Artificial intelligence in liver cancer surgery: Predicting success before the first incision

Abstract

Advancements in machine learning have revolutionized preoperative risk assessment. In this article, we comment on the article by Huang et al, which presents a recent multicenter cohort study demonstrated that machine learning algorithms effectively stratify recurrence-free survival, providing a robust predictive framework for maximizing surgical outcomes in intrahepatic cholangiocarcinoma. By leveraging interpretable models, the research enhances clinical decision-making, allowing for more precise patient selection and personalized surgical strategies. These findings highlight the growing role of artificial intelligence in optimizing surgical outcomes and improving prognostic accuracy in hepatobiliary oncology.

Key Words: Intrahepatic cholangiocarcinoma; Artificial intelligence; Machine learning; Surgery; Cancer

Core Tip: Machine learning is revolutionizing surgical planning for intrahepatic cholangiocarcinoma by enabling preoperative prediction of “textbook outcomes” through interpretable artificial intelligence models. This approach enhances precision in patient selection, optimizes surgical strategies, and reduces unnecessary procedures, paving the way for more personalized hepatobiliary oncology care.

INTRODUCTION

The landscape of surgical oncology is evolving rapidly with the integration of artificial intelligence (AI) into clinical decision-making. A groundbreaking study by Huang et al[1] highlights the power of AI in predicting surgical success even before the first incision is made. This multicenter cohort study demonstrates that machine learning models can effectively stratify recurrence-free survival, offering a valuable tool for preoperative planning. By analyzing a range of preoperative factors including Child-Pugh classification, Eastern Cooperative Oncology Group score, hepatitis B, and tumor size, this model provide interpretable predictions that enhance patient selection, improve surgical strategies, and ultimately optimize patient outcomes.

THE CHALLENGE OF INTRAHEPATIC CHOLANGIOCARCINOMA SUGERY

Intrahepatic cholangiocarcinoma (ICC) is a highly aggressive malignancy and the second leading cause of liver cancer-related mortality, characterized by variable prognoses and a critical need for early detection[2,3]. In the United States and Europe, major risk factors include primary sclerosing cholangitis, choledochal cysts, gallstones or chronic liver disease such as viral hepatitis or cirrhosis[4]. Despite these known risk factors, surgical planning for ICC remains challenging due to complexities in diagnosis and staging. ICC frequently presents at an advanced stage because of its insidious onset and non-specific symptoms, making early identification difficult. Due to challenges with early detection, only 50% of patients are initially considered for surgical resection at the time of diagnosis, but many of these patients are subsequently deemed ineligible even after neoadjuvant treatment[5]. The highly invasive nature of ICC leads to multifocality, lymph node metastasis, and vascular invasion, which further complicates achieving clear surgical margins. Even for those select few patients who are candidates for surgery, morbidity and mortality remain high, as extensive hepatectomies are often required but may be contraindicated in patients with underlying liver dysfunction or insufficient future liver remnant. Moreover, postoperative complications, including bile leaks and liver failure along with post-resection recurrence rates are notably high, with more than 60% of patients experiencing recurrence within a year post-resection[6]. This high recurrence rate significantly impacts long-term survival, which remains poor despite multimodal treatment approaches.

Although surgical resection remains the gold standard treatment for ICC, predicting which patients would be most likely to achieve favorable postoperative outcomes is difficult. Traditionally, surgeons rely on clinical experience and general prognostic factors to estimate the likelihood of complications or adverse events. However, the variability in patient physiology, functional status, tumor biology, and perioperative care makes this a complex task. This is where predictive analytics and machine-learning models can provide a much-needed advantage. The study by Huang et al[1] highlights how AI-driven algorithms can provide individualized risk assessments. Such tools aim to identify patients most likely to achieve a “textbook outcome,” characterized by a surgery free of major complications and a favorable postoperative recovery trajectory.

THE SIGNIFICANCE OF TEXTBOOK OUTCOME IN ICC SURGERY

A textbook outcome (TO) is a composite metric defined by negative surgical margins, no perioperative transfusion, no postoperative complications, no prolonged length of stay, no 30-day readmissions, and no 30-day mortality[7]. Achieving TO is not merely an academic exercise; it directly translates to improved patient well-being, reduced healthcare costs, and better long-term survival. Traditional indicators of surgical success, such as low complication rates or short length of stay, often provide only a partial assessment. In contrast, TO integrates multiple dimensions of the perioperative course, reflecting the quality of preoperative evaluation, surgical skill, and postoperative management. By prioritizing TO, clinicians are encouraged to optimize patient care across all stages of the surgical pathway. Given the difficulties in early diagnosis and accurate staging of ICC, achieving a TO indicates successful navigation of these initial hurdles. It indicates that the tumor was identified at a stage amendable to curative resection and that the surgical team could proceed without significant complications or the need for additional procedures.

AI-DRIVEN PRECISION IN SURGICAL PLANING

The AI approach not only enhances the precision of surgical interventions but also reduces unnecessary procedures for high-risk patients. As machine learning continues to advance, its role in hepatobiliary oncology will expand, paving the way for more personalized and data-driven treatment strategies. By leveraging machine learning for risk assessment, clinicians can move toward a more tailored and proactive approach, reducing the burden of ICC through early detection and prevention. Alaimo et al[8] applied AI to establish the optimal resection margin for ICC in a cohort of 600 patients. By employing both the optimal survival tree and optimal policy tree methods, they discovered that individuals with tumors measuring less than 4.8 cm and a surgical margin of at least 2.5 mm experienced a 37% relative increase in five-year overall survival compared to the broader patient group. These data-driven insights have implications for refining margin-width guidelines and ultimately improving survival rates.

Another study, Altaf et al[9] developed an AI-driven model to predict when immediate surgical intervention for ICC may prove futile. Utilizing a combination of machine learning and deep learning strategies, the model incorporated 10 preoperative factors to identify those at high risk for death or recurrence within 12 months post-surgery. The ensemble approach which integrates multilayer perceptron and gradient boosting classifiers achieved high accuracy, demonstrating an area under the curve of 0.830 (95% confidence interval: 0.798-0.861) in the training set and 0.781 (95% confidence interval: 0.707-0.853) in the testing set. Radiologic tumor burden score, serum carbohydrate antigen 19-9, and direct bilirubin levels were the factors most strongly predictive of futile surgery, assisting clinicians to avoid non-beneficial procedures.

Additionally, Takamoto et al[10] validated an AI-supported simulation for automated three-dimensional liver reconstruction in virtual hepatectomy. Their findings revealed that AI-based techniques substantially reduced processing time (2.1 minutes vs 35.0 minutes, P < 0.001) and provided accurate volumetric analyses for surgical planning. Taken together, these studies highlight the potential of AI-driven precision in improving surgical planning, optimizing resection margins, and predicting surgical outcomes for ICC, thereby enhancing patient management and prognosis. While machine learning-based predictive models are increasingly utilized in oncology and surgery, their “black box” nature remains a key concern. Clinicians may hesitate to adopt models whose underlying decision-making processes are not transparent. Huang et al[1] address this issue by integrating the Shapley Additive Explanations technique, which clarifies how each individual variable influences the prediction. This level of interpretability is essential for clinical use, as it enables surgeons to understand why a given patient is, or is not, predicted to achieve a TO.

The key findings of the study are compelling. First, logistic regression analysis identified Child-Pugh classification, Eastern Cooperative Oncology Group score, hepatitis B, and tumor size as significant preoperative predictors of TO. Second, the XGBoost model demonstrated strong predictive accuracy in both internal and external validation, achieving area under the curve values of 0.8825 and 0.8346, respectively, thus indicating a robust capacity to discriminate between patients likely to reach TO and those who are not. Lastly, the Shapley Additive Explanations algorithm enhanced model interpretability band transparency, allowing clinicians to understand the relative importance of each predictor and facilitate the application of the model’s results.

POTENTIAL LIMITATIONS AND FUTURE DIRECTIONS

While machine-learning techniques represent a noteworthy advance in predictive modeling for ICC surgery, several challenges persist. First, model generalizability and data set shift remain pertinent issues: Many AI models exhibit strong performance in their original training environment but fail to achieve comparable accuracy when tested on external datasets or deployed in distinct clinical settings with more diverse patient populations[11]. This “data set shift” can undermine both reliability and overall efficacy in real-world practice[11]. Second, data quality and bias significantly affect model accuracy. Inadequate data preparation, collection, labeling, and sampling can introduce biases that do not represent authentic clinical scenarios[12]. In addition, overfitting is a common concern, particularly in deep learning models, where extensive focus on training data - including noise and outliers - compromises generalizability to novel datasets[13]. Another major barrier as we previously mentioned is transparency and interpretability: Many AI algorithms operate as “black boxes”, making it difficult for clinicians to understand how decisions are made. This lack of transparency can hinder trust and acceptance among healthcare providers and complicate the integration of AI into clinical workflows[13].

Another key limitation is the lack of prospective studies. Most AI research in ICC has relied on retrospective datasets, which restricts the ability to draw causal inferences. Although retrospective analyses offer valuable preliminary insights, prospective validation in clinical trials is crucial for confirming whether AI-assisted surgical planning improves patient outcomes in real-world settings. Finally, effectively integrating AI into routine clinical practice, such as within electronic medical record systems, requires seamless collaboration between AI platforms and clinicians. This integration involves ensuring that AI-generated recommendations are both clinically actionable and easily interpretable. Equally important is providing clinicians with sufficient training and user-friendly tools, including automated risk scoring systems, so that AI outputs can be readily adopted in everyday practice.

CONCLUSION

Huang et al[1] demonstrate that machine learning can help predict the likelihood of achieving TO in patients with ICC. By refining preoperative planning, guiding treatment selection, and facilitating patient counseling, this approach promises to enhance surgical management. While further prospective research is needed to validate the findings externally and address the limitations of the study, this work represents a significant step forward in the pursuit of improved ICC outcomes. With further developments, machine learning-based TO prediction could increase surgical precision, minimize complications, and ultimately bolster survival. Integrating such tools into clinical workflows may lead to more personalized, data-driven interventions for patients with this challenging malignancy.