Introduction

Inflammatory bowel diseases (IBD) carry a high risk of complications if not diagnosed promptly and are often misdiagnosed as irritable bowel syndrome (IBS). Delayed IBD detection leads to an increase in complicated cases and missed opportunities for timely therapy initiation.

Aims & Methods

The study aimed to develop an AI-based clinical decision support system (CDSS) for IBD diagnosis.
We created three databases, containing clinical-laboratory data, endoscopic and histopathological images. Patients enrolled from 2014 to 2024 included those with confirmed IBD (ulcerative colitis (UC) and Crohn’s disease (CD) with colonic involvement) during flare-ups and IBS with normal endoscopic and histopathological findings.
Separate artificial neural networks were designed for IBD differentiation: multilayer perceptrons (MLPs) for clinical-laboratory data analysis and a pre-trained VGG16 convolutional neural network (CNN) for digital image analysis (initialized with ImageNet weights).
This work has been supported by the grants the Russian Science Foundation, RSF 24-25-20127.

Results

A retrospective database of 1,406 patients with clinical and laboratory data was compiled. For IBD vs. IBS differentiation, the cohort was reduced to 390 patients that had all the fulfilled data, while UC vs. CD analysis included 964 patients.
An MLP (13:13-5-1:1) achieved 92.7% and 99.0% accuracy in IBD diagnosis, incorporating stool frequency/consistency, absence of extraintestinal/perianal manifestations, functional disorders, family history of autoimmune diseases, prior surgeries, CRP, and fecal calprotectin (p≤0.019).
Another MLP (9:9-8-1:1) detected 70.16% of CD and 86.40% of UC cases, using data on constipation, diarrhea, bloody stools, absence of extraintestinal manifestations, fever, perianal lesions, strictures, and prior bowel surgeries (p<0.001).
The endoscopic database comprised 5,693 JPEG images: 3,400 of normal mucosa, 1,357 - CD, 935 - UC. Model sensitivity and specificity were for CD - 91%/79%; UC - 90%/97%; normal - 90%/97%.
The histopathological dataset included 398 whole-slide images (1280×960 px), segmented into 224×224 px tiles. The final set included 8,620 images: 2,360 - CD, 2,320 – normal mucosa, 3,950 - UC. Sensitivity and specificity were for CD - 91% and 95%, UC - 96%, 95%, normal - 96%, 93%, respectively.
The integrated clinical DSS was constructed using Bayesian belief networks implemented in Hugin Lite 8.3 software and integrated clinical, endoscopic, and histopathological nodes. The network architecture replicated the stepwise diagnostic process and incorporated the following nodes: clinical and laboratory data, endoscopy (IBD/IBS differentiation), histopathology (IBD/IBS differentiation), IBS, IBD, endoscopy (UC/CD diagnosis), histopathology (UC/CD discrimination), final diagnosis. The system was validated using six patients per diagnostic category (IBS, UC, CD), with complete sets of clinical, endoscopic, and histopathological data required for all component models. The integrated model demonstrated an overall diagnostic accuracy of 83.3%.

Conclusion

The developed CDSS shows promise for earlier IBD detection but requires further validation studies.