The clinical strategy for patients with PTMC remained controversial. Should we opt for surgery or just active surveillance only? If surgery was being performed, would it necessary to do the bilateral lobar resection (BLR), or only operating on the affected lobe (unilateral lobectomy)? For non-invasive, clinically node-negative (cN0) PTMC, would it have a better prognosis if the prophylactic central lymph node dissection (pCLND) was carried out? The availability of suitable prognostic models became important. Conventional LR model and ML each had their own advances and limitations. We established models for cervical LNM by conventional LR and ML algorithms with the goal of evaluating their performance in prediction of LNM in PTMC.
Risk factors for cervical LNM of PTMC
In the ML model and the traditional regression model analyses, four risk factors were found to be most closely associated with cervical LNM in PTMC. These includes, extrathyroidal metastasis (ETE), tumor size, age, and multifocality. Among them, ETE was the most important factor affecting the outcomes predicted by the models. It was confirmed by many clinical research that ETE predicts negative clinical outcomes in papillary thyroid cancer . All levels of extrathyroidal extension, including microscopic, were associated with a increased risk for nodal and distant metastasis . Unfortunately, minimal ETE was often difficult to identify before the operation, The utility of intra-operative frozen section for the evaluation of microscopic extrathyroidal extension in papillary thyroid carcinoma seemed important and the patients who were diagnosed as ETE in postoperative pathology without pCLND. might be recommended to be intensively followed up. Since tumor size was a simple parameter that can be determined with ultrasound images, it was widely used to predict the aggressiveness of PTMC to aid clinical decision-making. In PTC, the larger the tumor was, the greater the risk of cervical LNM was , Similar results were found in PTMC . In our study, patients with tumors ≥5 mm had a significantly higher risk of cervical LNM compared to those with tumors < 5 mm, which was consistent with the results of previous clinical studies and meta-analyses . Age was an important risk factor for thyroid cancer. Children and young adults often present aggressive disease patterns and advanced stages, and had a relatively high rate of PTMC lymph node metastasis . Thus, we put children and young adults into the same category (< 25 years old) with a risk value of 1. Compared to this group，for adults, with every 5 years of age increasing, the risk of PTMC cervical lymph node metastasis gradually decreased. The OR values were: 0.586, 0.362, 0.26, and 0.239 showing a relatively obvious and gradual decreasing trend. Therefore, for older patients, the extent of surgery should be conservative and less frequent follow-up was allowed. On the other hand, younger patients, especially children and adolescents, should be treated more aggressively and followed up closely to reduce the risk of recurrence. The multifocality of the tumor was also one of the factors closely related to PTMC cervical lymph node metastasis. Papillary thyroid carcinoma often occurs in the form of two or more independent lesions in the thyroid (18–87%) . Multifocality may arise from intrathyroidal metastases from a single malignant lesion or from multiple lesions of independent origin with intrathyroidal metastases . In our study, the incidence of multifocal PTMC was 35.05%. It was a major risk factor in both the ML and the conventional LR models. The results were consistent with previous studies . In contrast, although some previous studies had suggested that bilateral tumor was a risk factor for thyroid neck lymph node metastasis . Our study found that laterality was not associated with cervical lymph node metastasis. There were no significant differences between unilateral or bilateral tumors in either univariate or multivariate analysis. Histopathological subtypes were also closely linked to the lymph nodes of PTMC. Some histological subtypes of PTC were classified as aggressive variants of PTC (AVPTC), which included columnar/tall-cell variant (TCV), and diffuse sclerosis subtype . In our study, results consistent with previous studies were also obtained. However, among the above risk factors of cervical LNM, histiocytic subtype was a factor that could not be determined before surgery. ETE and multifocality were often hard to find during preoperative routine inspection. They largely depended on the accuracy of intraoperative frozen section pathology analysis. This may limit the applicability of the machine learning model in preoperative clinical prediction.
The relationship between tumor prognosis and population sociology has received increasing attention. In our study, in addition to age, patient gender was also found to be associated with cervical LNM in PTMC. It was well known that the prevalence of thyroid cancer in men was much lower than that in women, although men had a higher rate of cervical LNM and a poorer prognosis . In our study, we found that male sex was also a risk factor for cervical LNM. Compared with women, the OR was 2.279. The mechanism was unclear, although some studies suggested that estrogen might regulate the proliferation of thyroid cells by combining with estrogen receptor (ER) α and ERβ . Since ER expression levels differ between males and females, this may be one of the reasons for the difference in the sex ratio of cervical LNM in PTMC patients. Race and region of residence were also associated with the risk of cervical LNM in PTMC. Black race was a protective factor for cervical LNM (OR: 0.551) relative to white race, while other races were not significantly different from white race. In a previous study on race and PTC prognosis in the SEER database, black Americans had lower overall survival than white Americans (HR: 1.127). However, there were fewer lymph node metastases in classic papillary thyroid carcinoma (OR: 0.476) and follicular subtype papillary thyroid carcinoma (OR: 0.522) in black Americans . Genetic variation may be a possible mechanism for the differences. In addition, it is possible that the limitation in health care resources for black Americans might have leaded to less prophylactic neck dissections and/or less proper ultrasounds, causing an overt reduction in observed incidence rate. The distribution of PTMC varied by region in the United States. The Pacific coast accounts for the largest proportion (42.767%) and had the highest incidence of lymph node metastasis. (11.9%). This may be related to the uneven iodine intake and different racial distribution of residents. However the factors that effected the cancer disparities are complex, including lifestyle, income, health security and access to affordable health services of high quality . One of the thought-provoking results in our study was the effect of marital status on cervical LNM in PTMC. We found that being single was the only marital- status related risk factor for the cervical LNM. Even the divorced, widowed or separated had a better outcome than the single. It is possible that spouses may encourage patients to seek medical attention for alarming symptoms thus resulting an early diagnosis of the tumor [37, 38]. In addition, the support of the family especially the spouse might help reduce the stress and depression in the patients which might help recovery from the disease. Our analysis showed that social and family relationship was an important factor affecting tumor prognosis.
Predictive model performance comparison
There were several ML algorithms commonly used in predictive model construction. Different ML approaches had different advantages and disadvantages. In this study, compared with the other four ML algorithms (AdaBoost, RF, GNB and MLP), XGboost was found to be the best model for predicting cervical LNM in PTMC using a dataset derived from the SEER database. In this study, we found the XGboost algorithm performance best both in AUC and in the accuracy of model construction. Its accuracy was improved in optimized procession.
Compared with conventional LR model, machine learning methods performance better in predicting cervical LNM outcomes. Though the advantage in AUC value was not so obvious. This may be attributable to the variables that were selected into the model. There were no correlation or collinearity among all variables. In addition, the variables selected were simple and there were not so many features. The fact that the machine learning method show only limited advantage in our study indicates if the variables were simple and did not have any collinearity, conventional LR could also be a good choice for model construction. It was likely that only when there were complex, high-dimensional data available the ML method might show much more substantial advantages.
Visualization of feature importance
We visualize OR and 95%CI for variables identified by conventional LR in the form of nomogram and forest plot map to help understand the model . Nomogram and forest plot clearly showed the ETE, larger tumor size, histology of column cell, multifocality, male, single, and the region of Pacific Coast were all risk factors while older age, race of black and the histology of follicular variant and encapsulated were protective factors. ETE had the biggest OR ration. It was a very important positive risk factor in predicting the result of the cervical LNM in PTMC. In contrast, the feature of marital status and race were not as important as other factors.
For a long time the ML only provided a ranking of feature importance and did not specify whether each important factors was protective or dangerous in the way LR did. The “black-box” characteristics of ML algorithms made it difficult to understand. In this study we leverage SHAP to illustrate the factors in the predict model constructed . The map provided by SHAP helped to visualize the prediction power of valuables. With SHAP of XGboost, we can easily find that ETE, gender, multifocality and tumor size were clearly the positive risk factors for the cervical LNM. ETE was the most important risk factor of the cervical LNM. This agreed with what was shown by nomogram and forest plot from LR modeling.
The main limitations of this study were as follows: First, this study was mainly limited by the retrospective nature of the analysis, so confounding is inevitable. Second, the identification of cervical LNM was primarily derived from the collection of data on cases where therapeutic lymph node dissection was performed. The incidence of cervical LNM in this study was much lower than it was in some other studies. This suggested that the incidence of cervical LNM may be underestimated. In addition, we did not differentiate the central and lateral LNM of cervical LNM. There may be different characteristics of these two kinds of LNM which is also important for clinical strategy. Third, SEER database only included patients lived in the United States of America. The factor of residence in our model might only be representative of this particular cohort and reflected health care differences, such as access to proper prophylactic neck dissections and/or ultrasounds, existing in different geological locations in America. It was possible that the residence factor was either not relevant or contributed in a different way in other populations of the world. Fourth, some high-risk factors or characteristics associated with cervical LNM were not documented in the SEER database, such as autoimmune thyroid disease (AITD), preoperative ultrasound, imaging, fine-needle biopsy, or molecular analysis. We hoped that in the future, prospective multicenter studies with long-term follow-up data will help obtain additional useful clinical or social characteristics to further improve the model.