HCCA accounts for 50-60% of bile duct carcinomas characterized by poor prognosis, and surgery represents the optimal treatment approach15. Given the complexity of the surgical procedures of HCCA and the lack of current accurate prognostic models, accurate prediction of postoperative survival outcomes is important16. To address this need, we conducted retrospective studies using the ML algorithm to identify prognostic risk factors for patients with HCCA after curative resection. Additionally, predictive nomograms were developed based on the findings.
In this study, we discovered four statistically significant and most prominent risk factors with positive margins, lymph node metastasis, low TLNC, and poor tumor differentiation via analysis using five mL algorithms. Additionally, COX regression was employed to screen for prognostic risk factors. Previous studies have shown that ML is superior to Cox's regression in extracting prognosis-related risk factors10, 12, 13. However, our study shows minimal differences between ML and COX regression in identifying prognostic risk factors. This is due to four important prognostic indicators identified as being significantly more important than other variables with a significant importance gap. The ML algorithm allows for simultaneous screening of all variables and detecting nonlinear interactions, clearly indicating the importance of variables that traditional Cox regressions may overlook. Furthermore, unlike classic COX regression, the ML algorithm provides robust feature ranking and cross-validation so that nomograms contain the most impactful prognostic factors. In summary, ML excels in visualization, variable screening, identification of nonlinear interactions, ranking features, and cross-validation, ensuring the effectiveness of nomograms.
Identification of these four factors highlights the complexity and multifaceted nature of the prognosis caused by HCCA, highlighting the important role of surgical quality (resectomy status and degree of lymph nodeectomy) and tumor biology (tumor differentiation and involvement of lymph nodes). Previous studies have demonstrated that lymph node invasion, tumor differentiation, and marginal status are important prognostic indicators affecting clinical outcomes in patients with HCCA.17. This finding is consistent with our findings.
Positive margins predicting a decline in survival highlight the essential need for comprehensive preoperative assessment and in-depth surgical techniques to achieve complete tumor resection. The study found that positive margin status in patients with biliary malignant tumors can be further classified into their in situ and invasive cancers.18. Furthermore, many studies have shown that classifying positive margins (R1 resection) on the duct margin and radial margins is useful for more accurate prognostic layer degradation and patient selection of adjuvant therapy.19, 20, 21. Therefore, it makes sense to classify the state of positive surgical margins in detail.
In particular, recent studies have demonstrated that the effects of positive margins on OS in postoperative patients may be closely related to lymph node metastatic status.22,23. Specifically, in the presence of lymph node metastasis, margin status does not appear to affect survival outcomes. Another study has shown that lymph node metastasis primarily determines prognosis regardless of margin statustwenty four. This contrasts with findings that may be attributed to the limitations of single-centered study populations. However, Koca F. et al.twenty five We found that patients with positive margins had a worse prognosis after HCCA resection, regardless of lymph node metastasis. Therefore, future studies should further investigate the effects of margin status on survival outcomes in patients with lymph node metastasis.
Furthermore, based on the degree of tumor invasion, surgical resection of HCCA includes resection of the diseased bile duct and major liver resection. Previous studies have shown that the caudate lobe is frequently involved in 40-98% of HCCA cases, thus suggesting a combination of caudate lobe resection to achieve radical resection surgery.26. Furthermore, it has been consistently demonstrated that caudate lobectomy as a radical resection strategy for HCCA does not significantly increase postoperative morbidity and mortality.27,28,29,30. Given the high incidence of caudate lobe involvement in HCCA and the association between positive margins and reduced survival, our findings including caudate lobe resection in healing surgery to achieve negative margins, wherever it is safe and feasible, are supported.
Low survival outcomes associated with lymph node metastasis and low TLNC highlight the importance of proper lymph node resection and comprehensive staging in HCCA management. Previous studies have shown significant differences in survival rates showing ≥13 and <13 TLNCs among HCCA patients, indicating that the latter group has significantly worse prognosis.31. Recent systematic analyses demonstrated that 7 TLNCs or higher are sufficient for prognostic staging, whereas 15 TLNCs or higher do not enhance detection of lymph node-positive patients.32. In our study, we observed that TLNC plays an important role in the survival of patients with HCCA. Patients with TLNCs below 6 showed significantly lower survival rates compared to six patients. In particular, reaching ≥12 TLNC reduces the difference in prognostic shock, suggesting that approximately 12 TLNC may be sufficient to achieve satisfactory surgical outcomes. Currently, routine lymph node dissections include lymph nodes in the hila, lobe muscle, terminal abdomen, cerebral ventral and common hepatic artery areas.33.
In many cases of malignant tumors, poorly differentiated tumors generally show poor prognosis8,34,35. Similarly, this trend has been observed in HCCA. Recent studies have revealed that end-peritoneal deinvasion serves as a factor contributing to poor prognosis in HCCA, with incidence ranging from 56.0 to 88.0% of biliary tumors.36. Interestingly, periperitoneal ablation is more common in moderately and inadequately differentiated tumors37. This finding suggests that lower HCCA differentiation status is more likely to result in perinjury. Inadequately differentiated tumors showing higher metastatic potential further demonstrate their independent prognostic value. Although end-peritoneal invasion is also associated with poor prognosis, the ML algorithm prioritizes tumor differentiation of risk factor extraction. This is likely due to the stronger ability to affect prognosis, as tumor differentiation reflects broader biological aggression, including the tendency to perinvade peritoma stage invasion. In the future, it would be meaningful to conduct a detailed investigation into the association between tumor differentiation and peritumor invasion.
Additionally, ML-based nomograms were employed to stratify patients into three risk groups. This provides a clinically viable framework for postoperative management. In the training set, for patients in the high-risk, low-risk, and low-risk groups, OS was (39.3 ± 1.7) months, (24.1 ± 3.2) months, and (10.3 ± 1.4) months ((10.3 ± 1.4) months).p<0.001). Similarly, in the test set, OS was (43.4±5.4) months, (20.9±2.3) months, and (10.3±1.3) months, respectively (10.3±1.3)p<0.001). These findings highlight the usefulness of nomograms for postoperative risk assessments and allow for customized monitoring and intervention strategies. For patients classified within the low- and middle-risk groups, optimized postoperative care including enhanced follow-up imaging and symptom monitoring is sufficient. Adjuvant therapy administration may also be an important option among patients in the high-risk group. However, given the dependence of nomograms on postoperative variables, its usefulness is limited to postoperative risk stratification and cannot guide preoperative decisions. Future studies integrating preoperative imaging or molecular markers may fill this gap.
The Bismuth-Corlette classification has become the most widely adopted classification and surgical instruction of the HCCA since its proposal in 1975.16. The TNM staging system is the most commonly used traditional method for predicting the prognosis of patients with HCCA. However, TNM staging systems have limited accuracy in prognostic assessments, making it difficult to personalize postoperative patients' evaluations of HCCA38. The performance of ML-based nomograms was compared with established prognostic tools such as the TNM staging system, Bismuth-Corlette classification, and previously reported nomograms. Our results ultimately demonstrated that the performance of the proposed model outweighs the performance of previously established nomograms derived from TNM staging systems, bismuth calllet classification, and similar studies (Table 2). The C index value of the model validates this advantage. Specifically, compared to our studies, these previous studies have a smaller range of variables, relying only on a single analysis method, resulting in smaller actual sample sizes. As a result, these findings demonstrate a favorable improvement in the predictive accuracy and clinical utility of the model over existing methods.
However, as a retrospective study, there may be some limitations. First, sample sizes for HCCA patients after curative resection were limited, and they all came from a single center. A significant limitation, therefore, is the lack of external validation in an independent cohort. This is essential to confirm generalizability. Second, our study included a long period. Unfortunately, we did not include the year of surgery as a variable. During this period, pathological and molecular detection techniques, perioperative care techniques, adjuvant therapy strategies, tracking methods, and the social environment have continued to evolve. This may have resulted in a better prognosis for patients treated in recent years compared to previously treated patients. Third, defects in chemotherapy standardization and integrity can also bias outcomes. Given these limitations, future large-scale, multicenter prospective studies are guaranteed. These studies should employ a wider range of ML algorithms and integrate multiomic data such as emerging biomarkers, preoperative imaging, and pathology to improve the modality of adjuvant therapy. Furthermore, it is essential to further validate the performance of nomograms across diverse populations and treatment regimens, thereby improving prognostic accuracy.
