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

Abstract

Perianal fistulising Crohn's disease (PFCD) is found in up to 25% of patients with CD and is associated with decreased quality of life and increased healthcare expenditures. Approximately 11.5% of patients with CD present with perianal fistula (PAF) as their first symptom [1]. Furthermore, 4%–5% of patients have isolated PAF as their only manifestation [2]. Prolonged time to diagnosis of CD in patients with PAF may result in development of more complex fistula and progression of luminal disease. In one study, the median time to diagnosis of CD after PAF was 15 months with a range up to 4 years [3]. Earlier diagnosis and treatment is associated with higher rates of fistula closure and prevention of disease progression [3, 4]. However, differentiating PFCD from cryptoglandular fistula (CGF) can be challenging with two diverging treatment algorithms.

Xiang et al. [5] developed a web-based tool using the top five features in a machine learning model to predict the risk of CD in patients with PAF. Rectal wall ulceration, rectal wall thickening, submucosal fistula, T2 hyperintensity and age < 30 were independently associated with increased risk of CD and together achieved an AUROC of 0.94 (95% CI: 0.89–0.99). This is an excellent step towards risk-stratifying patients with PAF into those who require colonoscopic assessment for CD and those who likely have CGF unrelated to CD.

A major limitation of this study is the lack of endoscopic assessment in patients classified as having CGF to rule out subclinical CD. PAF can be challenging to treat, with a need for clinical decision tools and prediction models to determine which patients are more likely to have CD versus CGF. Xiang et al.'s model strongly focuses on rectal inflammation with two of the five features, including MRI findings seen in proctitis. Current MRI-based PFCD scoring systems all incorporate the items assessed in the model of Xiang et al. Important pieces of the history (chronic diarrhoea, abdominal pain, weight loss, rectal bleeding and family history of IBD), physical examination (irregular or hypertrophied anal skin tag) and laboratory testing (faecal calprotectin/lactoferrin), which are known predictors of CD could further improve the prediction model. Faecal calprotectin accurately distinguishes PFCD from CGF [6]. Furthermore, there is a strong correlation between fistula scraping calprotectin values and disease outcomes defined by the TOpCLASS classification system, suggesting a relationship with fistula prognosis [6]. Current ECCO guidelines recommend ileocolonoscopy in patients with an unexplained fistula and suspicion of CD [7]. In patients with negative ileocolonoscopy, capsule endoscopy can provide additional diagnostic yield [7].

Although an excellent start, further validation testing is needed in patients who have completed assessment to formally rule out CD before assigning a CGF diagnosis. A successful prediction model should incorporate patient history, examination and laboratory results in addition to MRI findings. Our goal must be to diagnose CD early in the fistula course to decrease complications and change the disease trajectory.