Zhu DL, Tulahong A, Liu C, Aierken A, Tan W, Ruze R, Yuan ZD, Yin L, Jiang TM, Lin RY, Shao YM, Aji T. Identification of key factors and explainability analysis for surgical decision-making in hepatic alveolar echinococcosis assisted by machine learning. World J Gastroenterol 2025; 31(37): 111038 [DOI: 10.3748/wjg.v31.i37.111038]
Corresponding Author of This Article
Tuerganaili Aji, PhD, Chief Physician, Director, Professor, Department of Hepatobiliary and Echinococcosis Surgery, The First Affiliated Hospital of Xinjiang Medical University, No. 137 South Liyushan Road, Urumqi 830054, Xinjiang Uygur Autonomous Region, China. tuergan1@163.com
Research Domain of This Article
Medical Informatics
Article-Type of This Article
Retrospective Study
Open-Access Policy of This Article
This article is an open-access article which was selected by an in-house editor and fully peer-reviewed by external reviewers. It is distributed in accordance with the Creative Commons Attribution Non Commercial (CC BY-NC 4.0) license, which permits others to distribute, remix, adapt, build upon this work non-commercially, and license their derivative works on different terms, provided the original work is properly cited and the use is non-commercial. See: http://creativecommons.org/licenses/by-nc/4.0/
Da-Long Zhu, Alimu Tulahong, Chang Liu, Ayinuer Aierken, Wei Tan, Rexiati Ruze, Zhong-Dian Yuan, Lei Yin, Tie-Min Jiang, Ying-Mei Shao, Tuerganaili Aji, Department of Hepatobiliary and Echinococcosis Surgery, The First Affiliated Hospital of Xinjiang Medical University, Urumqi 830054, Xinjiang Uygur Autonomous Region, China
Ren-Yong Lin, State Key Laboratory of Pathogenesis, Prevention, and Treatment of Central Asian High Incidence Diseases, Clinical Medical Research Institute, The First Affiliated Hospital of Xinjiang Medical University, Urumqi 830054, Xinjiang Uygur Autonomous Region, China
Author contributions: Aji T, Zhu DL, and Tulahong A designed the research study; Zhu DL, Tulahong A, Liu C, Aierken A, Tan W, Ruze R, Yuan ZD, and Yin L curated and investigated the data, as well as wrote the original draft; Zhu DL and Tulahong A conducted the formal analysis, while Zhu DL, Tulahong A, and Liu C visualized the results; Aji T administered the project and secured funding for the study; Aji T, Jiang TM, Lin RY, and Shao YM supervised the research, and Aji T, Jiang TM, Lin RY, and Shao YM reviewed and edited the manuscript; All authors have read and approved the final manuscript.
Supported by Natural Science Foundation of Xinjiang Uygur Autonomous Region, No. 2022D01D17; and State Key Laboratory of Pathogenesis, Prevention and Treatment of High Incidence Diseases in Central Asia, No. SKL-HIDCA-2024-2.
Institutional review board statement: The study was approved by the Ethics Committee of the First Affiliated Hospital of Xinjiang Medical University (approval number: K202202-03). All research was conducted in accordance with both the Declarations of Helsinki and Istanbul.
Informed consent statement: Informed consent was waived.
Conflict-of-interest statement: The authors declare that they have no competing interests.
Data sharing statement: The data from this study will be shared under reasonable requests and will be made available through our institutional repository in accordance with standard research data sharing policies. After approval from the Institutional Ethics Committee, researchers can obtain the anonymized datasets through a formal application process. Data sharing will adhere to strict privacy protection and academic ethics guidelines.
Open Access: This article is an open-access article that was selected by an in-house editor and fully peer-reviewed by external reviewers. It is distributed in accordance with the Creative Commons Attribution NonCommercial (CC BY-NC 4.0) license, which permits others to distribute, remix, adapt, build upon this work non-commercially, and license their derivative works on different terms, provided the original work is properly cited and the use is non-commercial. See: https://creativecommons.org/Licenses/by-nc/4.0/
Corresponding author: Tuerganaili Aji, PhD, Chief Physician, Director, Professor, Department of Hepatobiliary and Echinococcosis Surgery, The First Affiliated Hospital of Xinjiang Medical University, No. 137 South Liyushan Road, Urumqi 830054, Xinjiang Uygur Autonomous Region, China. tuergan1@163.com
Received: June 24, 2025 Revised: July 23, 2025 Accepted: September 9, 2025 Published online: October 7, 2025 Processing time: 95 Days and 22.3 Hours
Abstract
BACKGROUND
Echinococcosis, caused by Echinococcus parasites, includes alveolar echinococcosis (AE), the most lethal form, primarily affecting the liver with a 90% mortality rate without prompt treatment. While radical surgery combined with antiparasitic therapy is ideal, many patients present late, missing hepatectomy opportunities. Ex vivo liver resection and autotransplantation (ELRA) offers hope for such patients. Traditional surgical decision-making, relying on clinical experience, is prone to bias. Machine learning can enhance decision-making by identifying key factors influencing surgical choices. This study innovatively employs multiple machine learning methods by integrating various feature selection techniques and SHapley Additive exPlanations (SHAP) interpretive analysis to deeply explore the key decision factors influencing surgical strategies.
AIM
To determine the key preoperative factors influencing surgical decision-making in hepatic AE (HAE) using machine learning.
METHODS
This was a retrospective cohort study at the First Affiliated Hospital of Xinjiang Medical University (July 2010 to August 2024). There were 710 HAE patients (545 hepatectomy and 165 ELRA) with complete clinical data. Data included demographics, laboratory indicators, imaging, and pathology. Feature selection was performed using recursive feature elimination, minimum redundancy maximum relevance, and least absolute shrinkage and selection operator regression, with the intersection of these methods yielding 10 critical features. Eleven machine-learning algorithms were compared, with eXtreme Gradient Boosting (XGBoost) optimized using Bayesian optimization. Model interpretability was assessed using SHAP analysis.
RESULTS
The XGBoost model achieved an area under the curve of 0.935 in the training set and 0.734 in the validation set. The optimal threshold (0.28) yielded sensitivity of 93.6% and specificity of 90.9%. SHAP analysis identified type of vascular invasion as the most important feature, followed by platelet count and prothrombin time. Lesions invading the hepatic vein, inferior vena cava, or multiple vessels significantly increased the likelihood of ELRA. Calibration curves showed good agreement between predicted and observed probabilities (0.2-0.7 range). The model demonstrated high net clinical benefit in Decision Curve Analysis, with accuracy of 0.837, recall of 0.745, and F1 score of 0.788.
CONCLUSION
Vascular invasion is the dominant factor influencing the choice of surgical approach in HAE. Machine-learning models, particularly XGBoost, can provide transparent and data-driven support for personalized decision-making.
Core Tip: This study explored machine-learning applications in hepatic alveolar echinococcosis surgical decision-making. Through feature selection methods and model comparisons, we identified key factors such as vascular invasion type influencing surgical choices. The eXtreme Gradient Boosting model showed good performance and clinical benefit. Preoperative assessment combined with model assistance can enhance personalized surgical planning and improve patient outcomes, offering valuable evidence for clinical practice.
Citation: Zhu DL, Tulahong A, Liu C, Aierken A, Tan W, Ruze R, Yuan ZD, Yin L, Jiang TM, Lin RY, Shao YM, Aji T. Identification of key factors and explainability analysis for surgical decision-making in hepatic alveolar echinococcosis assisted by machine learning. World J Gastroenterol 2025; 31(37): 111038
Echinococcosis is a global zoonotic disease caused by parasites of the genus Echinococcus, with widespread prevalence in pastoral regions, posing a severe threat to human health and socioeconomic conditions. Based on clinical and etiological characteristics, echinococcosis is mainly categorized into three types: Cystic echinococcosis, alveolar echinococcosis (AE), and neotropical echinococcosis[1,2]. Among these, AE, caused by Echinococcus multilocularis, exhibits infiltrative growth characteristics similar to malignant tumors and is the most lethal type, primarily affecting the liver. Without timely treatment, the 10-year mortality rate of AE patients exceeds 90%. Currently, radical surgery combined with antiparasitic drug therapy is considered the optimal treatment for hepatic AE (HAE)[1,3-6]. However, due to the lack of specific symptoms in the early stages of HAE and its infiltrative growth, the majority of patients are diagnosed at advanced stages, with approximately 65% of patients losing the opportunity for radical surgery. For such patients, liver transplantation becomes the only curative option, but it is limited by organ shortages and the need for long-term postoperative immunosuppression[1,2,5,7]. Since 2011, when Wen et al[8] first reported the successful treatment of HAE using ex vivo liver resection and autotransplantation (ELRA), this breakthrough technique has brought new hope for cure to HAE patients who cannot undergo radical resection.
Surgical decision-making requires a comprehensive assessment of the patient's overall condition, clinical manifestations, laboratory, and imaging data. The traditional decision-making model, primarily based on hypothetico-deductive reasoning and the surgeon's personal experience, is prone to cognitive bias and has some subjectivity and limitations[9,10]. Recently, artificial intelligence (AI), particularly machine learning, has shown significant promise in medicine due to its robust data processing and pattern-recognition capabilities. Machine learning can reveal complex, nonlinear, and highly interactive relationships within large-scale data, thereby providing substantial support for disease prediction, diagnosis, and treatment decision-making[9,11,12]. Despite the widespread adoption of machine learning in medical imaging and prognostic prediction, its application to surgical strategy analysis in HAE remains largely uncharted. Recent studies have primarily focused on single-feature selection methods and black-box models, offering limited insight into the decision-making mechanisms behind these models. Thus, this study sought to bridge this gap by using multiple feature selection techniques alongside SHapley Additive exPlanations (SHAP) explainability analysis. Our objectives were to: (1) Accurately identify key factors influencing surgical decisions for HAE through cross-validation of multiple feature selection methods; (2) Develop a high-performance machine-learning model that provides data-driven decision support for personalized surgical strategies; and (3) Elucidate the decision-making mechanisms of the model to enhance the transparency and credibility of medical AI.
MATERIALS AND METHODS
Study design
This study retrospectively analyzed patients with HAE treated at the First Affiliated Hospital of Xinjiang Medical University from July 2010 to August 2024.
Inclusion criteria: (1) Preoperative clinical diagnosis and postoperative pathological confirmation of HAE; and (2) Surgical resection.
Exclusion criteria: (1) Presence of other tumors; (2) Concurrent cystic echinococcosis of the liver; and (3) Incomplete preoperative clinical data.
A total of 710 eligible patients were included in the analysis, with 545 undergoing hepatectomy and 165 receiving ELRA.
Variables and measurements
All baseline data were derived from electronic medical records during the patients' hospital stay, with surgery performed or guided by senior surgeons at our center. The collected information included patient demographics, preoperative laboratory indicators, imaging, and pathological data, such as gender, history of previous hepatectomy, lesion size and number, lesion location, vascular and biliary invasion, white blood cell count (WBC), red blood cell count (RBC), platelet count (PLT), hemoglobin (Hb), direct bilirubin, indirect bilirubin (IBil), alanine aminotransferase (ALT), aspartate aminotransferase (AST), alkaline phosphatase (ALP), γ-glutamyl transferase (GGT), albumin (ALB), globulin (GLO), total protein (TP), serum creatinine (Scr), blood urea, uric acid (UA), thrombin time, prothrombin time (PT), partial thromboplastin time, and other key clinical parameters. For variables with missing values, we used a complete case analysis method, excluding incomplete cases to ensure the reliability of the analysis results.
Data processing methods
To avoid multicollinearity among variables interfering with the analysis, Spearman correlation analysis was used. If the correlation coefficient exceeded 0.8, one of the variables was removed based on clinical significance. Continuous variables were processed using the P5-P95 capping method to reduce the impact of extreme values[13]. We chose three feature selection methods: Recursive feature elimination (RFE)[14,15], minimum redundancy maximum relevance (mRMR)[16,17], and least absolute shrinkage and selection operator (LASSO) regression, as each has its own advantages and complements the others. RFE iteratively removes the least important features; mRMR focuses on maximizing correlation with the target variable while minimizing redundancy among features; and LASSO achieves sparsity in feature coefficients through L1 regularization. By taking the intersection of these three methods, we can select the most critical features for prediction more reliably. To enhance model validity, the dataset was divided into training and validation sets in a 2:1 ratio, utilizing a stratified threefold cross-validation method. This ensured that the distribution of surgical types remained consistent across each subset, thereby bolstering the robustness of the analysis and preventing overfitting.
Machine-learning methods
The surgical decision-making for hepatobiliary echinococcosis is complex, with feature variables exhibiting both nonlinear and interaction effects. To fully explore these intricate relationships, it is necessary to leverage the advantages and disadvantages of different algorithms to enhance prediction accuracy. Accordingly, we systematically selected and compared 11 mainstream machine-learning methods: Random Forest (RF), Gradient Boosting Machine (GBM), C5.0 Decision Tree (C5.0), AdaBoost (Ada), eXtreme Gradient Boosting (XGBoost), k-Nearest Neighbors, Naive Bayes, Support Vector Machine, Gaussian Process, Logistic Regression, and Neural Network[18].
Bayesian optimization, combined with 10-fold stratified cross-validation, was used for hyperparameter optimization. The training data were divided into 10 subsets while maintaining consistent class distribution. Each set of hyperparameters was trained on nine subsets and validated on the remaining subset, repeating this process 10 times. The average validation performance [area under the curve (AUC)] served as the objective for optimization. Bayesian optimization iteratively selected potentially optimal hyperparameter combinations based on historical evaluation results, thus balancing exploration and exploitation efficiently.
To further analyze model interpretability, the SHAP method was utilized, elucidating the contributions of features to the prediction results. Model performance was comprehensively evaluated using the receiver operating characteristic (ROC) curve, calibration curve, and decision curve analysis (DCA).
Declaration of use of AI tools
In preparing this work, the authors used a chatbot to edit and proofread the manuscript. After using this tool or service, the authors reviewed and edited the content as needed and take full responsibility for the content of the published article.
Statistical analysis
Categorical variables are presented as frequencies (n) and percentages (%), with comparisons made using the χ2 test. Normally distributed continuous variables were expressed as mean ± SD, while non-normally distributed variables were presented as median (interquartile range) and compared using t tests or the Mann-Whitney U test. All statistical analyses were performed using R 4.4.2 and SPSS 26.0 software, with P < 0.05 indicating statistical significance.
RESULTS
Patient characteristics
In the collinearity analysis of the 26 variables included in the study, the correlations between RBC and Hb, ALT and AST, and GLO and TP were 0.837, 0.909, and 0.866, respectively - all exceeding 0.8. Based on clinical significance, the variables RBC, AST, and TP were removed. A total of 23 clinical variables for 545 patients who underwent hepatectomy and 165 patients who underwent ELRA were included in the analysis. After capping, the baseline characteristics are shown in Table 1 (raw data in Supplementary Table 1), indicating that ELRA patients had more severe disease, larger lesions, more involvement of liver lobes and important vascular structures, and poorer liver and kidney function and coagulation function, consistent with clinical reality. After threefold cross-validation, the dataset was divided into training and validation sets in a 2:1 ratio, with the training set (n = 474) and validation set (n = 236) having baseline characteristics as shown in Table 2 (raw data in Supplementary Table 2), with no significant differences in variable distribution between the two groups.
Table 1 Baseline characteristics of patients in the hepatectomy group and the ex vivo liver resection and autotransplantation group, n (%).
The RFE algorithm repeatedly trained models and removed the least contributing features, selecting 16 variables (Figure 1A). mRMR selected 15 variables, based on the relevance of features to the target and redundancy among features. During LASSO feature selection, all multilevel categorical variables were one-hot encoded into dummy variables, and the penalty parameter λ-min that minimized cross-validation error was chosen (Figure 1B). To maintain interpretative consistency, if any dummy variable coefficient was non-zero, its original variable was retained, and the selected levels were mapped back to the corresponding categorical variable, resulting in 16 variables (Supplementary Table 3). To ensure that the retained variables demonstrated significant predictive value across different models and to enhance the reliability and robustness of feature selection, the intersection of the results from these three methods was taken (Figure 1C), ultimately selecting 10 features (lesion size, type of vascular invasion by the lesion, degree of biliary invasion by the lesion, WBC, PLT, Hb, IBil, Scr, GGT, and PT) for machine-learning analysis.
Figure 1 Feature selection process.
A: Recursive feature elimination (RFE) algorithm results; B: Least absolute shrinkage and selection operator (LASSO) regression feature selection; C: Intersection of features selected by RFE, minimum redundancy maximum relevance, and LASSO. RFE: Recursive feature elimination; LASSO: Least absolute shrinkage and selection operator; mRMR: Minimum redundancy maximum relevance.
Model performance
The ROC curves of the 11 machine-learning models we constructed are shown in Figure 2A, and we ultimately selected the XGBoost model with an AUC of 0.872. Although the RF, GBM, C5.0, and Ada models had higher AUCs of 1.000, they were discarded due to the potential for overfitting. After parameter optimization and cross-validation, the ROC curves of the XGBoost model on the training and validation sets are shown in Figure 2B, with the training set AUC increasing to 0.935 and the validation set AUC at 0.734. The optimal threshold of the ROC curve was determined as 0.28 using the Youden index, with a corresponding sensitivity of 93.6% and specificity of 90.9%, indicating that this threshold can accurately identify ELRA candidates while minimizing misclassification. The calibration curve showed good consistency between predicted probabilities and observed proportions in most intervals, especially in the 0.2-0.7 probability range, closely fitting the ideal calibration line, indicating reliable predictive calibration of the model within this range (Figure 2C). DCA revealed that the prediction strategy based on the XGBoost model demonstrated higher net clinical benefit compared to full-intervention and no-intervention strategies, suggesting an advantage of this model in guiding clinical decision-making (Figure 2D). The model achieved an accuracy of 0.837, recall of 0.745, and F1 score of 0.788, underscoring its reliability in predicting surgical strategies for HAE.
Figure 2 Model performance and validation.
A: Receiver operating characteristic (ROC) curves of 11 machine learning models; B: ROC curves of the eXtreme Gradient Boosting (XGBoost) model on the training and validation sets after parameter optimization; C: Calibration curve of the XGBoost model, showing consistency between predicted probabilities and observed proportions; D: Decision Curve Analysis of the XGBoost model, demonstrating net clinical benefit. ROC: Receiver operating characteristic; XGBoost: EXtreme Gradient Boosting; DCA: Decision curve analysis; kNN: K-Nearest Neighbors; SVM: Support Vector Machine; GP: Gaussian Process; LR: Logistic Regression; MN: Neural Network; RF: Random Forest; GBM: Gradient Boosting Machine; C5.0: C5.0 Decision Tree; Ada: AdaBoost.
Model interpretability
Machine-learning models typically lack explainability and transparency, making it difficult to explain the influence of each variable and the decision-making process. Therefore, we used the SHAP method to interpret the output of the final model by calculating the contribution of each variable to the prediction. This interpretable method provides two types of explanations: Global explanations at the feature level and local explanations at the individual level[19]. Since the XGBoost model exhibited the best predictive performance among the 11 machine-learning models, it was chosen to interpret the output of the model. Global explanations describe the overall functionality of the model, as shown in the SHAP summary bar chart (Figure 3A), which assessed the contribution of features to the model using average SHAP values. These values are displayed in descending order: Type of vascular invasion by the lesion, Hb, and PLT were the top three most important features in the prediction model. Additionally, the SHAP summary dot plot (Figure 3B) intuitively shows the direction and magnitude of the influence of each feature on the prediction of the model: For example, features such as greater vascular invasion by the lesion, larger lesion size, higher degree of biliary invasion, and lower levels of Hb and PLT significantly increased the likelihood of a patient undergoing ELRA. Local explanations, by calculating and displaying the contribution of each feature to the prediction of a single sample, helped us understand the decision-making mechanism of the model. The SHAP waterfall chart (Figure 3C) shows the contribution of each feature to the model prediction for the 10th HAE patient. The specific values of each feature and their corresponding SHAP values in the chart indicated the positive and negative impacts of these features on the prediction result. For instance, the lesion simultaneously invaded three vessels, namely the hepatic vein, portal vein, and inferior vena cava (IVC), and the lesion size within the range of 15-20 cm made significant positive contributions of +0.131 and +0.094 to the prediction results. PT of 11.7 and PLT of 198 both made negative contributions to the prediction results. Other features also had varying degrees of impact. By accumulating SHAP values, the waterfall chart intuitively demonstrates the prediction decision-making process for an individual patient, revealing the positive and negative contributions of each feature to the specific prediction result, thereby enhancing the transparency and clinical usability of the model. Apart from the degree of bile duct invasion, features such as the type of vascular invasion, PLT, and PT show significant sensitivity. The wide range of fluctuation and strong influence of the SHAP values for these features indicate their high sensitivity to the prediction of the model, with the type of vascular invasion being the most pronounced. Analysis using Beeswarm Plot revealed that for lesions without vascular invasion or those with only portal vein invasion, liver resection is recommended. In contrast, for lesions that invade the hepatic vein, IVC, or multiple vessels, we tend to opt for ELRA.
Figure 3 SHapley Additive exPlanations analysis for model interpretability.
A: SHapley Additive exPlanations (SHAP) summary bar chart, displaying the average SHAP values of features in descending order; B: SHAP summary dot plot, showing the direction and magnitude of each feature's influence on the model's prediction; C: SHAP waterfall chart for the 10th hepatic alveolar echinococcosis patient, illustrating the contribution of each feature to the prediction result. SHAP: SHapley Additive exPlanations; XGBoost: EXtreme Gradient Boosting; WBC: White blood cell count; PLT: Platelet count; Hb: Hemoglobin; IBil: Indirect bilirubin; GGT: γ-glutamyl transferase; Scr: Serum creatinine; PT: Prothrombin time.
DISCUSSION
AE is a highly disabling and lethal zoonotic disease with a significant global burden. In 2010, the number of deaths worldwide due to AE reached 7771, with an annual disease burden exceeding 688000 disability-adjusted life years. China is the country with the highest number of AE patients globally, accounting for over 90% of the total[1,20]. Due to the extensive nature of AE lesions, only approximately 35% of patients are eligible for radical hepatectomy, while liver transplantation is limited by organ shortages. Untreated or improperly treated AE patients have a mortality rate of ≥ 90% within 10-15 years after diagnosis[1,3-6]. Therefore, the demand for new treatment modalities in China is greater than in other countries and regions. Since our center first implemented ELRA in 2010, this technique has brought new hope to end-stage HAE patients and has been rapidly developed and applied. Currently, several medical centers have applied ELRA in the treatment of end-stage HAE[21-23]. Although after more than a decade of relentless efforts by numerous medical and disease control personnel, the prevalence trend of echinococcosis in China has continuously declined, with significant reductions by 2022[20,24,25]. However, with globalization and the expansion of human activities, the epidemic range of AE shows a tendency to expand, with prevention and treatment still facing enormous challenges, and HAE treatment modalities (hepatectomy and ELRA) likely to see broader application[20,26].
The choice of surgical approach is a complex and crucial decision-making process that requires a comprehensive evaluation of the patient's baseline condition, disease severity, and financial considerations to ensure the safety and efficacy of treatment. Traditional decision-making models rely on the clinical experience and personal judgment of physicians, which have the following limitations: (1) Strong subjectivity, susceptible to individual capability and bias; (2) Limited evidence, with some decisions lacking clear evidence and relying on summaries of personal or small-scale cases; and (3) Inability to handle big data, as human capacity is limited when facing complex, high-dimensional patient data, making comprehensive analysis difficult[9,10]. The advantage of machine learning lies in its ability to analyze multimodal big data, integrating and analyzing large numbers of cases and genetic, imaging, and other multimodal data to uncover hidden patterns. It relies on algorithms and data, reducing human subjective bias, enhancing diagnostic consistency, and enabling precise matching of patient characteristics with treatment plans to achieve personalized medicine tailored to each individual[9-12]. Machine-learning technology can integrate multimodal big data, reduce human bias, and realize precision medicine. Although a large number of features may provide richer information for predictive models, too many features can limit the clinical application of the model, and the inclusion of noncausal features may reduce predictive accuracy. Therefore, we adopted a conservative "better safe than sorry" strategy, taking the intersection of the results from RFE, mRMR, and LASSO regularization regression to ensure that the variables ultimately included made significant contributions to the decision-making of surgical approaches[14-17]. The 10 core variables selected through cross-validation in this study-maximum lesion diameter, type of vascular invasion, degree of bile duct involvement, WBC, PLT, Hb, IBil, Scr, GGT, and PT-are all closely related to clinical decision-making and prognosis in HAE. Among these, vascular invasion is already recognized as an independent risk factor for poor long-term prognosis after liver resection, and accurate preoperative assessment can aid in developing personalized surgical strategies[27]. Hb and PLT levels not only reflect the patient's baseline hematopoietic and coagulation functions but also indirectly determine surgical outcomes by affecting postoperative liver regeneration and the risk of thrombosis. Furthermore, both are key components of the Hb-ALB-Lymphocyte-PLT composite index, which assists in implementing personalized risk assessments[28,29]. The remaining variables are closely associated with the inflammatory response in HAE, liver functional reserve, and the risk of perioperative complications. These features not only provide critical information for machine learning models but are also tightly coupled with preoperative decision-making, perioperative management, and prognosis evaluation in HAE.
In the analysis of baseline characteristics in Table 1, although indicators such as gender (P = 0.074) and history of previous liver resection (P = 0.438) are close to the significance threshold, these potential trends still hold exploratory value considering clinical heterogeneity and sample characteristics. Subsequent studies could further elucidate the clinical significance of these features by either increasing the sample size or designing more refined stratified analyses.
In our validation set, the XGBoost model achieved an AUC of 0.734, demonstrating robust discriminatory power for surgical type decision-making. A misjudgment in the surgical plan can lead to inappropriate surgical methods for the patient and delay the optimal intervention time. With its high predictive accuracy, the model can assist clinicians in making more precise choices between hepatectomy and ELRA, avoiding unnecessary trauma, maximizing the preservation of residual liver volume, thereby reducing postoperative complications and mortality risk, and enhancing overall efficacy.
Machine learning is often criticized for its black box problem; mainly because some machine-learning decision-making processes are opaque and lack clinical interpretability, making it difficult for models to gain clinical trust[11,30]. SHAP is a machine-learning interpretability method based on game theory, proposed by Lundberg and Lee[19] in 2017 and has become the gold standard for explaining complex model predictions. Its core idea originates from the Shapley value in cooperative game theory, which quantifies the contribution of each feature to model predictions to achieve transparent interpretation of black box models[19]. We used the SHAP method to achieve global explanations at the feature level and local explanations at the individual level, addressing the black box issue of machine-learning models. According to the SHAP values, we found that the type of vascular invasion by the lesion is the most important feature in the prediction model and the key factor influencing the choice of surgical approach. Vascular invasion is an important biological characteristic of liver malignancies (especially hepatocellular carcinoma), and its extent and type directly affect the selection of treatment strategies and prognosis[31,32]. For the two main radical surgical approaches, hepatectomy and liver transplantation, vascular invasion plays a decisive role in the decision-making process. Invasion of the main portal vein or its first-order branches by liver tumors is usually considered a contraindication for hepatectomy. In comparison, the criteria for liver transplantation are stricter, with multiple liver cancer treatment guidelines considering vascular invasion as a contraindication for liver transplantation. However, for patients with microvascular invasion, some criteria (e.g., the Hangzhou criteria) may still consider liver transplantation in combination with other factors (such as a-fetoprotein levels and histological grading), but postoperative adjustment of the immunosuppression regimen and combination with adjuvant therapy are required to reduce the risk of recurrence[32,33]. ELRA, as the ultimate option for unresectable lesions due to large tumor size, extensive invasion, or proximity to central vessels and/or bile ducts, involves a series of key vascular-related technologies, such as liver blood flow occlusion, extracorporeal venous bypass, vascular resection and reconstruction, reperfusion, and reimplantation. Compared with liver transplantation, the degree of vascular invasion is more critical for ELRA[34-36]. ELRA is still not satisfactory for treating cancer, and it remains controversial whether patients can benefit from it. However, HAE, a benign disease with infiltrative growth, known as parasitic cancer, is undoubtedly more suitable for ELRA, and its therapeutic value has been widely recognized[37,38]. Vascular invasion is indispensable in the treatment decision-making of end-stage HAE and occurs throughout the treatment process. First, the surgical approach must be assessed based on the condition of the patient's lesion. Previous research from our center proposed that ELRA could be considered when more than two hepatic veins are involved, the invasion extends to tertiary hilar branches, the length of IVC involvement behind the liver is ≥ 1.5 cm, and the circumference is ≥ 120°[21]. The team from West China Hospital has proposed stricter criteria: (1) Simultaneous invasion of at least two important hepatic portal structures; (2) IVC invasion with length > 3.0 cm and circumference > 180°; and (3) IVC invasion length < 3.0 cm but upper boundary reaching the pericardial level[22,39]. Our SHAP analysis suggests that for lesions without vascular invasion or those with only portal vein invasion, liver resection is recommended. Conversely, for lesions that invade the hepatic vein, IVC, or multiple vessels, we tend to favor ELRA. We align with the West China team in emphasizing the importance of invasion of the IVC and multiple vascular branches. Our research criteria are relatively straightforward, with a more detailed evaluation of the type of vascular invasion, but in terms of specifically quantifying the extent of invasion, they are not as meticulous as those established by West China Hospital. However, the standards set by the West China team lack an assessment of hepatic vein invasion. Our limitations mainly arose from the lack of imaging data, which restricts our ability to quantify the extent of lesion invasion. In the future, we will conduct prospective studies that integrate imaging, AI, and three-dimensional reconstruction technologies to develop a precise and personalized surgical decision-support system. Although the criteria vary among centers, all regard the degree of lesion involvement of important vessels as a core indicator for surgical decision-making.
The degree of vascular invasion not only affects the choice of surgical approach but also determines the specific strategies for vascular resection, repair, and reconstruction during surgery[40-42]. Our center has proposed a graded management plan for IVC reconstruction based on the degree of defect: (1) If the defect in the IVC wall after radical resection is < 120° of the luminal circumference, direct suturing can be performed; (2) If the defect is 120°-180°, patch repair with autologous vascular graft is the best choice; (3) If the defect exceeds 180° and autologous patch repair is difficult, the retrohepatic IVC should be replaced, but strict anticoagulation is required postoperatively; and (4) When the IVC is completely obstructed and collateral circulation has been fully established, retrohepatic IVC resection without reconstruction can be performed[43]. The West China team selects appropriate vascular control techniques based on the site of invasion to reduce bleeding and maintain hemodynamic stability during surgery[40]. These clinical practices fully demonstrate the key role of vascular invasion in HAE treatment decision-making. We believe that the machine-learning model can be deeply integrated with 3D printing technology for personalized vascular models in the future. By using machine learning to predict surgical plans and using 3D printing technology to accurately simulate the patient's vascular anatomy, precise preoperative planning and personalized vascular custom printing during surgery can be achieved. This approach enhances surgical accuracy and assists surgeons in preoperative simulation and risk assessment, thereby optimizing personalized treatment strategies.
Although this study provides important insights, several limitations require attention. Firstly, as a single-center retrospective study, the sample source is limited, making it difficult to avoid selection bias, and the patient characteristics may not comprehensively reflect the heterogeneity of HAE across different regions and healthcare systems. To improve the external validity and generalization of the model, future work should focus on conducting multicenter, large-sample prospective validations. By integrating data from diverse geographic regions and various levels of healthcare institutions, a systematic assessment of the robustness and clinical applicability of the model can be achieved. Currently, our center plans to initiate the construction of a specialized collaborative network for HAE with multiple medical institutions, and relevant multicenter validation studies will be systematically advanced in the future.
CONCLUSION
This study leveraged machine learning to enhance surgical decision-making for HAE. Among 710 patients (545 hepatectomy, 165 ELRA), key features influencing surgical choices were identified via multiple selection methods (RFE, mRMR, LASSO), converging on ten critical factors including lesion size, vascular and biliary invasion types, WBC, PLT, Hb, IBil, Scr, GGT, and PT. The optimized XGBoost model achieved an AUC of 0.935 in training and 0.734 in validation. SHAP analysis highlighted vascular invasion type as the most influential factor, with hepatic vein/IVC/multi-vessel involvement favoring ELRA over hepatectomy. This transparent, data-driven approach offers robust support for personalized surgical strategies in HAE.
Footnotes
Provenance and peer review: Unsolicited article; Externally peer reviewed.
Peer-review model: Single blind
Specialty type: Gastroenterology and hepatology
Country of origin: China
Peer-review report’s classification
Scientific Quality: Grade A, Grade A, Grade A
Novelty: Grade A, Grade A, Grade B
Creativity or Innovation: Grade A, Grade B, Grade B
Scientific Significance: Grade A, Grade A, Grade B
P-Reviewer: Denisov A, MD, PhD, Additional Professor, Researcher, Spain; Khajavian M, PhD, Academic Fellow, Postdoctoral Fellow, Malaysia; Zhao JP, MD, PhD, Associate Chief Physician, Associate Professor, China S-Editor: Li L L-Editor: A P-Editor: Yu HG
Torgerson PR, Schweiger A, Deplazes P, Pohar M, Reichen J, Ammann RW, Tarr PE, Halkic N, Müllhaupt B. Alveolar echinococcosis: from a deadly disease to a well-controlled infection. Relative survival and economic analysis in Switzerland over the last 35 years.J Hepatol. 2008;49:72-77.
[RCA] [PubMed] [DOI] [Full Text][Cited by in Crossref: 175][Cited by in RCA: 197][Article Influence: 11.6][Reference Citation Analysis (0)]
Wen H, Dong JH, Zhang JH, Zhao JM, Shao YM, Duan WD, Liang YR, Ji XW, Tai QW, Aji T, Li T. Ex vivo liver resection followed by autotransplantation for end-stage hepatic alveolar echinococcosis.Chin Med J (Engl). 2011;124:2813-2817.
[PubMed] [DOI] [Full Text]
Guyon I, Weston J, Barnhill S, Vapnik V. Gene Selection for Cancer Classification using Support Vector Machines.Mach Learn. 2002;46:389-422.
[PubMed] [DOI] [Full Text]
Lundberg SM, Lee S.
A unified approach to interpreting model predictions. Proceedings of the 31st International Conference on Neural Information Processing Systems (NIPS'17); 2017; Long Beach, CA, United States. Red Hook, NY, United States: Curran Associates Inc., 2017: 4768–4777.
[PubMed] [DOI]
Aji T, Dong JH, Shao YM, Zhao JM, Li T, Tuxun T, Shalayiadang P, Ran B, Jiang TM, Zhang RQ, He YB, Huang JF, Wen H. Ex vivo liver resection and autotransplantation as alternative to allotransplantation for end-stage hepatic alveolar echinococcosis.J Hepatol. 2018;69:1037-1046.
[RCA] [PubMed] [DOI] [Full Text][Cited by in Crossref: 59][Cited by in RCA: 79][Article Influence: 11.3][Reference Citation Analysis (0)]
Yang GJ, Ouyang HQ, Zhao ZY, Li WH, Fall IS, Djirmay AG, Zhou XN. Discrepancies in neglected tropical diseases burden estimates in China: comparative study of real-world data and Global Burden of Disease 2021 data (2004-2020).BMJ. 2025;388:e080969.
[RCA] [PubMed] [DOI] [Full Text][Cited by in RCA: 6][Reference Citation Analysis (0)]
Wang X, Wang P, Tang B, Xu J, Wang B, Gu L, Liang Y, Guo H, Liu H, Wu Y, Wang H, Zhou Y, Zeng Y, Diao Y, Yao L, Wang M, Li C, Pawlik TM, Shen F, Cai L, Yang T. Platelet count as a double-edged sword: The impact of thrombocytosis and thrombocytopenia on long-term outcomes after hepatic resection for hepatocellular carcinoma.Biosci Trends. 2025;19:351-360.
[RCA] [PubMed] [DOI] [Full Text][Cited by in RCA: 1][Reference Citation Analysis (0)]
Hashimoto K, Kanno H, Sakai H, Miyazaki D, Fukutomi S, Arai S, Goto Y, Akashi M, Isobe T, Mori N, Ishibashi N, Fujita F, Hisaka T. Hemoglobin, Albumin, Lymphocyte, and Platelet (HALP) Score As a Predictor of Prognosis After Hepatectomy.Anticancer Res. 2024;44:3669-3678.
[RCA] [PubMed] [DOI] [Full Text][Cited by in RCA: 2][Reference Citation Analysis (0)]
Sichuan Clinical Research Center for Echinococcosis; Echinococcosis Committee of Sichuan Medical Doctor Association. [Expert consensus on diagnosis and treatment of hepatic alveolar echinococcosis (2020 edition)].Zhongguo Puwai Jichu Yu Linchuang Zazhi. 2020;27:13-17.
[PubMed] [DOI] [Full Text]
