Artificial intelligence and inflammatory bowel disease: Where are we going?

Table 1 Main studies of artificial intelligence application in the various fields of inflammatory bowel diseases

Ref.	Field	Study features	Main finding
Isakov et al[26], 2017	IBD genetics	ML model to assess 16390 genes in IBD and healthy patients	Identified 347 IBD-risk genes (67 newly identified)
Cheng et al[27], 2019	IBD genetics	Software analysis to assess the genetics of 32713 IBD patients	Identified several genes potentially involved in UC; identification of 11 common Gene Ontology terms for UC
Yuan et al[28], 2017	IBD genetics	Software analysis to assess 12754 genes in IBD and healthy patients	Identified 41 genes closely associated with IBD
Mihajlović et al[30], 2021	IBD and microbiota	ML classification algorithm to identify IBD from 1638 fecal samples	Confirmed strong connection between IBD and specific fecal microbial species
Manandhar et al[31], 2021	IBD and microbiota	ML model analysis of fecal microbiota from 729 IBD patients and 700 healthy controls	Identified of 117 bacterial taxa with a potential role in diagnostic screening of IBD
Mossotto et al[32], 2017	IBD diagnosis	ML model to assess 287 pediatric patients with IBD	Accuracy of 83.3% of the combined endoscopy-histology ML model in the classification of pediatric IBD patients
Quénéhervé et al[33], 2019	IBD diagnosis	AI analysis of CLE images from 50 IBD patients and 9 healthy controls	AI analysis had 100% sensitivity and specificity for IBD diagnosis, 92% sensitivity and 91% specificity of IBD differential diagnosis
Ananthakrishnan et al[9], 2013	IBD diagnosis and data collection	NLP model trained and validated on 700 UC patients and 700 CD patients to improve case definition and identification from EMRs	NLP model provided better accuracy (AUC 0.94-0.95) than models using only the International Classification of Diseases 9th revision for IBD case definition and identification
Stidham et al[39], 2019	IBD endoscopy	DL model for UC severity trained on 16514 endoscopic images	Similar performance of the DL model and experienced human reviewers in grading UC endoscopic severity
Ozawa et al[40], 2019	IBD endoscopy	CNN-based CADe system for UC severity trained on 26304 endoscopic images	CADe system had AUCs of 0.86 and 0.98 in the identification of Mayo score 0 and 0-1, respectively
Maeda et al[41], 2021	IBD endoscopy	Endoscopic AI model used in real time on 135 UC patients in clinical remission	Endoscopic applications of real time AI predicted clinical relapse of UC with statistical significance
Gottlieb et al[42], 2021	IBD endoscopy	DL algorithm to assess UC severity on 795 full-length endoscopy videos	DL algorithm showed significant inter-rater agreement to human central readers for prediction of UC severity
Yao et al[43], 2021	IBD endoscopy	Endoscopic AI model (CNN) to assess UC grading used on 169 endoscopy videos and compared to dual central reader review	AI model approximated the scoring of experienced reviewers for grading of UC endoscopic activity
Byrne et al[44], 2021	IBD endoscopy	DL model (CNN) to detect and assess UC activity leveraged on > 375000 frames	DL model resulted in well aligned scoring guidelines and experts’ performances
Takenaka et al[45], 2020	IBD endoscopy	DL algorithm trained on endoscopic images and biopsy results and tested on 875 UC patients	DL model identified with an accuracy > 90% patients in endoscopic and histologic remission
Maeda et al[46], 2019	IBD endoscopy (endocytoscopy)	CADe system to predict persistent histologic phlogosis from endocytoscopy validated on 100 UC patient	CADe system provided a diagnostic accuracy of 91% with perfect reproducibility for identification of persistent histologic inflammation
Bossuyt et al[47], 2020	IBD endoscopy	AI algorithm based on pixel color data and pattern recognition from endoscopic images tested on 55 patients	AI algorithm (“red density”) provided an objective computer-based assessment of UC disease activity with good correlation with endoscopic and histological scoring systems
Aoki et al[51], 2019	IBD endoscopy (VCE)	AI system (CNN) tested on 10440 small bowel images for detection of erosions and ulcers in CD	AI system showed an accuracy of 90.8% for detection of erosions and ulcers
Klang et al[52], 2020	IBD endoscopy (VCE)	DL algorithm applied on 17640 VCE images for ulcer detection in CD	DL algorithm provided an accuracy ranging from 95.4% to 96.7% with an AUC of 0.99 for ulcer detection
Klang et al[53], 2021	IBD endoscopy (VCE)	DL model applied on 27892 VCE images for identification of intestinal strictures in CD	DL model showed an accuracy of 93.5% in stricture identification and excellent differentiation between strictures and other lesions
Ferreira et al[54], 2022	IBD endoscopy (VCE)	DL model trained and validated on 8085 VCE images for detection of erosions and ulcers in CD	DL model provided an accuracy of 92.4% and a precision of 97.1% for lesion detection
Aoki et al[55], 2020	VCE	Comparison between standard endoscopist reading and reading after AI model screening of 20 full-length VCE videos	The mean VCE video reading time was significantly shorter after AI model (CNN) screening compared to standard reading
Maeda et al[56], 2021	IBD endoscopy (surveillance)	Case report of dysplasia detection by AI system in a patient with long standing UC	AI system (EndoBRAIN) identified 2 colonic lesions that harbored low-grade dysplasia upon histological examination
Fukunaga et al[57], 2021	IBD endoscopy (surveillance)	Case report of dysplasia detection by AI system in a patient with long standing UC	AI system (EndoBRAIN) identified rectal lesions that harbored high-grade dysplasia upon histological examination
Reddy et al[72], 2019	IBD prognosis	ML model employed on 82 CD patients’ EMRs to predict disease course	ML model predicted inflammation severity with high accuracy (AUC 92.8%) from EMR data
Takenaka et al[73], 2022	IBD prognosis	ML model validated on endoscopic images and biopsy results from 875 UC patients to predict disease course	Histologic remission detected by the ML model correlated with a significant reduction in clinical relapse, steroid use, hospitalization, and colectomy
Li et al[75], 2021	Response to treatment	ML model employed on 174 CD patients to predict response to infliximab	ML model based on clinical and serological parameters showed an accuracy of 0.85 for prediction of response to infliximab
Waljee et al[76], 2018	Response to treatment	AI model employed on 472 CD patients to predict response to vedolizumab	AI model based on clinical and serological parameters was able to identify patients that achieved a corticosteroid-free biologic remission at week 52 of vedolizumab
Waljee et al[77], 2018	Response to treatment	ML algorithm employed on 491 UC patients to predict response to vedolizumab	ML algorithm based on clinical and serological parameters was able to identify patients that achieved a corticosteroid-free biologic remission at week 52 of vedolizumab
Doherty et al[78], 2018	Response to treatment	AI model to assess response to treatment with ustekinumab in 306 patients with CD	AI model detected patients in remission based on clinical data and fecal microbiota at week 6 and 22 of ustekinumab

AI: Artificial intelligence; AUC: Area under the curve; CADe: Computer-aided detection; CD: Crohn’s disease; CLE: Confocal laser endomicroscopy; CNN: Convolution neural networks; DL: Deep learning; UC: Ulcerative colitis; EMR: Electronic medical record; IBD: Inflammatory bowel disease; ML: Machine learning; NLP: Natural language processing; VCE: Video capsule endoscopy.

Table 2 Artificial intelligence application for histological assessment of ulcerative colitis

Ref.	Study design	Population	Outcome	Results
Vande Casteele et al[64], 2022	Cohort study	Colonic biopsies from 88 UC patients with histologically active disease	To assess a DL machine in quantifying eosinophils in colonic biopsies and validate against a pathologist’s count	The AI system highly agreed with manual eosinophil count by pathologists (ICC 0.81-0.92)
Peyrin-Biroulet et al[67], 2022	Cohort study	200 histological images of UC biopsies	To evaluate an AI algorithm in assessing histological disease activity according to the Nancy index	The CNN model had an excellent agreement with pathologists in the assessment of the Nancy index (ICC 0.84)
Villanacci et al[66], 2022	Cohort study	614 biopsies from 307 UC patients	To test a CNN-based CADe system for evaluating HR based on PHRI, Robarts, and Nancy indexes	The CADe system accurately assessed HR (sensitivity 89%, specificity 85% for PHRI) and similar performance for Nancy and Robarts

AI: Artificial intelligence; CADe: Computer-aided detection; CNN: Convolution neural networks; DL: Deep learning; HR: Histological remission; ICC: Interclass correlation coefficient; PHRI: PICaSSO histologic remission index; UC: Ulcerative colitis.