnoses and treatment, then there should be significant cost savings to health systems in the long-term which may offset the cost of AI implementation[132]. One example of these cost savings is demonstrated in a recent study by Thiruvengadam et al[133] who developed a microsimulation model to show the potential cost savings of CAD in screening colonoscopies. This study demonstrated that even small improvements in adenoma detection rate could lead to long-term cost savings to the payor through reduced rates of colorectal cancer. Though similar studies are lacking in IBD, the potential use of AI to reduce cost through improvements in endoscopic efficiency and reduced invasive interventions are plausible.

Liability

In addition to cost, legal liability is also a concern. For example, if a patient receives a colonoscopy and AI misses an adenoma, but then later develops malignancy, who is responsible? One way of assigning liability could be through the level of AI autonomy[134,135]. In the case of the use of a CAD, the physician who misses the adenoma would likely remain liable. But with higher levels of AI autonomy, for example, autonomous without human backup, liability may be upon several parties including the company, individual developer, or institution[134]. It may be possible to reduce misdiagnoses and liability using explainable AI, which could allow for more oversight by both physicians and regulatory agencies[24,136]. However, to date, there remains a lack of legal precedent, likely due to the rapid development of technology in this area.

A different but equally important question arises when AI is used by patients to self-diagnose. Studies have shown that patients are at least willing to use LLMs such as ChatGPT to self-diagnose[137]. However, with any tool such as AI or even internet search engines, there is a risk that patients delay care based upon their own research. As this technology advances, and more AI-tools are approved under the Food and Drug Administration Software as a Medical Device (SaMD), further safeguards are warranted to reduce the risk of patient harm and liability. For example, potential SaMD could be incorporated into healthcare institution’s clinical models and trigger an alert to a healthcare professional under specific circumstances, which may allow for earlier engagement with patients while allowing for physician oversight of approved medical devices.

Privacy

Patient privacy remains a major issue in healthcare, where data breaches are increasing. Data use by technology companies is becoming commonplace and may be shared with third party companies. However, training AI models requires large amounts of data. With involvement of commercial companies, questions about how this data is handled are paramount. Another risk is unintended disclosure[138]. For example, performing a prior authorization using a system such as ChatGPT could inadvertently upload sensitive patient information into a system which is not Health Insurance Portability and Accountability Act compliant, increasing the risk of data breaches[139]. Reidentification, where de-identified data is traced back to the original individual, represents a major risk for patient confidentiality. A recent survey showed that many patients do not trust health systems to use AI appropriately or provide adequate protection of their privacy[140]. Thus, privacy remains a major hurdle for large scale acceptance of AI in healthcare.

Nonetheless, several potential solutions exist which may allow us to overcome the vulnerabilities of AI use in healthcare. Blockchain technology, which uses a decentralized ledger to store information, may allow for enhanced patient protection. It also has the potential to meet several of the Health Insurance Portability and Accountability Act requirements including encryption and audit logging. However, concerns have been highlighted about the scalability of these systems, latency, and implementation with legacy electronic health records[141]. Other methods such as differential privacy, which adds random noise during model training, may help to address privacy concerns with AI development. In fact, one recent study demonstrated that reconstruction attacks under worse case-scenarios were < 0.05% in large datasets such as RadImageNet, leading the authors to recommend that differential privacy should be used in all sensitive data sets[142]. Regardless of the tool, data management and cybersecurity should be discussed at all stages of AI development, which will help build trust in these systems and safeguard against unwarranted release of private healthcare information.

Disparities

Disparities in healthcare are commonplace. It is well known that minorities are less likely to participate in clinical research and undergo screening colonoscopies[143]. This increases the risk that models will be under-trained on under-represented populations, perpetuating these inequalities[144]. To ensure adequate model performance, future studies should incorporate both diversity in disease phenotypes and demographics into their training models. This is especially important to consider given that one of the benefits of AI could be the dissemination to rural and underserved communities where health care resources are scarce.

FUTURE DIRECTIONS AND EMERGING TRENDS

Early studies in AI have demonstrated the potential impacts it may have in the management of IBD. However, many technical barriers remain which should be addressed by future research efforts[145]. Establishing a ground truth in image annotation remains challenging, for example, with MES where neighboring values are difficult to discern[36]. However, this problem is not necessarily the result of poor performance of DL models, but rather, models are unable to surpass human performance in these areas due to subjectivity of scoring systems with which these models are initially trained. Thus, future studies may benefit by either developing new scoring systems or moving toward less subjective scoring indices for model development.

Additionally, many models to date are trained and validated in small retrospective cohorts. Caution must be used when interpreting many of these models as they may be overfit and generalize poorly to larger populations. Thus, further rigorous prospective validation is needed. To reduce the risk of overfitting and poor generalization, future studies should focus on multi-center trials with large diverse datasets. Though the cost associated with these types of studies may impede validation of these models, newer strategies are emerging to help reduce the number of patients needed for model development. For example, generative AI models such as GANs are capable of synthesizing realistic images which may help to augment smaller data sets, rare disease phenotypes, and under-represented populations. Newer models such as diffusion models, which synthesize data through denoising, have shown improved performance compared to GANs albeit at the cost of higher computational demands. Nonetheless, this technology is promising and may help to overcome a major barrier in model development[28]. Federated learning could also help to address sample size and diversity. In this context, federated learning would allow for multi-institution engagement in model development, with patient data being stored locally and local model updates being sent to a central model for aggregation. Due to local storage of patient data, federated learning also helps to address some of the privacy concerns regarding storage of patient data in a centralized system[146].

Cross platform compatibility of these models will also be important for future dissemination as many centers often have access to different equipment and computing capabilities. Future studies should focus on incorporating images and data from different instruments and manufacturers such as that seen in a prior study on multi-domain CEs which may help to reduce cost of implementation and improve access to medical centers[69].

As we continue to move toward personalized medicine, further understanding IBD phenotypes is an important goal. Endo-Histo-Omics offers a novel approach for further delineating these phenotypes[147]. However, other novel technologies such as quantum computing may further allow us to understand how a patients unique biology contributes to therapeutic response[148]. For example, though quantum computing has yet to be applied in the field, it may one day allow us to further understand the molecular interactions between biological therapies and their receptors. The data generated from these models, when combined with clinical and omics data, may allow for new insights into therapeutic response.

Finally, as AI continues to advance in the field of IBD, the ultimate goal is to move these platforms from development to deployment. In doing so, it will be critical that future models follow the recent framework provided by the Food and Drug Administration for SaMD approval including strict safeguards for privacy, robust model validation, and transparent reporting (Figure 3)[149].

Figure 3 Food and Drug Administration roadmap for approval of artificial intelligence as software as a medical device: There are four major areas of software development for software as a medical device. Figure created with BioRender.com (https://BioRender.com/t8xxcz4). In pre-development, the architecture and infrastructure behind the model is discussed, and involves key stake holders such as the Food and Drug Administration. In development, focus shifts towards model validation while ensuring usability of the software and refining the user interface. During approval and deployment, models may be submitted to the Food and Drug Administration through different regulatory pathways such as 501 (k). During this time, a predetermined change control plan may be submitted, which allows for premarket approval of future modifications to a device. In post-marketing, feedback is gained on the model in the real-world setting with periodic model refinements per predetermined change control plan. FDA: Food and Drug Administration; PCCP: Predetermined change control plan.

CONCLUSION

The management of IBD is complex, time intensive, costly, and is dependent upon clinician expertise. Many studies to date have demonstrated the potential impact AI could have on the way we manage IBD in the future. Most promising are developments of AI in endoscopy, histology, and non-invasive imaging, where autonomous systems may one day standardize diagnostic processes and therapeutic monitoring. Systems such as these could have broad clinical implications in IBD. Improved sensitivity for detecting changes in disease burden could improve power in clinical trials and characterize patients who were once thought to be therapeutic non-responders. Automation in these processes could also improve efficiency and diagnostic accuracy in IBD, while allowing for expert-level evaluation of disease activity in underserved areas which do not have access to IBD specialists. LLM-based chatbots are improving the way health systems interact with their patients, while also acting as another resource for medical advice to patients with IBD. Moreover, therapeutic prediction models such as VDZ-CDST are impacting the way clinicians choose therapeutic agents, with the goal of moving away from the one-size-fits-all paradigm. Though these advances are promising, more work is needed to overcome several technical, ethical, and practical barriers before AI in IBD is integrated clinically. However, despite these limitations, the field of IBD will likely continue to benefit from the development of these novel technologies with hopes that AI will one day improve access, efficiency, and precision in the management of IBD.