Machine learning as a tool predicting short-term postoperative complications in Crohn’s disease patients undergoing intestinal resection: What frontiers?

Abstract

The recent study, “Predicting short-term major postoperative complications in intestinal resection for Crohn’s disease: A machine learning-based study” investigated the predictive efficacy of a machine learning model for major postoperative complications within 30 days of surgery in Crohn’s disease (CD) patients. Employing a random forest analysis and Shapley Additive Explanations, the study prioritizes factors such as preoperative nutritional status, operative time, and CD activity index. Despite the retrospective design’s limitations, the model’s robustness, with area under the curve values surpassing 0.8, highlights its clinical potential. The findings align with literature supporting preoperative nutritional therapy in inflammatory bowel diseases, emphasizing the importance of comprehensive assessment and optimization. While a significant advancement, further research is crucial for refining preoperative strategies in CD patients.

Key Words: Machine learning; Crohn’s disease; Intestinal resection; Postoperative complications; Preoperative assessment; Nutritional optimization; Predictive model; Gastrointestinal surgery; Surgery

Core Tip: In this editorial on the abovementioned study, a machine learning model predicts major postoperative complications within 30 days for Crohn’s disease (CD) patients undergoing intestinal resection. Prioritizing factors include preoperative nutritional status, operative time, and CD activity index. The model’s robustness, with area under the curve values exceeding 0.8, emphasizes the clinical significance of comprehensive preoperative assessment and nutritional optimization in CD. These findings, discussed in the editorial context, align with existing literature and endorse European Society for Clinical Nutrition and Metabolism guidelines. Further research is warranted to refine preoperative strategies for this patient population.

INTRODUCTION

We have read with interest the recent study titled “Predicting short-term major postoperative complications in intestinal resection for Crohn’s disease: A machine learning-based study” published in the World Journal of Gastrointestinal Surgery[1]. This intriguing retrospective study assessed the predictive efficiency of a machine learning model in patients with Crohn’s disease (CD) undergoing intestinal resection in relation to the development of major postoperative complications within a short time (i.e., within 30 days of surgery). The definition of these complications correctly followed the Clavien-Dindo classification[2], identifying the study’s target complications as equal to or greater than grade III. Another variable considered in this model, as a secondary endpoint of the study, was the duration of hospital stay.

Inflammatory bowel diseases (IBDs) are chronic conditions with a typically relapsing-remitting course. Their natural history is far from predictable, especially in CD, where optimal treatment of one disease location does not preclude reactivation in the same site (or in a perianastomotic location if the site has undergone surgical resection) or in additional locations[3,4]. CD, by definition, can, in fact, affect any section of the gastrointestinal tract from mouth to anus[5]. This has always been an apparent differentiating factor between CD and its twin IBD, ulcerative colitis, where definitive surgical techniques like total restorative proctocolectomy with ileal pouch construction are available when indicated, albeit with potentially long-term complications such as chronic pouchitis[6]. Therefore, this work is of extreme relevance as it has predictive variables for significant outcomes (as in this case, related to post-surgical outcomes in CD) allowing better preoperative risk stratification and the adoption of potential additional measures during perioperative and postoperative phases in high-risk subjects.

INTEGRATING ARTIFICIAL INTELLIGENCE INTO THE PERIOPERATIVE MANAGEMENT OF CD

Despite the significant limitation of a retrospective design, this study has an interesting strength in its analytical methodology. The authors employed a machine learning-based random forest analysis[7] rather than traditional regression models to evaluate variables with predictive power towards their predefined endpoints. Each predictive parameter underwent evaluation using Shapley Additive Explanations to assess its contribution to the predictive model quantitatively. Furthermore, the enrolled patients (over two hundred) were randomized into training and validation cohorts, adding value to the study. Additionally, the area under the curve (AUC) of this nomogram predictive model was remarkably higher than 0.8 (precisely, 0.916 in the training cohort and 0.96 in the validation cohort). These AUC values, subjected to the random forest model, maintained and surpassed their performance (0.965 and 0.924, respectively, for the two cohorts).

From the authors’ results, in the initial phase of uni- and multivariate traditional analysis, it emerged predictably that the group of patients who retrospectively manifested a higher rate of major postoperative complications presented with an unfavourable clinical-biochemical profile preoperatively (i.e., with reduced albumin levels and a higher CD activity index greater than or equal to 220). This was accompanied by a higher rate of conversion to laparotomic surgery, indicating a predictably higher likelihood of postoperative complications and longer hospital stay.

The authors’ analysis facilitated the stratification of feature importance among the weighed variables, culminating in a prioritized list of the foremost five factors. The leading contributors to this selection were the preparatory CD activity index, operative time, serum albumin, body mass index, and intraoperative blood loss. Furthermore, following the Shapley Additive Explanations analysis, serum albumin, operation time, and preoperative CD activity index were identified as paramount in stabilizing the predictive model.

This investigation has adeptly identified variables inherently linked to an adverse post-surgical outcome, notably incorporating them into a sturdy predictive model. Undoubtedly, the findings underscore the imperative role of preoperative nutritional optimization. It is widely acknowledged that optimizing nutritional status before surgery in CD diminishes the risk of septic complications within the intra-abdominal milieu, consequently mitigating the necessity for temporary diverting stomas[8].

A systematic review conducted in 2017, encompassing literature spanning the past two decades, unequivocally affirms the advantages associated with preoperative nutritional therapy[9]. These advantages extend beyond mitigating infectious complications, demonstrating superior efficacy even when employing enteral formulations compared to the more intrusive total parenteral nutrition[9]. Furthermore, it is imperative to underscore that malnutrition serves as an inherent risk factor for postoperative complications in abdominal surgery[10].

Hence, the rationale behind the European Society for Clinical Nutrition and Metabolism (i.e., European Society of Parenteral and Enteral Nutrition) guidelines on IBD was elucidated, with this study providing further substantiation. These guidelines strongly advocate for a comprehensive nutritional assessment before planned surgery in IBD patients and an appropriately tailored dietary and nutritional intervention for individuals at nutritional risk or with established malnutrition[11]. Moreover, the guidelines recommend deferring surgery for up to fourteen days whenever feasible in cases of preoperative malnutrition[11]. Nonetheless, this study reaffirms what has been previously reported[12-14], that scheduling elective surgery in CD (when feasible) under well-controlled disease activity is preferable to bringing a patient to the operating table with active disease. Similar considerations can be extended to the operative duration, highlighting its correlation with the risk of postoperative complications[15].

Conversely, upon scrutinizing the variables included for importance ranking in the random forest model, it was surprising to note that the age at the time of CD diagnosis indirectly reflected the patient’s age at the time of surgery for those diagnosed after the age of fifty, and, in general, the age at the time of surgery, held remarkably diminished significance in the model. Indeed, a previous study revealed that elderly CD patients undergoing their first surgery display a more indolent disease course, while younger patients, particularly those in the ≤ 19 years category, tend to experience a potentially more aggressive trajectory[16]. This underscores the imperative for tailored surgical strategies and vigilant long-term follow-up, especially for those undergoing their initial abdominal surgery at a younger age.

Conversely, it is crucial to contextualize this finding with the understanding that the authors primarily focused on short-term complications, even though elderly patients generally face a higher risk of postoperative complications following abdominal surgery[17,18]. Therefore, such a trend might have been accentuated in the retrospective analysis had long-term complications been more comprehensively assessed. In the analysis conducted by the authors, another aspect that struck us was the absence of preoperative therapy as a conditioning factor for major postoperative complications, especially in the case of steroids.

The preoperative safety of medical therapies in CD has been the subject of a significant debate, resulting in increased confidence regarding the role of steroids[19] in postoperative infectious complications and less certainty concerning biologics. Despite guidelines leaning towards a favourable safety profile for biologics in the preoperative phase, the discourse has left some uncertainties[4]. Our team recently published data concerning small molecules, specifically tofacitinib, in ulcerative colitis[20]. This agent has demonstrated a safety profile in the preoperative phase comparable to that of biologics. In any case, the intrinsic limitation of the retrospective design necessitates reproducing this model in a prospective study setting. This approach would mitigate the biases related to the study design and allow the assessment of risk factors using prospective risk measures. Furthermore, a prospective study design would enhance the potential generalizability of the results, improving both internal and external validity and providing a clearer understanding of the impact magnitude of certain preoperative variables, such as steroid therapy. In conclusion, another aspect in the authors’ data that has caught our attention is how the creation of a stoma has not played a role in complications despite such behaviour described in previous experiences[21,22].

CONCLUSION

In conclusion, utilising a machine learning-based approach, the recent study on predicting short-term major postoperative complications following intestinal resection for CD contributes significantly to gastrointestinal surgery. Despite the inherent limitations of a retrospective design, the study’s strength lies in its analytical methodology, employing a random forest analysis and Shapley Additive Explanations to evaluate predictive variables. The robustness of the predictive model, with AUC values exceeding 0.8, underscores its potential clinical utility. The prioritized list of contributing factors, including preoperative nutritional status, operative time, and CD activity index, highlights the importance of comprehensive preoperative assessment and nutritional optimization. The study aligns with existing literature advocating for preoperative nutritional therapy in IBD, as endorsed by European Society of Parenteral and Enteral Nutrition guidelines. While the current study represents a significant step, pursuing further research remains paramount to continually improve the precision and clinical utility of preoperative optimization in CD patients.