Development and Validation of a Multimodal Machine Learning Model for Diagnosing and Assessing Risk of Crohn's Disease in Patients With Perianal Fistula.

Abstract

Listen

Translate article

Diagnosis of Crohn's disease (CD) can pose challenges, particularly when perianal fistula is the initial presentation. To develop and validate a predictive model, establishing a visual web tool for early diagnosis of CD in patients presenting with perianal fistula. This retrospective, multicentre validation study involved patients diagnosed with either perianal fistulising CD or cryptoglandular fistula who underwent initial perianal fistula surgery subsequent to rectal MRI at three Chinese centres from September 2016 to December 2020. A random forest classification model was trained on the derivation cohort (n = 550), randomly split into training and test sets at a 7:3 ratio. Validation utilised data from two external centres (n = 300). Model interpretation employed the Shapley Addictive explanation (SHAP) framework. The validated model was integrated into a web tool for calculating patient-specific risk. In the derivation cohort, SHAP analysis highlighted rectal wall ulceration, rectal wall thickening, submucosal fistula, and T2 hyperintensity as risk factors, while age was identified as protective. A random forest classification model developed using these top 5 features achieved an AUROC of 0.9425 (95% CI: 0.8943-0.9906). In the validation cohort, the model performed well with AUROC values of 0.9187 (95% CI: 0.8620-0.9754) and 0.9341 (95% CI: 0.8876-0.9806), respectively. We developed a publicly accessible web-based application. We have developed a multimodal machine learning model and a web tool that can predict and present CD risk in patients initially presenting with perianal fistula.