Machine learning in predicting postoperative complications in Crohn’s disease

Abstract

Crohn's disease (CD) is a chronic inflammatory bowel disease of unknown origin that can cause significant disability and morbidity with its progression. Due to the unique nature of CD, surgery is often necessary for many patients during their lifetime, and the incidence of postoperative complications is high, which can affect the prognosis of patients. Therefore, it is essential to identify and manage postoperative complications. Machine learning (ML) has become increasingly important in the medical field, and ML-based models can be used to predict postoperative complications of intestinal resection for CD. Recently, a valuable article titled “Predicting short-term major postoperative complications in intestinal resection for Crohn's disease: A machine learning-based study” was published by Wang et al. We appreciate the authors' creative work, and we are willing to share our views and discuss them with the authors.

Key Words: Crohn’s disease; Intestinal resection; Postoperative complications; Machine learning; Explainability

Core Tip: Crohn's disease (CD) is a condition characterized by chronic, recurrent inflammation, which can involve any part of the digestive tract. As the disease progresses, many patients require bowel resection and ostomy, with subsequent postoperative complications. Machine learning (ML) has emerged as a valuable tool for predicting these complications. A recent study published by Wang et al presented an ML approach for predicting major postoperative complications in CD patients undergoing intestinal resection. While acknowledging the merit of the study, we would like to express our opinions and engage in a discussion with the authors.

TO THE EDITOR

The prevalence of Crohn's disease (CD) is increasing, posing a growing health and economic challenge[1]. While the management of CD is primarily aimed at symptom relief and avoidance of surgery, a proportion of patients will inevitably undergo bowel resection. Consequently, the anticipation and prevention of postoperative complications have received increasing attention. Machine learning (ML) has gained prominence in medicine due to its ability to robustly navigate complex non-linear relationships. In an article, Wang et al[2] used ML techniques to predict short-term major postoperative complications in bowel resection for CD. In this article, the authors used logistic regression and random forest for predictive modeling, and random forest has a better performance than logistic regression. In particular, preoperative CD activity index (CDAI), serum albumin levels, and duration of surgery emerged as key predictive factors. While the innovative application and effort of ML are commendable, it is imperative to address some methodological issues that may affect the credibility of the study's conclusions.

First, the significant sample imbalance, with only 5% (13/259) of patients experiencing major surgical complications, could potentially compromise the training effectiveness and generalizability of the model[3,4]. The model may prefer to predict no complications when the area under the the receiver operating characteristic (ROC) curve (AUC) is the only evaluation metric. To address this issue, the use of oversampling, undersampling, or alternative metrics beyond AUC may be warranted[5-7].

Second, the inclusion of preoperative CDAI as a categorical variable (≥ 220 or < 220) rather than a continuous numerical variable may be due to limitations in data accessibility or quality in retrospective studies. The use of a continuous CDAI variable might provide a better response of CD activity[8].

Last, in the evolving landscape of ML, while prioritizing interpretability is not always mandatory[9], there is a growing emphasis on explainability[10]. In this article, the summary plots generated by SHapley Additive exPlanations (SHAP)[11] attempt to elucidate the dynamics of the random forest model. However, a more comprehensive explanation, including detailed SHAP scores for each variable, scatterplots depicting the SHAP scores, and corresponding curve fitting, would improve the interpretability of the study.

In conclusion, the use of ML to predict major postoperative complications after intestinal resection for CD is innovative and requires significant effort. However, we hold different opinions on some methodological issues and would like to discuss them with the authors to enhance the robustness and credibility of future research in this area.