Application and challenges of artificial intelligence in predicting perioperative complications of colorectal cancer

Abstract

Colorectal cancer (CRC) is a prevalent malignancy, with surgery playing a key role in its treatment. However, perioperative complications, such as anastomotic leaks, infections, and mortality, can significantly affect surgical outcomes, extend hospital stays, and increase healthcare costs. Traditional risk prediction models often lack precision, leading to increased interest in artificial intelligence (AI) for improving risk stratification. This review examines the application of AI, particularly machine learning and deep learning, in predicting perioperative complications in CRC surgery. AI models have been employed to predict a variety of postoperative complications, including readmissions, surgical-site infections, anastomotic leakage, and mortality, by analyzing diverse data sources such as electronic health records, medical imaging, and preoperative markers. Despite the promising results, several challenges remain, including data quality, model generalizability, the complexity of clinical data, and ethical and regulatory concerns. The review emphasizes the need for multicenter, diverse datasets and the integration of AI into clinical workflows to improve model performance and adoption. Future efforts should focus on enhancing the transparency and interpretability of AI models to ensure their successful implementation in clinical practice, ultimately improving patient outcomes and surgical decision-making in CRC surgery.

Key Words: Artificial intelligence; Colorectal cancer; Perioperative complications; Machine learning; Predictive models

Core Tip: Artificial intelligence (AI), including machine learning and deep learning, is increasingly applied to predict perioperative complications in colorectal cancer surgery. By analyzing diverse data sources such as electronic health records, medical imaging, and preoperative markers, AI models can improve risk stratification, predict complications like anastomotic leakage and mortality, and enhance clinical decision-making. However, challenges such as data quality, model generalizability, and ethical concerns must be addressed. Future efforts should focus on developing interpretable models, utilizing multicenter datasets, and integrating AI into clinical workflows to optimize patient outcomes and ensure successful clinical adoption.

INTRODUCTION

Colorectal cancer (CRC) is a prevalent malignancy, and surgery remains a cornerstone of its treatment. Despite the advances in surgical techniques, perioperative complications such as anastomotic leaks, infections, and thromboembolic events continue to affect patient outcomes, may lead to increased healthcare costs, extended hospitalizations, and higher likelihood of patient mortality. Traditional risk prediction models often lack precision, primarily because they rely on static, population-level data and predefined scoring systems that may not fully capture individual patient variability or real-time physiological changes. Consequently, they often provide only broad risk stratifications rather than precise, individualized risk estimations. This limitation has driven the interest in artificial intelligence (AI) as a potential solution to enhance risk stratification and provide more accurate predictions. AI, including machine learning (ML) and deep learning, has the ability to analyze vast amounts of clinical data to improve the prediction of perioperative complications in CRC surgery. However, there are several challenges in AI implementation, such as data quality, model generalizability, and ethical concerns. This review examines the applications of AI in predicting perioperative complications, the challenges faced, and the future directions necessary for successful clinical integration.

APPLICATIONS OF AI IN PREDICTING PERIOPERATIVE COMPLICATIONS

Predictive models for postoperative complications

AI and ML have been successfully applied to predict postoperative complications in CRC patients. These models analyzed a variety of perioperative complications, including anastomotic leakage, surgical-site infections, and unplanned readmissions. The neural network model incorporated electronic health record data, including preoperative laboratory values (e.g., albumin, hemoglobin), intraoperative factors (e.g., operative time, blood loss), and demographic data to enhance prediction accuracy. For instance, a recent study developed an artificial neural network (ANN) to predict major postoperative complications and readmission in CRC surgery. This ANN model demonstrated superior performance over traditional regression models, highlighting its potential to improve risk prediction accuracy[1]. Another study developed an improved ML model, MGA-extreme gradient boosting (XGBoost), to predict postoperative infectious complications in elderly patients, achieving an area under the curve (AUC) of 0.862, outperforming other ML algorithms tested[2].

Preoperative risk assessment

AI models have also shown significant promise in preoperative risk assessment for complications such as anastomotic leakage. A study using a feed-forward multilayer perceptron network found that preoperative immune-inflammatory markers like the systemic immune-inflammation index and neutrophil-to-lymphocyte ratio were significant predictors of postoperative complications[3]. Additionally, a nomogram model incorporating preoperative factors such as the prognostic nutritional index, albumin levels, American Society of Anesthesiologists (ASA) scores, and tumor diameter was developed to predict postoperative complications. This model achieved an AUC of 0.835 in the training cohort, underscoring the importance of integrating multiple clinical markers in risk prediction[4].

Prediction of specific complications

AI has been effectively utilized to predict specific complications following CRC surgery, including anastomotic leakage, surgical-site infections, and low anterior resection syndrome (LARS). For example, a Random Forest model was developed to predict anastomotic leakage, achieving an AUC of 0.78, and identified key risk factors such as operation time and ASA scores[5]. Another ML model predicted LARS by incorporating factors such as anastomotic height and neoadjuvant therapy. This Random Forest model achieved an AUC of 0.852 in an external validation set[6], demonstrating the utility of AI in predicting complex, site-specific complications.

Mortality prediction

AI models have also been utilized to predict both short-term and long-term mortality following CRC surgery. A study using a least absolute shrinkage and selection operator logistic regression model predicted 30-day mortality, achieving an AUC of 0.871. This model demonstrated excellent discrimination and calibration, offering potential for its integration into clinical practice. Another study, using an XGBoost model, predicted in-hospital mortality in emergency colorectal surgery. This model achieved an AUC of 0.814, with significant predictors including age, CRC diagnosis, and serum lactate dehydrogenase levels.

Enhanced decision support systems

AI-based decision support systems have been developed to aid clinicians in preoperative and perioperative decision-making. For example, a ML -based calculator was created to predict five perioperative risk events, including surgical site infection and mechanical ventilation, in emergency general surgery. This calculator demonstrated stronger predictive performance compared to traditional scoring systems[7], highlighting the potential of AI to support clinical decision-making in complex cases.

CHALLENGES OF AI IN PREDICTING PERIOPERATIVE COMPLICATIONS

Data quality and availability

One of the major challenges in developing AI models for perioperative complication prediction is the quality and availability of data. Many studies rely on retrospective data, which often contains missing or incomplete information, potentially introducing bias and affecting model performance[5,8]. For example, a study predicting anastomotic leakage noted that the performance of a Random Forest model decreased from an AUC of 0.78 on the internal test set to 0.60 on the external validation set, indicating the importance of using high-quality, multicenter data[5]. High-quality data sources should be characterized by standardized data collection processes, minimal missing data, and diverse patient populations to improve model generalizability. While a higher AUC in an external validation set is often a positive indicator, it does not inherently guarantee clinical applicability. Real-world implementation requires prospective validation, clinician interpretability, and integration into decision-making workflows to confirm utility beyond statistical performance.

Model generalizability

AI models trained on single-center or specific patient populations may lack generalizability to diverse clinical settings. A study using a Random Forest model for anastomotic leakage prediction found that the model’s performance significantly decreased when applied to an external validation set, emphasizing the need for multicenter, diverse datasets[5]. Similarly, AI models trained on data from specific geographic locations may be biased by regional differences in patient demographics, limiting their applicability in diverse populations[9]. Incorporating data from diverse populations, including different racial, geographic, and socioeconomic backgrounds, is essential for improving the external validity of AI models. This approach mitigates biases inherent to single-center datasets and ensures that AI-driven predictions are broadly applicable across different healthcare settings.

Complexity of clinical data

The complexity of clinical data, including nonlinear relationships and interactions between variables, presents a challenge for AI models. Although ML algorithms like neural networks and gradient boosting are well-equipped to handle such complexity, they require careful tuning and validation to ensure robust performance[1,3]. Strategies such as hyperparameter tuning, feature selection using recursive feature elimination, and ensemble learning (e.g., combining multiple models like random forest and XGBoost) can improve model robustness. Cross-validation and external validation on independent datasets further ensure model reliability in clinical applications.

Regulatory and ethical considerations

Integrating AI into clinical practice requires addressing significant regulatory and ethical considerations. Issues such as data privacy, algorithm transparency, and accountability must be resolved to ensure the safe and ethical use of AI in predicting perioperative complications[9,10].

Clinical validation and adoption

AI has been used in surgical settings to enhance precision and reduce complications. For instance, AI-assisted robotic systems have been employed in minimally invasive surgeries to improve accuracy and reduce recovery times[11]. Despite promising results from AI models, their routine adoption in clinical practice remains challenging. Many models need further validation across diverse clinical settings, and clinicians may be hesitant to rely on AI-driven predictions without robust evidence of their clinical utility[7,9].

FUTURE DIRECTIONS

Multicenter and diverse datasets

To improve the generalizability of AI models, future studies should focus on developing models using multicenter and diverse datasets. Collaborative efforts between institutions, as well as the use of national or international registries, can help address the challenges of model bias and improve the robustness of AI predictions[5].

Integration with clinical workflow

Efforts should be made to integrate AI models into clinical workflows by developing user-friendly tools for real-time decision-making. Web-based calculators and decision support systems can be designed to predict perioperative risks and guide surgical planning, ultimately improving clinical outcomes[7,10]. AI-driven risk prediction models must provide actionable insights that influence perioperative management. For example, an AI model predicting a high risk of anastomotic leakage may prompt preoperative interventions such as selective use of diverting stomas or closer postoperative monitoring. Similarly, AI-driven preoperative risk stratification could guide intraoperative decision-making, such as adjusting surgical techniques based on individualized risk profiles.

Enhanced model transparency and interpretability

Improving the transparency and interpretability of AI models is crucial for their adoption in clinical practice. Techniques such as SHAP values and feature importance analysis can provide clinicians with a clearer understanding of how models arrive at their predictions, thereby increasing trust in AI-driven recommendations[7].

Addressing ethical and regulatory issues

To ensure the safe and effective use of AI in healthcare, ethical and regulatory issues must be addressed. This includes ensuring data privacy, ensuring transparency in algorithms, and establishing regulatory frameworks to guide the development and deployment of AI models in clinical practice[9,10].

CONCLUSION

AI has shown significant promise in predicting perioperative complications in CRC surgery, with applications ranging from preoperative risk assessment to mortality prediction. However, challenges such as data quality, model generalizability, and regulatory considerations must be addressed to fully realize the potential of AI in clinical practice. Future directions include the development of multicenter datasets, integration of AI into clinical workflows, and enhanced model transparency and interpretability. By overcoming these challenges, AI has the potential to become a valuable tool in improving surgical outcomes and patient care in CRC surgery.