TO THE EDITOR
We are deeply interested in the paper entitled “Identification of key factors and explainability analysis for surgical decision-making in hepatic alveolar echinococcosis assisted by machine learning” by Zhu et al[1], published in the World Journal of Gastroenterology. This study innovatively integrates multiple machine learning methods with feature selection strategies to systematically identify the key factors that influence the choice of surgical approach for hepatic alveolar echinococcosis (HAE). SHapley Additive exPlanations (SHAP) revealed that vascular invasion type is a critical determinant in selecting between hepatectomy and ex vivo liver resection and auto-transplantation (ELRA). Additionally, factors such as low hemoglobin levels, low platelet counts, and large lesion sizes significantly increase the likelihood of patients undergoing ELRA. This study demonstrates notable methodological innovation by introducing artificial intelligence (AI) into preoperative decision-making for HAE, enhancing predictive accuracy through the integration of multiple machine learning approaches. Particularly noteworthy is the adoption of three feature selection techniques: Recursive feature elimination, minimum redundancy maximum relevance, and least absolute shrinkage and selection. These methods effectively resolve multicollinearity issues, significantly enhancing the robustness and interpretability of the model. However, it is debatable whether the model relies entirely on structured data, which often present limitations such as missing values, measurement errors, and difficulty in capturing complex nonlinear relationships. For example, simplifying vascular invasion to the number of invaded vessels types, representing lesion size with a single diameter measurement, such overly simplified quantification methods may fail to accurately reflect the true pathological complexity, thereby compromising the reliability of the model’s predictions.
HAE is a multi-chambered cystic structure exhibiting invasive growth characteristics. Its biological behavior resembles that of malignant tumors[2]. Radical surgical resection remains the only potentially curative approach[3]. Treatment decision-making involves a complex multidisciplinary process requires a comprehensive evaluation of three critical factors: Is complete resection achievable? Is the residual liver volume sufficient to maintain function? Can the patient’s systemic condition tolerate surgery? Furthermore, due to the frequent involvement of major hepatic vessels and bile ducts, HAE lesions present greater surgical challenges than primary hepatic malignancies, with a higher incidence of postoperative complications (35%-59.7%), primarily including hepatic failure, hemorrhage, and bile leakage[4,5]. Particularly in extensive liver resections, the incidence of postoperative hepatic failure can reach as high as 58.22%[6]. Therefore, accurate preoperative assessment of the anatomical features of HAE lesions and their adjacent structures is crucial in selecting an appropriate surgical approach and thus reducing the risk of complications. The lack of a precise preoperative assessment is among the major factors contributing to the high complication rates and increased postoperative mortality[7].
A detailed assessment of vascular involvement serves as the core basis for judging the resectability of HAE and formulating surgical strategies[8]. However, when quantifying this critical variable, Zhu et al[1] employed a simplified ordinal classification system, categorizing the involvement of hepatic veins, portal veins, and the inferior vena cava as “no abnormality”, “invaded only one type”, “invaded two types” and “invaded three types”. While this approach facilitates model calculations, it sacrifices the morphological information critical for preoperative decision-making. In clinical practice, vascular contact, direct invasion, and secondary thrombosis constitute the three key dimensions for assessing the extent and nature of vascular involvement[9]. Specifically, the degree to which vessels are encircled by lesions directly affects the feasibility of safe vessel dissection. The longitudinal extent of vascular invasion determines whether vascular reconstruction should be performed via direct anastomosis or if vascular grafting is needed. This information is indispensable in preoperative assessments. The number of affected vessel types alone is insufficient for a comprehensive evaluation of surgical feasibility. Furthermore, the presence of collateral circulation is also critical in preoperative assessments. As a chronic progressive disease, HAE can have a latency period of up to 10 years, with vascular involvement typically accompanied by the formation of collateral circulation. Patients with portal vein involvement and established collateral circulation face a high risk of uncontrollable hemorrhage during hepatectomy; conversely, ELRA can effectively control bleeding and achieve radical resection[2]. However, these critical insights remain underrepresented in the current modeling variables. Contrast-enhanced computed tomography (CT) serves as the primary imaging modality for assessing hepatic and peripheral vascular involvement, with its core value lying in multiphase dynamic evaluation (including arterial, portal venous, and equilibrium phases) to provide a comprehensive, layered assessment of the hepatic artery, portal vein, hepatic veins, and inferior vena cava. Therefore, for HAE patients, this imaging modality not only clearly delineates the number of involved vessels, the extent of longitudinal invasion, and their anatomical proximity to the lesion but also enables precise assessment of collateral circulation formation and compensatory status in advanced disease, these data provide critical therapeutic evidence and decision support for developing individualized surgical plans.
HAE typically manifests as a single, irregularly shaped, large cystic-solid mass. In the study of Zhu et al[1], lesion size was simplified into categorical variables such as < 5 cm, 5-10 cm and ≥ 10 cm. Although this grouping provides a quick, intuitive preliminary impression of the lesion and facilitates rapid stratification, it fails to accurately reflect the true three-dimensional burden of the lesion. Particularly for irregularly shaped lesions, such measurements are insufficient to guide individualized surgical planning. Furthermore, visual measurements of lesion size, which are dependent on surgeon experience, also introduce subjective errors and uncertainties[10]. In fact, the volume of a HAE lesion is a key determinant of the extent of hepatic resection during surgery and the postoperative functional liver reserve (FLR). Accurate preoperative assessment of residual liver volume is crucial for planning surgical strategies. However, accurately predicting three-dimensional resection margins and the FLR from single-plane measurements of lesions is difficult. Currently, volumetric measurements based on CT or magnetic resonance imaging (MRI) serve as the gold standard for FLR assessment. Advances in AI technology have enabled the precise segmentation of the entire liver[11] and its internal structures, including lesions and vascular systems[12,13]. Automated methods can yield high-precision measurements of HAE lesion volume, total liver volume, and FLR. These measurements demonstrate high consistency and objectivity and are unaffected by operator or time variations.
The formulation of HAE surgical strategies relies not only on the precise assessment of the lesion and liver morphology but also on the comprehensive analysis of liver function. These two aspects are complementary and indispensable. In addition to precise FLR measurement, liver function assessment must fully account for persistent structural and functional damage caused by the HAE lesion itself and its associated pathological changes, including displacement effects, formation of inflammatory granulation tissue, and fibrosis. Liver volume, texture, and blood supply constitute the core elements of functional status evaluations. Currently, commonly used clinical methods for assessing FLR include blood-based liver volume measurement and scoring systems, such as the Model for End-stage Liver Disease score, Child-Turcotte-Pugh score, albumin-bilirubin classification and the indocyanine green clearance test (ICG-R15). These scoring systems were originally designed to assess liver function in patients with cirrhosis. However, the mechanism of liver injury caused by HAE differs significantly from that seen in typical viral or alcoholic cirrhosis. Liver damage in HAE is caused by parasite-driven granulomatous inflammation and focal fibrosis, resulting in an infiltrative, space-occupying growth pattern akin to that of malignancy. This frequently results in functional loss within the affected area, with compensatory hyperplasia occurring in the unaffected hepatic lobes. In contrast, viral or alcoholic cirrhosis manifests as diffuse, panhepatic injury with overall functional decline. Therefore, for the pathological physiological pattern of HAE, which is characterised by local liver dysfunction and compensatory function in the unaffected side, the scores obtained through tools for evaluating overall liver function neither accurately reflect the selective dysfunction caused by HAE nor reliably predict the functional reserve of the remaining liver parenchyma after lesion resection. Consequently, these scores have certain limitations in the preoperative assessment of HAE, particularly in predicting postoperative compensatory capacity. ICG-R15 has long been regarded as one of the gold standards for liver function assessment, yet its application has some limitations[14]. The degree of liver fibrosis is a core factor influencing liver tissue hardening. Currently, elastography techniques based on MRI and ultrasound have become the preferred noninvasive imaging modalities for clinically assessing liver fibrosis, both of which have demonstrated good performance in diagnosing fibrosis with different etiologies. However, ultrasound elastography is susceptible to operator variability and artifacts, with diagnostic performance [area under the curve (AUC) = 0.75-0.93][15] typically inferior to magnetic resonance elastography (AUC = 0.91-0.97)[16]. Enhanced CT or MRI not only clearly visualizes vascular involvement and collateral circulation formation in HAE but also objectively reflects liver physiological function through blood flow perfusion and contrast agent metabolism dynamics. Studies show that machine learning models based on enhanced CT radiomics achieve liver function prediction accuracy (C-index = 0.773) approaching that of models based on biochemical markers (C-index = 0.771)[17]. Furthermore, machine learning models integrating enhanced CT and MRI radiomics data effectively assess hepatic reserve capacity and classify ICG-R15[18]. Thus, integrating these multimodal data enables noninvasive, quantitative, and comprehensive precision assessment spanning from “morphological structure” to “functional status”, providing more comprehensive and objective evidence for preoperative planning and treatment decisions in HAE patients.
The integration of medical imaging and machine learning technologies offers new opportunities in liver disease management, with high potential for improving liver fibrosis staging, portal hypertension assessment, liver tumour characterization, and prognostication[10,19,20]. In HAE studies, machine learning-based segmentation algorithms can accurately quantify the lesion volume, liver volume, and FLR. Furthermore, magnetic resonance elastrography (MRE) enables the noninvasive assessment of hepatic parenchymal fibrosis. When combined with dynamic contrast-enhanced CT or MRI, MRE allows the comprehensive evaluation of vascular invasion, collateral circulation establishment, and hepatic reserve function, gradually establishing a new “morphology-function” integrated preoperative assessment paradigm. To strengthen the linkage between imaging, functional assessment, and surgical decision-making, a multimodal data intelligent integration system must be established. After standardized image acquisition and preprocessing, this system employs deep learning segmentation models to achieve automatic segmentation of the entire liver, lesions, and vascular structures. It then extracts morphological, textural, and spatial relationship features. Concurrently, quantitative histograms and spatial distribution analyses are performed on perfusion and physiological parameters. Furthermore, functional reserve indicators are either directly computed or derived through constructed metrics. All the features are uniformly encoded into structured mathematical vectors and then integrated via fusion architectures such as multibranch deep learning networks. Importantly, all the aforementioned key technologies have now shown feasibility for clinical translation. This enables the model to simultaneously analyse the anatomical invasiveness and local microenvironment alterations of the lesion and the overall compensatory potential of the liver. Ultimately, model outputs are directly coupled with critical decision points such as surgical procedure selection, driving a comprehensive evolution of surgical decision-making from experience-dependent approaches towards quantitative, personalized prediction.
In summary, Zhu et al[1] provided important insights for surgical decision- making in HAE and pioneered a new pathway for AI-based auxiliary assessment. Their work is characterized by rigorous methodology. In this letter to the editor, we discussed some of the limitations inherent in models constructed solely from structured data. First, the quantification of some variables is overly simplified, which compromises the reliability and stability of the predictive model. Second, the model fails to incorporate critical details essential for surgical assessment, such as the anatomical relationship between lesions and blood vessels. Third, the included variables do not effectively evaluate preoperative liver function. Additionally, some variables carry risks of subjective judgment errors, and their reproducibility requires further validation. Finally, Zhu et al[1] employed SHAP for model interpretation as another highlight of this study. Further exploration of the deeper interpretability challenges that arise when such an excellent predictive model is deployed in real clinical decision-making scenarios is warranted. As the authors explain in the discussion section, to mitigate the potential reduction in predictive accuracy due to noncausal features, they adopted a conservative “better safe than sorry” strategy. Consequently, the variables identified by the final predictive model represent statistical associations rather than causal relationships. Features with high SHAP values cannot serve as direct targets for clinical intervention, potentially leading to biased decision-making. Therefore, we advocate for the following approaches in scenarios involving information-rich, large-scale multimodal data: Employing causal analysis to identify variables with genuine causal relationships for model building and enhancing SHAP analysis through counterfactual explanations, partial dependency graphs, and case-based reasoning to bolster clinical credibility and practical utility.
Future endeavours should strive to advance beyond the limitations of models reliant on simplified, structured data by actively integrating multidimensional clinical and imaging information. This paradigm shift will enable the systematic analysis of complex variable interactions and nonlinear relationships, paving the way for a comprehensive and precise personalized management system for HAE that encompasses diagnosis, surgical selection, treatment, and prognosis. Ultimately, such an approach will advance precision medicine tailored to specific patient characteristics.
