Editorial: Development and Validation of a Multimodal Machine Learning Model for Diagnosing and Assessing Risk of Crohn's Disease in Patients With Perianal Fistulae: Authors' Reply

Abstract

We thank Drs Parian and Schwartz for their insightful editorial on our study [1]. Patients presenting with perianal fistulae experience a significantly worse long-term prognosis when there is a delay in diagnosing Crohn's disease (CD) [2]. However, clinical decision-making tools for the early identification of patients with perianal fistulae at risk of CD are currently lacking in practice. The present model aims to facilitate the early diagnosis of CD in patients where perianal fistula is the primary manifestation, thereby enabling timely endoscopic evaluation [3].

The incidence of CD in China is increasing rapidly [4], but the training of colorectal specialists has not kept pace. This discrepancy poses challenges for surgeons in identifying perianal fistulising Crohn's disease (PFCD) and in implementing optimal management strategies. The surgical protocols and treatment objectives for PFCD differ substantially from those for cryptoglandular fistula (CGF); treating PFCD in the same way as CGF often has detrimental consequences. Magnetic resonance imaging (MRI) is widely regarded as the gold standard for evaluating perianal fistulae [5]. There is potential for developing a diagnostic prediction model based on MRI that offers a user-friendly tool for clinicians. We can use it to identify patients with PFCD at an early stage and provide appropriate recommendations for subsequent treatment.

We agree with the study's limitation of the absence of endoscopic assessment. As an invasive procedure, endoscopy is performed on all patients with fistula and may be considered excessive, particularly given the higher prevalence of CGF. Although the activity of perianal CD generally mirrors luminal disease activity, perianal symptoms may occasionally manifest independently [6]. The role of endoscopy in validating diagnostic models for CD is of paramount importance. While current models are largely based on MRI features, this is only the first stage of validation. We appreciate the suggestions regarding the expanded dimensions of the model. In the next phase, we plan to optimise the diagnostic process by combining the model results with current ECCO guideline recommendations and synthesising the MRI features with data from the medical history, physical examination, and laboratory investigations to form an integrated diagnostic process framework. We expect this approach to minimise diagnostic bias and enhance the accuracy of identifying subclinical CD. Prospective clinical trials involving multiple medical centres in China, including patients from different regions and covering more clinical features, are underway to further test the robustness and generalisability of the model. Meanwhile, the model will be integrated into an easy-to-use clinical decision support system (i.e., mobile application) to improve its usability and accessibility for clinicians.

Texture analysis based on conventional MRI images is also being used in the early diagnosis of PFCD, showing good sensitivity and specificity [7]. The deep convolutional neural network classifier developed based on MRI shows more robustness in distinguishing between PFCD and CGF compared to radiologists' assessments [8]. Future research endeavours will centre on the advancement of MRI radiomics modelling, the dynamic monitoring of longitudinal patient data and the iterative refinement of the model based on emerging evidence.