Artificial intelligence use for precision medicine in inflammatory bowel disease: a systematic review

Abstract

Introduction: Inflammatory Bowel Disease (IBD), encompassing Crohn’s disease and ulcerative colitis, presents significant clinical challenges due to its heterogeneous nature and complex etiology. Recent advancements in biomedical research have enhanced our understanding of IBD’s genetic, microbial, and biochemical aspects. However, persistent issues in clinical management, including treatment non-response, surgical interventions, and diagnostic uncertainties, underscore the need for more targeted approaches. This review examines the convergence of artificial intelligence (AI) and precision medicine (PM) in IBD management. By leveraging AI’s capacity to analyze complex, multi-dimensional datasets, this emerging field offers promising applications in improving diagnostic accuracy, predicting treatment responses, and forecasting disease progression, potentially transforming IBD patient care. Method: The systematic review (SR) was conducted by searching the following databases: PubMed, PubMed PMC, BVS, Scopus, Web of Science, Embase, Cochrane, and ProQuest up to February 2024. Studies that employed AI in IBD applied to precision medicine were included. Results: 139 studies on applying AI in precision medicine for IBD were identified. Most studies (>70%) were published after 2020, indicating a recent surge in interest. The AI applications primarily focused on diagnosis, treatment response prediction, and prognosis. Machine learning algorithms were predominantly used, particularly random forest, logistic regression, and support vector machines. Omics data were frequently employed as predictors, especially transcriptomics and microbiome analyses. Studies demonstrated good predictive performance across all three areas, with median AUC values ranging from 0.85 to 0.90. Conclusion: AI applications in IBD show promising potential to enhance clinical practice, particularly in disease prognosis and predicting treatment response. However, clinical implementation requires further validation through prospective studies. Future research should focus on standardizing protocols, defining clinically significant outcomes, and evaluating the efficacy of these tools.

Introduction

Inflammatory Bowel Disease (IBD) encompasses two primary subtypes: Crohn’s disease (CD) and ulcerative colitis (UC). The etiology, diagnosis, and management of IBD present significant clinical challenges due to its heterogeneous nature and the complex interplay of genetic factors, environmental triggers, and immunological dysregulation.

Over the past few decades, biomedical and bioinformatic research breakthroughs have enhanced our understanding of IBD complexity. High-throughput genomic sequencing has facilitated the identification of numerous genetic susceptibility loci [1]. Gut microbiome studies have shown the role of microbial dysbiosis in disease pathogenesis [2]. Metabolomic and proteomic analyses have uncovered IBD-specific biochemical signatures [3].

Despite these advancements, several challenges persist in clinical management. Approximately one-third of patients treated with anti-tumor necrosis factor-alpha (TNF-α) agents fail to respond during induction therapy, and among initial responders, about 50% experience a loss of response within a few years [4]. Around 80% of patients with CD will require surgical intervention over their lifetime [5]. Although the risk of post-surgical recurrence remains high [6], effective predictors for this outcome are still lacking [7]. Furthermore, in up to 15% of all IBD cases, a definitive distinction between UC and CD cannot be made during the initial diagnosis [8]. This difficulty in diagnosis occurs more frequently in pediatric populations compared to adults [9]. Diagnostic reclassification occurs in some patients, primarily involving a shift from UC to CD diagnosis [10].

Precision medicine (PM) has emerged as a promising approach in healthcare, aiming to tailor medical interventions to each patient’s characteristics [11]. Given the disease’s heterogeneity, this approach is particularly relevant for IBD. It can identify common factors that define subgroups likely to benefit from specific therapeutic strategies [12].

Artificial Intelligence (AI), especially machine learning (ML) algorithms, provides robust tools for big data analysis and pattern recognition. AI’s capacity to integrate and analyze complex, multidimensional datasets, including genomic, metabolomic, microbial, and clinical information, holds the promise of uncovering intricate patterns [13]. In IBD, AI applications may improve diagnostic accuracy, predict individual patient treatment response, and forecast disease progression. Figure 1 shows the workflow of AI-assisted precision medicine in IBD.

Figure 1.

Workflow of AI-assisted precision medicine in IBD: This figure illustrates the integration of clinical data, omics (microbiome, genomics, metabolomics), and behavioral factors related to inflammatory bowel disease (IBD) into artificial intelligence (AI) processing aimed at achieving precision medicine goals such as treatment response, disease course prediction, diagnosis, and biomarker identification.

The synergy between PM and AI presents a transformative opportunity to improve patient outcome in IBD. This approach addresses the current challenges in IBD management and paves the way for more targeted and effective interventions.

This systematic review (SR) explores the intersection of AI and precision medicine in IBD. By analyzing the current literature, this study investigates the potential applications of AI-driven PM in enhancing IBD management.

Materials and methods

We conducted a SR following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) recommendations [14]. A protocol was registered in PROSPERO (CRD42023373601). We performed a comprehensive search across PubMed, PubMed PMC, BVS, Scopus, Web of Science, Embase, Cochrane, and ProQuest databases, using the following search string: (“Inflammatory Bowel Diseases” OR “Colitis, Ulcerative” OR “Crohn Disease” OR Colitis) AND (“Artificial Intelligence” OR “Machine Learning” OR “Supervised Machine Learning” OR “Unsupervised Machine Learning” OR “Big Data” OR “Deep Learning” OR “Precision Medicine”).

Inclusion criteria

The articles were included in this review based on the following criteria: (1) Human studies on IBD (UC and/or CD) patients; (2) AI application in PM; (3) Comparative, randomized, cross-sectional, retrospective, prospective, or cohort studies were included.

Exclusion criteria

Articles that met any of the following criteria were excluded from the review: (1) reviews, letters, editorials, and conference presentations; (2) studies on non-IBD subjects; (3) studies not distinguishing between CD and UC; (4) incomplete data; (5) AI applied solely to imaging data; and (6) IBD data used only for method validation.

Study selection

Study selection was conducted in two stages. First, two independent reviewers (HDC, TAOS) assessed titles and abstracts based on the inclusion or exclusion criteria, with a third reviewer (LMG) resolving any conflicts. Subsequently, the same two independent reviewers assessed the full text of the previously selected articles, with a third reviewer resolving conflicts. We utilized Rayyan software for organization and data storage [15].

Results

A total of 10,293 studies were initially identified, yielding 5,633 unique records after deduplication. Following title and abstract screening, 453 studies were selected for full-text review. Upon application of the eligibility criteria, 139 studies were included in the SR. Figure 2 illustrates the PRISMA flow diagram detailing the study selection process.

Figure 2.

Study selection flow diagram for the systematic review.

We observed a significant upward trend in publication frequency among the included studies over recent years, with more than 70% of the studies published from 2020 onwards (Figure 3A).

Figure 3.

Distribution of the studies included in the systematic review. A. Number of publications per year up to February 2024 (N=139). No papers were selected in 2014 and 2016. B. Geographic distribution of studied populations, after excluding post hoc studies (N=82).

Regarding the geographical distribution of study populations, excluding post hoc analyses of publicly available databases due to lack of information, we noted a concentration of studies in Asian countries (primarily China and South Korea), North America (the United States and Canada), and European nations. No studies were conducted in African countries; only one was done in Latin America (Figure 3B).

Our analysis of AI application objectives in PM allowed us to classify the studies into three main categories: (i) diagnosis, (ii) prognosis, and (iii) treatment response. Eleven publications [16-26] were concurrently classified into two categories.

Critical factors in the application of AI include the data used, algorithms employed, distribution of the study population (balanced or unbalanced data), training process, and performance metrics.

The median sample size was 237 patients (10 to 95,878). Of note, most studies presented unbalanced data in their analyses. Machine learning algorithms were predominantly used across all three categories, with random forest, logistic regression, support vector machines, artificial neural networks, and boosting algorithms being the most utilized (Figure 4A). Ensemble methods combining multiple algorithms were used in three studies [34,89,105].

Figure 4.

Circular bar plots illustrate AI algorithms and data types used across Diagnosis (D), Prognosis (P), and Treatment (T) in precision medicine applications. A. Random Forest and logistic regression demonstrate predominance across categories. Other methodologies, such as artificial neural networks, boosting techniques, and support vector machines, exhibit variable utilization, highlighting diverse applications of AI algorithms. B. Transcriptomics (gene expression, non-coding RNA, and single nucleotide polymorphisms) and non-omics data (clinical, laboratory, and demographic) predominate, particularly in diagnosis and prognosis/treatment response, respectively. ANN, artificial neural networks; SVM, support vector machines; PLS-DA, Partial least squares-discriminant analysis.

The analyses used various types of data. We observed a clear distinction between diagnosis, prognosis, and treatment response studies (Figure 4B).

Omics data, including transcriptomics (gene expression profiles from various tissues, non-coding RNA (Ribonucleic Acid) analysis, and single nucleotide polymorphisms), microbiome (primarily through bacterial and fungal sequencing), and metabolomics (from blood, feces, and other tissues), were often employed for diagnosis (Figure 4B). In contrast, clinical, laboratory, and demographic (non-omics) data were more frequently employed for predicting treatment response and disease prognosis (Figure 3B). Several studies integrated these non-omics data with omics information [19,20,27-33].

For model development, researchers employed various training and testing techniques (Supplementary Table 1). The most used technique was the random split of the dataset into training and test sets. Cross-validation was also frequently used, either in combination with the train-test split or alone. For post hoc analyses, it was common to combine databases for training while holding out one or more databases for subsequent model validation.

The area under the receiver operating characteristic curve (AUC) was the most commonly used performance evaluation metric, although some analyses reported only accuracy values as their primary evaluation metric. Excellent performance could be seen regardless of the PM category (Table 1). Comprehensive information is detailed in Supplementary Table 1.

Table 1.

Model performance across precision medicine applications

Reference	Precision medicine category	Metrics	Minimum	Median	Maximum
[16-25,34-87]	Diagnosis	AUC (N=45)	0.55	0.90131	1
		Accuracy % (N=17)	73	84.2	100
[16-19,23,24,26-28,122-154]	Prognosis	AUC (N=33)	0.575	0.857	0.9864
		Accuracy % (N=6)	77	92.76	100
[20-22,25,26,29-33,88-121]	Treatment response	AUC (N=38)	0.56	0.854	1
		Accuracy % (N=4)	73	81.05	98.1

N, number of studies; AUC, area under the ROC curve.

Artificial intelligence for diagnosis

Diagnostic evaluation in IBD was the most frequently studied category, with 64 publications. As previously presented, most studies implemented ML algorithms; however, six studies used them to identify diagnostic predictors [82-87].

The majority of studies focused on predicting CD or UC versus control groups. Only 18 [16,17,19,34-36,38,40,45,52,54,61-65,73,81] evaluated CD vs. UC or UC vs. CD, obtaining a median area under the Receiver Operating Characteristic curve (AUC) of 0.8395 and accuracy of 80.92%. Four studies were unable to distinguish between IBD phenotypes [16,17,61,73].

One study [59] developed a multi-class model based on fecal microbiome data to differentiate several diseases. This model achieved an AUC of 0.93 for both UC and CD in multi-class prediction. However, external validation using data from different geographic locations showed lower performance, with AUCs of 0.693 and 0.798 for UC and CD, respectively.

Eleven studies focused on the pediatric population [23,38,39,46,47,55,61,68,69,71,80], most comparing disease states to control groups.

Artificial intelligence for treatment adherence and response

A total of 44 studies implemented AI for predicting treatment response and one for adherence to azathioprine [92]. The predictive analysis used diverse data types, including clinical information, laboratory data, and gene expression analysis, with a predominance of non-omic data (Figure 4B).

A key element with clinical implications is the definition of treatment response outcomes. Researchers evaluated several metrics for this purpose, including clinical and endoscopic condition-specific scores, such as the Crohn’s Disease Activity Index, Pediatric Ulcerative Colitis Activity Index (PUCAI), Mayo score and sub-score, and Simple Endoscopy Score for Crohn’s Disease (SES-CD).

Three studies evaluated treatment responses in the pediatric population: one analyzed the response to corticosteroids (AUC 0.77) [29], another examined exclusive enteral nutrition (AUC 0.90) [95], and a third assessed the response to various treatments, including corticosteroids, exclusive enteral nutrition, and mesalazine (accuracy of 77.8%) [22].

Acute severe ulcerative colitis (ASUC), a life-threatening manifestation of UC characterized by the rapid onset of severe inflammation and often requiring hospitalization, was evaluated in terms of general response to various treatments [30], response to corticosteroids, infliximab, and cyclosporine separately [33], and response to corticosteroids alone [93]. Overall, the prediction for treatment response AUCs ranged from 0.703 to 0.97.

Biologics and small molecules are pivotal in treating inflammatory bowel disease, offering targeted therapeutic options that modulate the immune response. More than half of the studies evaluated the response to these classes: anti-TNF agents -including infliximab, adalimumab, and golimumab [20,21,25,26,31,32,89,90,99-106] demonstrated average AUC values of 0.903 for CD and 0.882 for UC. Studies on the anti-integrin agent vedolizumab [116-118,120,121] yielded average AUCs of 0.75 for CD and 0.708 for UC. Ustekinumab, targeting the interleukins IL-12 and IL-23, exhibited average AUCs of 0.808 for CD and 0.839 for UC. Tofacitinib, a Janus kinase inhibitor [109,110], showed an average AUC of 0.83 for UC.

In one of the studies, Telesco et al. [99] conducted a phase 2a clinical trial with UC patients to validate a predictive model of Golimumab response based on gene expression data from the intestinal mucosa. The model achieved an AUC of 0.688 for endoscopic response at week six and an AUC of 0.671 for response at week thirty, lower than the initial performance. Some hypotheses for the model’s poorer performance were raised, such as differences between training and validation populations.

Artificial intelligence for prognosis in inflammatory bowel disease

Forty-two studies focused on applying AI to the prognostic evaluation of IBDs. While supervised algorithms predominated, the authors employed unsupervised techniques in three articles [141,145,147], focusing on finding molecular markers of disease subtypes.

Assessing disease activity in IBD is essential for making effective treatment decisions, monitoring patient progress, and improving overall clinical outcome. Twenty-two studies [16-19,23,26,122,126-140] predicted disease activity based on various data with overall good performance.

Nine studies examined the clinical course of IBD [24,141,142,145,147,148], with three focusing on the stricturing phenotype in CD [122,143,144]. In one of the studies, Lee et al. [141] identified, from the gene expression of CD8+ T cells, a panel capable of segregating CD and UC carriers into two groups, called IBD1 and IBD2, with the first group being related to worse outcome (measured through the need for treatment escalation, need for immunomodulator). Biasci and Lee et al. [142] used whole blood gene expression to predict IBD1 (IBDhi) and IBD2 (IBDlo) groups in both CD and UC, making it more feasible from a clinical perspective and constructing a prognostic assay (PredictSURE IBD).

While medical management aims to control inflammation, maintain remission, and improve quality of life, surgery becomes necessary in cases of refractory disease, intestinal obstruction, and severe complications, which are common challenges in IBD. Five studies evaluated the need for surgical intervention [27,28,122-124], while post-surgical recurrence was the subject of analysis in four investigations [88,151-153].

Additionally, three studies addressed specific aspects of IBD prognosis: postoperative complications [125], diagnosis of intra-abdominal abscess [146], and sarcopenia [154].

Discussion

This SR revealed a growing interest in applying AI, particularly ML, for PM in IBDs. There has been a significant increase in publications in recent years, reflecting a broader trend observed across disciplines where AI is gaining prominence.

As mentioned earlier, the assessment of the geographical distribution of studies indicates a predominance of research from the United States of America, European countries, and Asian nations. Considering the multifactorial etiology of IBDs, which involves complex interactions between genetic and environmental factors, this geographic concentration limits the generalizability of results to populations from underrepresented regions, especially if stool microbiome or metabolomics are involved [155].

Almost no studies shared how the analysis was done in-depth, like a code or a GitHub project with analysis. This could be an essential step in facilitating validation studies in the future.

Regarding the strategies for model development, many studies employed only train-test splits, which, in theory, requires an extensive sample size to ensure adequate generalization. However, given that several studies presented limited sample sizes, a more restrictive interpretation of the results is advised. Applying techniques such as cross-validation and its variants, in conjunction or not with the training split, can significantly increase the robustness and reliability of the obtained results and should be used.

To enable practical application by clinicians, machine learning algorithms must provide insights into the reasoning behind their decisions. Although simpler models like logistic regression offer greater interpretability, their predictive performance often falls short of more advanced algorithms, such as random forest or gradient boosting algorithms. Currently, techniques that assist us in interpreting complex models, such as Shapley values [156], are available, facilitating the use of more complex models, especially for medical practice.

While the practical application of AI in clinical settings remains challenging, many studies in our review employed advanced omics data (such as gene expression, microbiome sequencing, and metabolomics) for various clinical purposes, likely due to reduced processing costs and the development of analytical techniques [157]. Given that the integration of such data in routine clinical practice is still emerging, we believe these studies may represent the first steps toward translational research in the field.

The identified PM categories in this SR are highly relevant to clinical practice and represent significant gaps in current knowledge in managing IBD.

Accurate and early diagnosis offers the possibility of altering the natural history of the disease, since a delayed diagnosis is associated with complications in CD and the need for surgery in both UC and CD [158]. Among the studies that used AI for IBD diagnosis, the majority evaluated UC or CD versus a control group, which has less clinical relevance, and in only one study were other clinical conditions incorporated into the predictive model [59]. From a practical standpoint, studies that evaluate IBD diagnosis against the primary differential diagnoses have significant clinical relevance and may be the focus of future studies.

Analysis of treatment response primarily evaluated the use of biologics and small molecule classes, which represent the main targeted therapy currently available and, as discussed, are very relevant in the clinical context. The lack of pre-treatment tests for predicting response remains a significant challenge, and the availability of such analysis is crucial in clinical practice. Even without validation, the results presented are promising, and well-conducted validation studies are necessary.

The IBD prognosis has seen a wide range of applications, with particular emphasis on disease activity prediction. Identifying relapse is a crucial aspect of patient follow-up. Currently, this information is obtained either subjectively (through clinical scales) or objectively (by fecal calprotectin, endoscopic, or radiological evaluation) [159-161]. Given the limitations and challenges associated with objective methods [162], developing and validating a model to predict disease activity holds significant practical application, and validation studies should be carried out.

Another relevant application was for predicting disease course. Recently, Noor et al. [163] published data from a randomized clinical trial that evaluated the previously described prognostic assay [142]. Two main outcomes were analyzed: evaluating the marker and comparing step-up versus top-down treatment in newly diagnosed CD patients. In this scenario, the assay could not predict the course of the disease.

Although Noor et al. and Telesco et al. did not validate the disease course and treatment response models, these results provide crucial information for future research, highlighting the importance of robust models built from systematic data collection and well-defined outcomes. In light of these findings, Wyatt et al. [164] published a prospective study protocol to analyze multi-omics data to develop predictive tools for treatment response in IBD.

This study presents limitations that should be considered when interpreting the results. First, the methodological heterogeneity among the included studies prevented the performance of a meta-analysis, limiting the quantitative synthesis. The exclusion of studies focused on image analysis may have omitted relevant information about AI applications in imaging diagnosis in IBDs. However, this topic has already been fully discussed in the literature, and we would like to raise other aspects of AI and PM in IBD that have not yet been covered. Furthermore, the variability in sample sizes, validation methods, and performance metrics across studies makes direct comparisons difficult. Almost all studies are observational, post hoc analysis, or retrospective; therefore, clinical trials are needed to validate the results.

Conclusion

The application of AI in IBD shows significant potential to enhance clinical practice. The results demonstrated promising predictive performance, particularly in disease prognosis and predicting treatment response. However, clinical implementation of these models requires additional validation in other cohorts of patients and a more significant number of participants. Successful integration of AI in IBD management depends on developing standardized protocols, clearly defining clinically significant outcomes, and fostering interdisciplinary collaboration. Future studies should focus on validating and evaluating the actual clinical impact of these tools. The evolution of AI in IBD has the potential to refine our understanding of the pathophysiology and to personalize therapeutic interventions to improve outcomes.

Disclosure of conflict of interest

None.

Abbreviations

AI

Artificial Intelligence

AUC

Area Under the Receiver Operating Characteristic Curve

ASUC

Acute severe ulcerative colitis

BVS

Biblioteca Virtual em Saúde

CD

Crohn’s Disease

IBD

Inflammatory Bowel Disease

ML

Machine Learning

PM

Precision Medicine

PRISMA

Preferred Reporting Items for Systematic Reviews and Meta-Analyses

PUCAI

Pediatric Ulcerative Colitis Activity Index

RNA

Ribonucleic Acid

SES-CD

Simple Endoscopic Score for Crohn’s Disease

SR

Systematic Review

TNF

Tumor Necrosis Factor

UC

Ulcerative Colitis