Machine learning model predicting factors for incisional infection following right hemicolectomy for colon cancer

Abstract

Background and aim

Colorectal cancer is a prevalent malignancy worldwide, and right hemicolectomy is a common surgical procedure for its treatment. However, postoperative incisional infections remain a significant complication, leading to prolonged hospital stays, increased healthcare costs, and patient discomfort. Therefore, this study aims to utilize machine learning models, including random forest, support vector machine, deep learning models, and traditional logistic regression, to predict factors associated with incisional infection following right hemicolectomy for colon cancer.

Methods

Clinical data were collected from 322 patients undergoing right hemicolectomy for colon cancer, including demographic information, preoperative chemotherapy status, body mass index (BMI), operative time, and other relevant variables. These data are divided into training and testing sets in a ratio of 7:3. Machine learning models, including random forest, support vector machine, and deep learning, were trained using the training set and evaluated using the testing set.

Results

The deep learning model exhibited the highest performance in predicting incisional infection, followed by random forest and logistic regression models. Specifically, the deep learning model demonstrated higher area under the receiver operating characteristic curve (ROC-AUC) and F1 score compared to other models. These findings suggest the efficacy of machine learning models in predicting risk factors for incisional infection following right hemicolectomy for colon cancer.

Conclusions

Machine learning models, particularly deep learning models, offer a promising approach for predicting the risk of incisional infection following right hemicolectomy for colon cancer. These models can provide valuable decision support for clinicians, facilitating personalized treatment strategies and improving patient outcomes.

Introduction

Colorectal cancer is one of the most prevalent malignancies globally, and right hemicolectomy is a frequently performed surgical intervention [1, 2]. Despite advancements in surgical techniques and perioperative care, surgical site infections (SSI) remain significant complications following right hemicolectomy. These infections not only prolong hospital stays but also escalate healthcare expenditures and compromise patient well-being [3, 4]. Therefore, there is an urgent need to identify predictive factors associated with incisional infections to mitigate their occurrence and improve postoperative outcomes.

Traditionally, risk assessment for SSI following right hemicolectomy has relied on clinical judgment and established risk factors. Nakamura et al [5]. assessed risk factors for incisional infections in 1144 patients undergoing laparoscopic colon cancer surgery using multifactorial logistic regression, identifying preoperative serum albumin levels, anastomotic techniques, and suture types as risk factors. However, the researchers did not evaluate the efficacy of the logistics. Moreover, previous studies assessing incisional risk have utilized conventional regression, yielding disparate risk factors [6, 7]. Therefore, the complexity and multifactorial nature of SSI necessitate more sophisticated methods for accurate prediction. In recent years, machine learning (ML) techniques have emerged as powerful tools for analyzing complex datasets and identifying patterns that may not be readily apparent through conventional methods.

This study aims to leverage the capabilities of various ML models, including random forests, support vector machines (SVM), deep learning, as well as traditional logistic regression, to predict factors contributing to incisional infections following right hemicolectomy for colon cancer [8, 9]. By harnessing various ML algorithms, we seek to develop a robust predictive model capable of discerning subtle relationships between patient demographics, preoperative variables, surgical factors, and postoperative incisional infections. Previous research by Chen et al [10]. found that predictive models using machine learning can improve the prediction of surgical site infections following colorectal surgery, achieving higher accuracy than logistic regression.

The findings of this study are expected to make a significant contribution to the field of colorectal cancer surgery by providing clinicians with reliable tools for preoperative risk assessment and personalized decision-making. By identifying high-risk patients susceptible to incisional infections, clinicians can implement targeted preventive strategies and optimize perioperative care protocols, ultimately improving patient prognosis and alleviating healthcare burdens.

Materials and methods

Study design and data collection

This retrospective study included data from 322 colon cancer patients who underwent right hemicolectomy between January 2018 and January 2021 at Shengjing Hospital of China Medical University. The data collection spanned three years, from January 2018 to January 2021, to ensure the comprehensiveness of the dataset and to minimize potential biases related to changes in clinical practice or treatment protocols over time. Patient demographics, preoperative chemotherapy status, body mass index (BMI), surgical duration, and other relevant clinical variables were collected from electronic medical records. Inclusion criteria comprised (1) patients with postoperative pathological diagnosis of colon cancer and (2) all underwent right hemicolectomy, while exclusion criteria included (1) previously diagnosed colon cancer, (2) prior treatment history, (3) incomplete medical records, and (4) patients undergoing emergency surgery for other conditions. This study was conducted in accordance with the principles outlined in the Declaration of Helsinki and approved by the Institutional Review Board of the Ethics Committee of Shengjing Hospital of China Medical University. Patient confidentiality and data privacy were strictly maintained throughout the study process.

In this study, some patients with right-sided colon cancer received chemotherapy or radiotherapy before surgery [11,12,13]. The criteria for selecting preoperative treatment primarily included the tumor stage and size, the overall health status of the patient, and the comprehensive evaluation by a multidisciplinary team. Specifically, for some advanced or larger tumors, neoadjuvant therapy before surgery helps to shrink the tumor, reduce the difficulty of surgery, and improve the success rate of the operation. Additionally, for patients in poor health who cannot undergo immediate surgery, preoperative treatment can control the disease and improve prognosis. These treatment decisions were based on the latest clinical guidelines and the discussions of the multidisciplinary team.

Machine learning models

After collecting and analyzing demographic characteristics, preoperative laboratory findings, and surgical details of the patients, all 16 variables were included in the model. Univariate analysis was conducted for feature selection, filtering potential predictor variables with a p-value less than 0.05. Subsequently, variables were included in a multivariate logistic regression model, and a random forest algorithm was employed for further feature selection to compute the importance scores of features.

Following the feature selection process, three different machine learning models were developed and trained to predict the risk of incisional infection after right hemicolectomy: random forest, support vector machine (SVM), and deep learning models. Random forest is an ensemble method combining multiple decision trees to improve accuracy. Support vector machine identifies the hyperplane that maximally separates classes, while the fully connected neural network model utilizes interconnected layers of neurons to learn complex relationships between inputs and outputs.

The deep learning model consists of two hidden layers with 96 and 48 neurons, respectively. All layers except the last one utilize the Relu activation function, while the last layer adopts the Sigmoid activation function.

Data preprocessing

Prior to model training, the dataset was divided into training and testing sets in a 7:3 ratio using stratified sampling to ensure balanced representation of incisional infection cases in both sets. Missing data were handled through imputation techniques such as mean imputation or using domain-specific knowledge for categorical variables.

Model training and evaluation

The training set was used to train each machine learning model, where hyperparameters were optimized using techniques such as grid search or random search. The performance of each model was evaluated using the testing set through metrics including accuracy, precision, recall, F1 score, and area under the receiver operating characteristic curve (ROC-AUC).

Definition of incisional infection and follow-up time

In this study, we defined incisional infection based on the Centers for Disease Control and Prevention (CDC) guidelines for surgical site infections. We categorized incisional infections into superficial incisional infection and deep incisional infection. Superficial incisional infection refers to an infection occurring in the skin and subcutaneous tissue of the incision, while deep incisional infection refers to an infection occurring in the fascial and muscle layers. We conducted a 30-day follow-up for all patients to monitor and record any incisional infection events that occurred postoperatively.

Preoperative preparation, preventive measures, and wound closure suture materials

Before surgery, all patients underwent standard preoperative preparation and preventive measures. The preoperative preparation included routine mechanical bowel preparation (MBP), where all patients received oral laxatives the day before surgery to empty the bowels. All patients received prophylactic antibiotic treatment within 30 min to 1 h before the surgery, typically using cephalosporin antibiotics. For patients allergic to cephalosporins, fluoroquinolones or clindamycin were used as alternatives.

During the surgery, different types of suture materials were used for wound closure based on the surgical team’s choice. Usually, absorbable sutures (such as PDS-plus) or non-absorbable sutures (such as No-PDS) were used. The specific choice depended on the surgical site, the patient’s individual condition, and the surgeon’s preference.

Statistical analysis

Descriptive statistics were used to summarize patient demographics and clinical variables. Additionally, univariate and multivariate analyses were conducted to identify factors associated with incisional infections. Statistical significance was determined at p < 0.05.

Results

Baseline information table for colon cancer patients in training group and validation group

322 patients with colon cancer who underwent right hemicolectomy were randomly assigned to a 7:3 ratio for training and validation cohorts, with 226 patients in the training set and 96 patients in the validation set. The inclusion and exclusion process is shown in Fig. 1. In the training cohort, there were 84 patients aged over 60, accounting for 37.2% of the total. Preoperative chemotherapy was administered to 46 patients, representing 20.4%, while 27 patients, or 11.9%, received preoperative radiotherapy. The mean tumor diameter was 3.2 cm, and the majority of patients underwent laparoscopic surgery. No statistically significant differences were observed in any variables between the two groups (p > 0.05). Furthermore, a total of 14 patients experienced incisional infections. Detailed data for both patient groups are presented in Table 1. Ninety patients (28%) received chemotherapy or radiotherapy before surgery. The T4 stage patients shown in Table 1 are classified based on postoperative pathological staging (ypT4). These patients were identified as advanced or locally advanced through preoperative imaging and clinical evaluation, and thus received neoadjuvant therapy. All patients were divided into two groups based on laparoscopic and open surgery. All variables were evenly distributed between the two groups (P > 0.05) (Supplementary Fig. 1).

Fig. 1

Flow chart for inclusion and exclusion of patients with Colon cancer who underwent Right Hemicolectomy

Table 1 Baseline characteristics of patients with Colon cancer undergoing resection in the training and validation cohorts (n = 322)

Univariate and multivariate logistic regression to identify factors influencing incisional infections

Univariate regression identified statistical significance (P < 0.05) for ASA stage, preoperative chemotherapy, operation time, ALB, and surgical approach. These variables were included in the multivariable regression model. Multivariable Cox Regression Analysis Indicates that ASA Stage (HR = 1.081 [1.033–1.204], p = 0.028), Operation Time (HR = 1.823 [1.430–2.150], p < 0.001), ALB (HR = 0.854 [0.517–0.984], p = 0.022), Surgical Approach (HR = 1.780 [1.378–2.355], p < 0.001), and Sutures Used at Closure (HR = 0.710 [0.455–1.916], p < 0.001).(Table 2).

Table 2 Multivariate and Univariate logistic regression analysis of patients with Incisional Infection after right hemicolectomy in the training cohort

Model characteristics of different predictors of incisional infections and the ROC curve

After multiple automatic parameter adjustments in the random forest model, and seeking the best compromise between training and testing set performance, a final selection of 100 decision trees with a maximum depth of 8 and a minimum sample requirement of 2 for each split was made. The top 5 most important variables in the model were determined as ASA, operation time, ALB, sutures used at closure, and surgical approach (Fig. 2A). ROC curve analysis demonstrated an area under the curve (AUC) of 0.844 for the training cohort and 0.757 for the validation cohort (Fig. 2B). Similarly, for the Support Vector Machine (SVM) model, an RBF kernel with C = 2 and γ = 0.2 was utilized, yielding an AUC of 0.786 for the training cohort and 0.737 for the validation cohort (Fig. 3). Lastly, for the Deep Neural Network (DNN) model, two hidden layers with 96 and 48 neurons were employed. The model architecture diagram is shown in Fig. 4A, with AUC values of 0.885 and 0.879 for the training and validation sets, respectively (Fig. 4B).

Fig. 2

Performance of the Random Forest Model. A represents the five most important variables, and B represents the ROC curve of the model on the training and validation sets

Fig. 3

ROC curves of the Support Vector Machine model on the training and validation sets

Fig. 4

Performance of the Deep Learning Model. A represents the model architecture diagram, and B represents the ROC curve of the model on the training and validation sets

Performance of multiple models

The model with the best performance overall is the deep learning model, with F1-scores of 0.906 and 0.858 on the training and validation sets, respectively, and AUCs of 0.885 (0.781 − 0.670) and 0.879 (0.768–0.963), respectively. In addition, the support vector machine model exhibits relatively poorer performance, with ROC-AUCs of 0.786 and 0.737. The logistic regression model performs second-best, also demonstrating good predictive efficacy, with F1-scores of 0.882 and 0.791 on the training and validation sets, respectively, and AUCs of 0.863 and 0.796 (Table 3).

Table 3 Evaluation indicators for each model

Discussion

Postoperative complications such as surgical site infection (SSI) in patients undergoing colon cancer surgery can lead to prolonged hospital stays, increased postoperative antibiotic usage, higher rates of reoperation, and heightened psychological stress for patients [14, 15]. Additionally, SSI can result in increased healthcare costs and decreased rates of both recurrence-free survival and overall survival (OS) [16, 17]. This is particularly pertinent for patients undergoing right hemicolectomy for colon cancer, as the anatomical complexity of the right colon and the technical and experiential demands of the procedure may pose increased risks for postoperative complications compared to left hemicolectomy. Furthermore, right hemicolectomy involves more colon segments, potentially elevating the risk of postoperative complications [18, 19]. Given the limitations of traditional predictive models, our study aims to develop a predictive model using machine learning techniques alongside traditional logistic regression to identify factors influencing postoperative incisional infections after right hemicolectomy for colon cancer. Our results demonstrate the effectiveness of various machine learning algorithms, including random forests, support vector machine (SVM), and deep learning, in predicting the risk of postoperative infections.

One key finding of our study is the identification of several important predictive factors associated with incisional infections. The top five important variables in the random forest model include ASA classification, duration of surgery, preoperative serum albumin level (ALB), surgical approach, and type of suture used for wound closure, all of which were found to be significant predictors of incisional infections. These findings align with previous literature; for example, Nakamura et al [5]. investigated risk factors for wound infections after laparoscopic colon cancer surgery and found that preoperative serum albumin level of 2.5 g/dl, FEA, and non-PDS sutures were risk factors. Sutton et al [20]. found a delayed onset of incisional infections in colon cancer patients undergoing laparoscopic surgery. Since preoperative serum albumin level indicates poor nutritional status of patients, wound recovery postoperatively tends to be slower. Lai et al [21]. suggested that hypoalbuminemia is a predictor of poor outcomes in colon cancer surgery and recommended nutritional therapy for malnourished patients 7 to 14 days before surgery. Given that laparoscopic surgery is minimally invasive and characterized by smaller incisions compared to open surgery, the widespread adoption of laparoscopic surgery makes it preferable for patients undergoing minimally invasive procedures. Previous studies have also made predictions about Surgical Site Infections (SSI). Wu et al [22]. developed a predictive model for SSIs following gastrointestinal surgery using multicenter national data. This model demonstrated good discriminative ability, calibration, and clinical utility, making it a valuable reference tool for predicting the risk of SSIs in patients. Additionally, Smith et al [23]. reviewed predictive models for SSIs and suggested that further research is needed to develop more purpose-specific risk tools.

In recent years, artificial intelligence models such as machine learning have become increasingly popular, and our research has further highlighted the superiority of deep learning models in predicting incisional infections compared to traditional logistic regression and other machine learning algorithms [24, 25]. Deep learning models demonstrated the highest performance metrics, including F1 score and area under the receiver operating characteristic curve (ROC-AUC), on both the training and validation sets. This indicates that deep learning algorithms have the ability to learn complex patterns and relationships from data, potentially providing higher predictive accuracy in identifying high-risk patients.

Incorporating machine learning techniques into predictive modeling of surgical site infections represents a significant advancement in personalized medicine and perioperative care. By utilizing these advanced analytical tools, clinicians can better assess individual patient risks and accordingly devise prevention strategies. For instance, high-risk patients identified by the predictive model can undergo closer postoperative monitoring, receive targeted interventions to optimize wound healing, and be prescribed personalized antibiotic prophylaxis regimens to reduce infection risks. Researchers such as Zhu et al [26]. achieved similarly high accuracy by developing machine learning algorithms to predict surgical complications and validating them at other centers.

Furthermore, our research underscores the importance of comprehensive data collection and analysis in predictive modeling of surgical outcomes. By incorporating a wide range of demographic, clinical, and surgical variables in our analysis, we were able to develop robust predictive models capable of capturing the complexity of factors influencing postoperative infections. This highlights the potential of machine learning models to leverage diverse datasets and uncover hidden insights that may not be easily discernible through traditional statistical methods.

Machine learning models, particularly deep learning models, provide an effective method for predicting the risk of incisional infection after right hemicolectomy. To effectively apply these AI tools in clinical practice, we propose the following: AI can first be used as a preoperative risk assessment tool: AI models can be integrated into preoperative assessment systems to help surgeons evaluate the risk of incisional infection for each patient before surgery. By combining individual patient characteristics and surgical variables, AI tools can provide personalized risk predictions, aiding in the development of more precise preoperative preparations and preventive measures. Then, real-time decision support system: During surgery, AI tools can function as part of a real-time decision support system, alerting the surgical team to potential high-risk situations and offering corresponding management suggestions. This helps enhance the safety and effectiveness of the surgery and reduces the incidence of postoperative complications. Next, postoperative monitoring and management: AI models can also be used for postoperative monitoring, assisting doctors in identifying high-risk patients and taking timely intervention measures. For instance, for patients predicted to have a higher risk of infection, doctors can intensify postoperative care and follow-up to prevent the occurrence of infections. Finally, education and training: AI tools can be used in medical education and training, helping medical students and junior doctors understand and apply advanced predictive models, thereby improving their clinical decision-making skills. For patients predicted to be at high risk post-surgery, preoperative preparations can be enhanced. For instance, malnourished patients can undergo nutritional therapy before surgery to improve serum albumin levels. During the surgery, selecting appropriate suturing materials and techniques can reduce the risk of infection. For example, using absorbable sutures may help lower the incidence of incision infections. High-risk patients should receive closer postoperative monitoring and care, including personalized antibiotic prophylaxis to prevent infections.

While the results are promising, our study has several limitations that need to be considered. Firstly, the retrospective nature of the study design may introduce biases and confounding factors that could affect the results. Additionally, the study was conducted at a single institution, which may limit the generalizability of the findings. Future research should validate the predictive accuracy and applicability of the developed model in larger multicenter cohorts and prospective designs. In the future, we will increase the sample size and further validate our research findings. A larger sample size will help improve the predictive accuracy of our model and ensure the robustness of the results. Additionally, by collaborating with multiple centers, we will collect data from different regions and patient populations through partnerships with other medical centers. This will not only validate the applicability of our model in various clinical settings but also provide a broader clinical reference. Finally, we will conduct prospective studies to further minimize the potential biases introduced by retrospective studies. Prospective data collection and analysis will provide higher-quality evidence to support our research conclusions.

However, it must be acknowledged that in the era of laparoscopic and robotic surgery, the clinical consequences of SSIs following right hemicolectomy are relatively minor due to the use of the Pfannenstiel incision. In contrast, other potential postoperative complications in colorectal cancer surgery, such as anastomotic leakage and low anterior resection syndrome (LARS), have a more significant impact on patient prognosis and quality of life. These more severe complications require highly accurate predictive models to guide clinical decision-making and patient care. Future research should focus on applying these advanced analytical techniques to these more serious clinical scenarios to further enhance the level of personalized medicine.

In conclusion, our study demonstrates the utility of machine learning techniques in predicting factors related to incisional infections following right hemicolectomy for colon cancer. The development of accurate predictive models holds the potential to improve risk stratification, guide clinical decision-making, and ultimately enhance postoperative outcomes in colon cancer surgery. Further research and validation of these models in diverse clinical settings are necessary to fully realize their clinical utility and impact on patient care.