Liver cancer, which is a serious disease that greatly endangers human health, has a variety of treatment methods. The operative trends for HCC demonstrate that the proportion of cases performed laparoscopic hepatectomy is increasing for the advantage of minimally invasive compared to open surgery15–17. LH and MWA are both curative minimally invasive treatment methods, which occupy a major position in the treatment of tumors, especially for primary HCC18–21. Previous studies showed that thermal ablation is a more minimally invasive, safer and more cost-effective therapeutic approach for HCC compared with LH10,22,23. And for patients who cannot perform laparoscopic hepatectomy, such as multiple tumors located in multiple hepatic segments, tumors that cannot be resected due to visual field limitation (such as the upper right posterior segment), and patients with extremely poor coagulation and liver function, thermal ablation is a good option. Appropriate treatment methods should be able to prolong the survival time of patients while minimizing the pain and financial burden of the treatment process. The problem of choosing a better treatment in a variety of cases has always been an urgent problem to be solved in clinical work.
In this study, we focused on the 3–5 cm primary HCC patients with VER after receiving LH or MWA treatment. We compared the OS and CSS between the patients with VER and those without VER respectively in LH group and MWA group, and found that the occurrence of VER will significantly reduce the survival time of patients (p < 0.0001). Therefore, regardless of LH or MWA, benefit for patients with VER from the initial treatment will discount largely.
In the comparison of OS and CSS of all patients between LH group and MWA group, the survival time of LH group is significant better (p < 0.05). However, the advantage disappeared when come to the comparison outcome in patients with VER between the two groups, regardless before or after PSM (p > 0.05). Therefore, the difference of survival between the two treatment methods was mainly caused by patients without VER. And compared with MWA group, LH group has more significant difference in survival time between patients with and without VER, which also makes it more important for LH group to correctly identify the occurrence of VER before operation. Meanwhile, previous studies have also shown that MWA has advantages over LH in terms of safety, minimally invasive and cost performance10,22,23. Therefore, for 3-5cm HCC patients with VER after LH treatment, if they can be correctly identified before treatment and take MWA instead, the treatment pain and economic burden of patients can be reduced while maintaining OS and CSS.
The rapid development of prediction models makes it possible to predict the outcome of treatment methods before operation24. With the increasingly close combination of artificial intelligence and clinical care, the method of establishing prediction model based on machine learning has gradually replaced the traditional method based on logistics regression. Machine learning has a good performance in improving the prediction ability of the model, dealing with missing values, and correcting the deviation of positive and negative sample sizes25. The use of machine learning algorithms can significantly improve predictive ability26; The 4 techniques used to model these outcomes covered a broad range of artificial intelligence methods. Of the techniques used, the GBM model performed the best, followed by the kernel XGB model. GBM algorithm is a kind of boostboosting algorithm, in which the pseudo-residuals are calculated according to the initial model, and then a basic learner is established to explain the pseudo-residuals. After that, the basic learner is multiplied by the weight coefficient (learning rate) and the original model for linear combination to form a new model. This iteration will result in a model that minimizes the expected value of the loss function27. GBM has the advantages of excellent predictive ability, being able to fit complex nonlinear relationships, deal with missing values, and effectively control overfitting. In the comparison of various machine learning models in this study, the model established based on GBM has no obvious overfitting in the training set, but has the highest AUC value in the verification set (AUC = 0.722). Moreover, its advantage of being able to deal with missing values is also beneficial for the application of the model.
In our study, the baseline indicators(AFP, tumor number, ALT, total bilirubin and direct bilirubin)were different in patients with and without VER in LH group, while ALT, total bilirubin and direct bilirubin had higher contributions in the GBM model. At the same time, we found that patients with the earliest onset of VER in the LH group had a RFS of only 32 days during the follow-up. According to the phenomena above, we believe that the occurrence of VER may be due to the existing scattered lesions in the liver that could not be detected by either preoperative imaging or intraoperative surgeons, but continued to grow after treatment until were found in postoperative imaging review. Before LH treatment, multiple tumors may indicate a higher possibility of undetected scattered lesions in addition to the visible tumors. The increasement of AFP, ALT, TB and DB level indicates worse liver function, which may indicate the presence of potential invisible lesions in patients serologically.
Some limitations do exist in our study. First of all, this study is a retrospective study, all associated bias risks existed. Although cases from several hospitals were included, preoperative and postoperative image and laboratory examination equipment, experience of doctors and equipment used in the treatment of patients were not strictly controlled, which may lead to deviation of the research results. Secondly, this study only included the preoperative text information of patients, failed to give full play to the advantages of artificial intelligence in image omics and multi-dimensional data processing due to the lack of image data. Therefore, there is still room for improvement in the prediction power of the model. And this part of the study will be carried out in the future. Finally, this study failed to balance the interference of subsequent treatment that may affect the patients' OS and CSS after the occurrence of VER. This may affect the final study results.
We conclude that, the survival of patients with primary HCC 3-5cm in diameter who developed VER after receiving LH was comparable to that of the patients receiving MWA. Therefore, such patients can receive more minimally invasive and cost-effective MWA as the initial treatment instead. At the same time, a model based on patients' preoperative text information was established to predict whether VER would occur after receiving LH treatment.