Interpretable machine learning-based clinical prediction model for predicting lymph node metastasis in patients with intrahepatic cholangiocarcinoma

Intrahepatic cholangiocarcinoma originates from the malignant transformation of the bile ducts epithelium, and represents more aggressive compared to HCC [1], with the 5-year overall survival ranging from 15% to 40% [1, 6]. The incidence of LNM in ICC is much higher than that in HCC [29].Indeed lymph node status is critical for therapy selection and has been identified as one of the most important factors for prognosis [6]. A few of studies demonstrated that lymphadenectomy (LND) improved long-term survival outcome of ICC patients [30, 31], thus, LND should be a routine method for radical resection in ICC [32, 33]. Whereas other studies reported that LND didn’t improve survival outcome of ICC patients , with associated surgery-related complications [34, 35]. It’s reported that approximately 50% of the patients did not dissect lymph node dissection [36], which may result in mis-or under-staging and further compromised their outcomes [32, 36]. For ICC patients, accurate prediction of LNM will facilitate clinical treatment decision-making for the appropriate diagnosis and surgical planning.

Accordingly, we used a novel type of AI-machine learning-to predict LNM in ICC patients. Using ML algorithms, we developed and validated six models to predict LNM in 345 patients with ICC. We found that XGB model (average AUC=0.908) had greatest predictive performance in internal validation. Unlike some nomogram models [14, 19], we further provided dynamic construction. Consequently, based on the XGB model, a web calculator has been established to estimate visually individual probability of LNM and improved the applicability of the model.

In our study, multivariate logistic regression analysis founded that ALD, smoking, boundary, diameter, and WBC were independent predictive factors of LNM in patients with ICC (Table 2). As an independent risk factor, the influence of WBC on prognosis has been reported. Shirono et al [37] found that the serum WBC level was negatively associated with survival time in ICC patients, furthermore illustrated that patients with the WBC level was more than 6800/µL had a short survival time. In this study, we demonstrated that WBC was an independent predictor for the presentation of LNM in ICC patients. We also revealed that the risk of LNM was significantly increased when serum WBC level was more than 7180/µL. According to the permutation importance of variables in Fig. 3, WBC ranks first among the five prediction models and deserves the most attention when predicting LNM. WBCs include monocytes, lymphocytes and neutrophils. Monocytes have roles in promoting tumor invasion and angiogenesis [38]. In addition, tumor-associated macrophages developed from monocytes, can promote tumor lymphangiogenesis by the secretion of pro-lymphangiogenic factors and trans-differentiation into lymphatic endothelial cells [39]. Subimerb et al. reported that the monocyte in patients with Cholangiocarcinoma is correlated with a poor prognosis [40]. On the other hand, lymphocytes play an essential role in immune response, low counts may result in an insufficient immunological reaction against tumor progression and metastasis [38]. Previous research has revealed that lymphocyte to monocyte ratio (LMR) was associated with N stage and distant metastasis [41]. Peng et al. reported that the pre-LMR served as a predictor for early recurrence of Cholangiocarcinoma [42]. Meanwhile, a high neutrophil count was associated with poor prognosis and recurrence in ICC [43]. Stefan et al. reported that neutrophil to lymphocyte ratio was independently associated with worse overall survival among ICC patients [44]. In the present study, a high WBC level maybe reflect increasing in monocytes or neutrophil. The effects of monocytes, lymphocytes and neutrophils on lymph node metastasis should be further studied.

In addition, we concluded that tumors with diameter less than 5cm were less likely to occur LNM, which is similar to previous conclusion [20]. What’s more, we performed more detailed studies for tumor (diameter>5cm), according to multivariate logistics regression analysis results, compared to tumor with 5-10cm, larger tumor (diameter more than 10cm) had a higher metastatic risk to lymph nodes (OR:5.89 VS 3.14). Due to the biological growth behavior of ICC, larger tumor volume means that the tumor has a longer growth cycle and further increases the possibility of lymph node invasive risk.

In addition, the present study found that the type of ICC boundary on radiological image was closely related to LNM, a distinct boundary played a protective role in reducing the likelihood of LNM occurrence, similar result has been reported previously [20]. Microinvasion may reveal a possible mechanism of tumor aggressiveness to lymph nodules [45]. As showed in Fig. 4, boundary served as the second important feature after WBC. Two other independent predictive factors were ALD and smoking. A meta-analysis of eight studies [46] reported that alcohol was major risk factors for ICC. Drinking alcohol causes alcoholic liver disease, which is greatly associated with increased ICC risk [47], as smoking dose [48]. Nonetheless, the relationship between ALD, smoking and LNM in ICC patients was comprehended poorly. Interestingly, we found that ALD was a protective factor for LNM. This finding seems to contradict the existing literature identifying ALD as a risk factor for various cancers, including ICC [46, 47]. To reconcile this apparent paradox, we propose several hypotheses. First, ALD-induced immunosuppression may alter the host’s immune landscape, reducing the attack of immune cells on cancer cells and thus reducing the spread of lymphoid tumors (Gao & Bataller, 2011). Second, liver pathology associated with ALD, particularly cirrhosis, may adversely alter the hepatic microenvironment, impeding tumor cell migration and invasion due to tissue reorganization and vascular changes [49] . Third, there may be a potential selection for survival bias, whereby ALD patients who die prematurely due to liver disease complications do not have sufficient time to develop LNM, leading to an underestimation of the risk factors associated with LNM in long-lived populations. Finally, the chronic inflammatory state associated with ALD may inhibit tumor spread, contrary to the generally accepted view that inflammation promotes cancer progression [50, 51]. These considerations highlight the complexity and individual variability of tumor biology and underscore the need for further research to elucidate the mechanisms by which ALD affects ICC metastatic behavior, thereby providing new insights into therapeutic approaches and patient management. Smoking was significantly associated with LNM and was an independent risk factor for LNM. Therefore, in people with a preliminary diagnosis of ICC, we recommend smoking cessation. However, whether quitting smoking can reduce the risk of LNM in patients with a history of smoking needs to be further verified.

To our knowledge, this paper is the first study to develop and validate a predictive models for predicting LNM in ICC applying machine learning algorithms. The model distinguishes from linear models adopted by previous studies, which can maximize clinical parameters and improve the diagnosis accuracy.

The XGB model initially proposed by Chen et al. in 2016 possessed the best prediction performance [22], it has a high accuracy and fast processing time and has been regarded as a more reliable algorithm when the sample size is limited [52]. Therefore, XGB is suitable for our study which is a small sample from a single medical center.

Finally, we established a concise, visualizable and dynamic online application based on XGB model, the real-time risk of LNM can be calculated and more rational and specific treatment regimens for patients can be tailored according to the personal information. For example, when an ICC patient presented with the following clinical characteristics: tumor diameter less than 5 cm, no boundary, no smoking, ALD and serum WBC count is 5000/µL. We inputted above data into the web calculator, then the application integrated each factor and calculated automatically total probability of LNM, the output result was approximately 6.5% (Fig. 5), indicating that the patient had a low risk to lymph node metastasis. Therefore, we do not recommend further PET/CT monitoring and lymph node dissection.

To sum up, we constructed a machine learning-based predictive model with a good performance to predict LNM in patient with ICC based on independent factors, including ALD, smoking, boundary, tumor diameter and WBC level. In addition, we did an attempt to translate research outputs into clinical practices by builting an online calculator, and the real-time predictive tool may aid in decision-making and management of ICC patients.