Preliminary Evaluation of a Novel Artificial Intelligence-based Prediction Model for Surgical Site Infection in Colon Cancer
An increasing amount of evidence suggests that postoperative infection is associated with poorer long-term outcomes in various malignancies (1,2). Surgical site infection (SSI) is a common postoperative complication that occurs in 5-40% of patients undergoing colorectal surgery (3,4). Postoperative SSI, an important marker of surgical quality, increases treatment costs, delays the initiation of adjuvant therapy, affects quality of life, and may be associated with premature mortality (2).
Preoperative weakening of immunity and undernutrition are listed as risk factors for SSI in the US Centers for Disease Control and Prevention (CDC) guidelines, and various systemic inflammatory and nutritional scores have been reported as useful SSI predictors (5,6). The roles of the systemic inflammatory response and nutritional status in SSI have been increasingly recognized, as reflected in the numerous immunological and nutritional markers shown to affect SSI incidence in different types of cancer, including colon cancer. These markers include the modified Glasgow Prognostic score (mGPS), C-reactive protein to albumin ratio (CAR, which is an index of cancer cachexia), Controlling Nutritional Status score (CONUT score, which detects potential malnutrition), prognostic nutritional index (PNI, which is calculated using the serum albumin levels and peripheral lymphocyte count), and lymphocyte to monocyte ratio (LMR, which reflects the balance between the tumor-promoting environment and antitumor immunity) (7-9).
Many studies have attempted to predict SSI following colon surgery by using prediction models, which include various clinicopathological factors. Such models rely on conventional statistical analyses, such as multivariate analyses and nomograms. The predictive ability of these models was high, as assessed by a receiver-operating characteristic (ROC) curve analysis or the concordance index (C-index), and their area under the curve (AUC) values were all approximately 0.75 (10-14). This led us to consider the applicability of artificial intelligence (AI)-based prediction models, which have developed rapidly in recent years.
The third AI boom is arriving, and AI is evolving rapidly. AI, particularly machine learning and deep learning, has been applied in clinical cancer research, and the cancer prediction performance has reached new heights. However, no studies have been published on building AI-based prediction models for colon cancer using clinicopathological factors.
Therefore, in this study, we aimed to build a novel AI-based prediction model (AI model) for SSI in stage II-III colon cancer using immunological and nutritional markers and to perform a preliminary evaluation of its performance.
Patients and Methods
This study involved a preliminary evaluation of an AI-based SSI prediction model that included various clinicopathological factors observed in patients with colon cancer. In previous models based on conventional statistical analysis, all the reported values for the AUC and C-index were approximately 0.75 (10,11,13,14). The accuracy of our AI model, with an AUC value of 0.731, was as good as that of the previously reported models. Moreover, the AI model uses easily obtainable data on simple clinicopathological factors, and the cost of constructing the model is low.
Postoperative infections contribute to poor survival (1,2). SSI is one of the most important complications of colorectal surgery (3,4). Therefore, the identification of SSI predictors, which can be used in clinical practice, is a pressing issue. A cancerous state often activates a systemic inflammatory reaction, and inflammation in turn reduces immunity (7). In addition, impaired nutrient intake is occasionally seen during the perioperative period in patients with colorectal cancer, which leads to an immunocompromised state (8). Therefore, nutritional management is indispensable for the success of colorectal surgery. Although various systemic inflammatory and nutritional scores have been used to predict postoperative complications (9), no useful SSI predictors have been reported for colon cancer. In this study, LMR and CAR were ranked high in terms of IOV using Prediction One. This finding suggests that inflammatory and nutritional scores are significantly associated with SSI.
The use of machine learning has become widespread in clinical research. The types of learning used by computers are subclassified into categories, such as supervised and unsupervised learning. Supervised learning begins with the prediction of a known output or target. Therefore, it is often used for the estimation of risk in medical research (15,16). In deep learning, unsupervised learning is initially used to identify robust features, and subsequently, these features are refined and can ultimately be used as predictors in the final supervised model. Deep neural networks (DNNs), also known as deep learning networks, are used in many AI applications (17). Multiple or multivariate logistic regression fits multiple parameters in prediction models by assuming that predictors are linearly and additively related to an outcome. However, nonlinear problems commonly occur when these models are applied in the field of human physiology because of complex interactions. Therefore, linear models might not be capable of adequately predicting outcomes, which may explain the differences observed in predictive accuracy between the multivariate and AI models.
In general, conventional statistical analysis involving a conventional linear model focuses on explaining data and is said to be inferior to AI in terms of its predictive ability. Although AI is useful for prediction, there are important concerns owing to the opaque, black-box nature of most AI algorithms. Building prediction models with explainable AI mechanisms is a powerful and transparent alternative to black-box AI models (18). In the present study, Prediction One calculated not only the accuracy of the model but also the contribution of each factor to the outcome, thereby enabling a better understanding of the model and its constituents. The IOV values and independent risk factors identified through multivariate analysis do not exactly correspond, but IOV is an informative variable for describing the model.
This study has several limitations worth noting. First, this was a retrospective, single-center study. Second, we did not determine whether the patients had hematological or autoimmune disorders, which may have influenced the preoperative laboratory data. Third, our data did not distinguish between superficial and deep-incision infections. Given that this is the first study of its kind that used Prediction One, SSI was defined broadly in order to improve prediction accuracy. Fourth, external validation was not performed. Prospective studies and external validation are needed to improve the performance of the AI model.
In conclusion, based on our preliminary analysis, the AI model was found to be useful for predicting SSI in patients with stage II-III colon cancer. In this study, SSI was found to be associated with worse oncological outcomes; therefore, the prediction of SSI occurrence is important in colon cancer.
Conflicts of Interest
The Authors have no conflicts of interest to disclose in relation to this study.
Yuki Ohno and Junichi Mazaki conceived the idea of the study. Yuki Ohno developed the statistical analysis plan and conducted statistical analyses. Ryutaro Udo, Tomoya Tago, Kenta Kasahara, Masanobu Enomoto and Tetsuo Ishizaki contributed to the interpretation of the results. Yuki Ohno drafted the original manuscript. Yuichi Nagakawa supervised the conduct of this study. All Authors reviewed the manuscript draft and revised it critically on intellectual content. All Authors approved the final version of the manuscript to be published.