Illuminating the black box: Machine learning enhances preoperative prediction in intrahepatic cholangiocarcinoma

Abstract

The study by Huang et al, published in the World Journal of Gastroenterology, advances intrahepatic cholangiocarcinoma (ICC) management by developing a machine-learning model to predict textbook outcomes (TO) based on preoperative factors. By analyzing data from 376 patients across four Chinese medical centers, the researchers identified key variables influencing TO, including Child-Pugh classification, Eastern Cooperative Oncology Group score, hepatitis B status, and tumor size. The model, created using logistic regression and the extreme gradient boosting algorithm, demonstrated high predictive accuracy, with area under the curve values of 0.8825 for internal validation and 0.8346 for external validation. The integration of the Shapley additive explanation technique enhances the interpretability of the model, which is crucial for clinical decision-making. This research highlights the potential of machine learning to improve surgical planning and patient outcomes in ICC, opening possibilities for personalized treatment approaches based on individual patient characteristics and risk factors.

Key Words: Intrahepatic cholangiocarcinoma; Textbook outcome; Machine learning; Predictive model; Shapley additive explanations; Preoperative assessment; Surgical outcomes; Disease-free survival; Extreme gradient boosting; Clinical decision-making

Core Tip: The extreme gradient boosting model, used in conjunction with the Shapley additive explanation algorithm, as described by Huang et al, offers a revolutionary outlook into the future of surgical oncology for patients with intrahepatic cholangiocarcinoma. This model identifies crucial preoperative factors that influence patient outcomes, enhances understanding of disease progression and treatment efficacy, and underscores its utility in clinical decision-making for patient care and surgical interventions. Moreover, its accurate predictive prognostic potential offers insights into successful treatment mechanisms and personalized care strategies.

TO THE EDITOR

The 21^st century has witnessed significant improvements in surgical oncology, especially with the integration of artificial intelligence (AI) modalities, which have significantly improved patient outcomes, surgical precision, and diagnostic accuracy[1-4]. Clinically, intrahepatic cholangiocarcinoma (ICC) is primarily characterized by significant challenges owing to its aggressive nature and high recurrence rate[5,6]. Moreover, conventional prognostic models have relied significantly on postoperative pathological findings in ICC care, limiting their utility in surgical decision-making. These challenges highlight the growing need for more advanced, complex, and highly effective prognostic models to increase the accuracy of postoperative risk assessment[7,8].

To this effect, a recent publication in the World Journal of Gastroenterology by Huang et al[1], which assessed the use of machine learning algorithms in surgical intervention for ICC, represents a significant advancement in treating and managing this disease. Huang et al[1] sought to develop an interpretable machine-learning model capable of predicting textbook outcomes (TO) in patients with ICC before surgical intervention to aid in clinical decision-making and improve patient management. The model developed by Huang et al[1] has significant implications for the management of ICC in clinical practice. It enables preoperative prediction of TOs, allows personalized treatment planning, and improves patient counseling. The model’s high accuracy and interpretability through Shapley additive explanation (SHAP) analysis provides clinicians with valuable insights into the key factors influencing surgical outcomes, potentially leading to more informed decision-making and optimized perioperative care strategies for patients with ICC. The research by Huang et al[1] is particularly noteworthy for its novel application of advanced computational techniques to address complex medical problems by identifying subtle patterns and correlations in patient data that may not be apparent through conventional analysis. This approach can potentially revolutionize ICC treatment and offer a more personalized, data-driven strategy for patient management[1,8].

Understanding TOs

Tsilimigras et al[9] define TOs as “the optimal course following surgery that is better aligned with patients’ expectations around optimal care”. This composite metric encompasses several factors, including clear surgical margins, absence of perioperative blood transfusion, severe complications, prolonged hospitalization or readmissions, and postoperative mortality for 30 days[10-12]. Predicting a patient’s likelihood of achieving these outcomes before surgery is vital for effective surgical planning and optimal resource management. Moreover, research has shown that achieving TOs improves both short- and long-term patient outcomes. Nathan et al[13] highlighted that patients who met TO criteria had higher survival rates compared to those who did not.

Huang et al[1] built on this assertion by employing an extreme gradient boosting (XGBoost) model to assess preoperative clinical parameters from 376 patients with ICC undergoing liver resection between 2011 and 2017. Among the studied cohort, 150 patients (52.3%) in the training group and 49 patients (55.1%) in the validation group achieved TO. Patients who achieved TO had a disease-free survival (DFS) rate of 64.2% at 1 year, 56.8% at 2 years, and 43.4% at 3 years. In contrast, patients who did not achieve TO had substantially lower DFS rates of 44.7%, 32.5%, and 25.2% at the same time points (P < 0.05), underscoring the clinical utility of TO as a valuable predictor of postoperative recovery and long-term prognosis. The four key preoperative factors were the Eastern Cooperative Oncology Group score, hepatitis B status, Child-Pugh score, and tumor size. Corresponding logistic regression analysis revealed a significant correlation between these variables and TO achievement.

The ability to accurately predict preoperative TO has immense clinical value. Typically, ICC prognosis has been evaluated using postoperative pathological markers; however, this study shifts the paradigm by employing an XGBoost model to assess preoperative clinical factors. The model showed high predictive accuracy, with area under the curve values of 0.8825 (internal validation) and 0.8346 (external validation)[1]. The XGBoost machine learning model was highlighted by Xu and Lu[14] in a SEER-based comparative study of 1055 patients with ICC in conjunction with the American Joint Commission on Cancer (AJCC) system. The study reported AUC, sensitivity, specificity, and a positive predictive value of 0.811, 0.573, 0.890, and 0.849, respectively, compared to 0.713, 0.478, 0.814, and 0.726, respectively, for the AJCC system. These findings underscore the superior predictive performance of machine learning models over traditional staging systems in the prognosis of ICC.

Lim et al[15] applied the XGBoost model in a multi-institutional survival estimation analysis of 993 patients with ICC in Korea, analyzing survival probabilities at 3, 6, and 12 months. The model demonstrated strong predictive performance, with AUC values of 0.778, 0.794, and 0.784 for the respective time points. Compared to these two studies, a distinguishing feature of the approach used by Huang et al[1] was the incorporation of the SHAP algorithm.

The SHAP algorithm

Rasheed et al[16] describe the SHAP algorithm as a game-theory-based approach designed to improve the transparency of machine learning models by assigning importance scores to each predictor variable. This algorithm quantifies the contribution of each feature to the prediction of a model, offering an intuitive visualization of the influence of specific factors on individual patient outcomes. Unlike traditional black-box AI models, SHAP values allow clinicians to understand the rationale behind predictions and increase their trust in machine learning-assisted decision-making. This algorithm, in combination with XGBoost, has previously been used for risk assessment and individual prediction of central cervical lymph node metastasis in patients diagnosed with papillary thyroid carcinoma, as demonstrated by Zou et al[17]. The study utilized an XGBoost model combined with the SHAP algorithm. XGBoost is an advanced machine-learning technique that enhances predictive accuracy through iterative tree-based learning, while SHAP improves model interpretability by quantifying the contribution of each feature to predictions, which is crucial for clinical decision-making in the management of ICC.

In the study by Huang et al[1], SHAP analysis ranked the relative importance of variables influencing TO prediction, identifying tumor size as the most impactful factor, followed by Child-Pugh classification, hepatitis B status, and ECOG score. Moreover, the SHAP summary plots revealed that larger tumor sizes negatively influenced TO probability, whereas better Child-Pugh scores and the absence of hepatitis B infection increased the likelihood of achieving TO. Additionally, SHAP dependence plots highlighted the interactions between variables, demonstrating how different variables interplay to influence outcomes.

Implications for surgical oncology

The findings of Huang et al[1] findings have significant implications for surgical oncology and hepatobiliary surgery. First, by identifying high-risk patients preoperatively, clinicians can implement tailored interventions such as prehabilitation programs, enhanced perioperative monitoring, and more aggressive multimodal management strategies. Additionally, accurate TO prediction can improve patient counseling, allowing for more realistic expectations regarding postoperative recovery and long-term prognosis.

Furthermore, the strong correlation between TO achievement and improved DFS emphasizes the importance of perioperative quality metrics in the management of ICC. Patients who achieved TO exhibited significantly better DFS rates at 1, 2, and 3 years compared to those who did not, thereby reinforcing the clinical utility of TO as a composite outcome measure.

Challenges and future directions

Despite its strengths, the study by Huang et al[1] has several limitations. As this was a retrospective analysis, inherent selection bias cannot be ruled out. Additionally, excluding other potentially influential variables, such as body mass index, comorbidities, and inflammatory markers, may limit the model’s predictive capacity. Expanding the dataset to include diverse populations, particularly from non-hepatitis B virus-endemic regions, could enhance the generalizability of the findings. Future research should explore the integration of radiomic and genomic biomarkers to further enhance predictive accuracy. Additionally, prospective validation in a randomized controlled setting would be instrumental for confirming the clinical applicability of this model.

CONCLUSION

Machine-learning framework that advances the preoperative assessment of patients with ICC is innovative. The model not only enhances the prognostic accuracy but also bridges the gap between AI and its clinical applicability. As machine learning continues to revolutionize oncologic care, studies like this pave the way for more personalized, data-driven surgical strategies, ultimately improving patient outcomes in ICC.