Differentiated thyroid cancer (DTC) is the most common endocrine malignancy, with the global incidence increasing dramatically in recent decades (Cabanillas et al., 2016; Siegel et al., 2020). Most DTC patients have a good long-term prognosis because of biological characteristics and effective responses to treatment modalities (Lim et al., 2017; Liu et al., 2018). For TC patients with distant metastasis (DM), however, the overall prognosis is significantly worse (Durante et al., 2006; Sugitani et al., 2008; Farooki et al., 2012; Nies et al., 2021).

The main histological subtypes of DTC include papillary thyroid carcinoma (PTC), and follicular thyroid carcinoma (FTC), <10% of people with DTC will develop DM (Durante et al., 2006). Most DTCDM is asymptomatic and only detected during systematic surveillance or systemic metastatic examination of malignant lymph nodes. The common site of distant metastases from thyroid cancer is the lung, followed by the bone, brain, and liver (Chesover et al., 2021; Liu et al., 2021). Because the incidence of DM is low and there is often an absence of symptoms, it is frequently overlooked at the time of initial TC diagnosis. Ten-year survival rates are often used to assess DTC treatment efficacy and characterize risk factors. For patients with metastatic thyroid cancer, treatment is often individualized and usually consists of thyroid surgery with adjuvant radiotherapy or chemotherapy (Sampson et al., 2007). While prognostic models have been developed to study factors influencing different primary cancer subtypes (Zhao et al., 2019; Zhou et al., 2020; Jin et al., 2021; Kong et al., 2021), the 10-years overall survival of patients with DTCDM is unclear, especially because OS has not changed significantly in recent decades (Goffredo et al., 2013). In addition, there are few prognostic models for DTCDM that can help to inform patient follow-up and treatment decisions (Chen et al., 2021).

Machine learning (ML) is an emerging multi-disciplinary approach used to correlate multiple discrete variables and accurately predict outcomes. Following the development of evidence-based medicine and the need for more advanced tools to collect medical data with complex structures and large sample sizes, ML emerged as an alternative approach to disease diagnosis and prognosis, with high predictive performance and a wide range of applications (Goecks et al., 2020; May, 2021). ML algorithms are now being successfully applied to predict cancer survival (Angraal et al., 2020). The XGBoost algorithm (XGB), in particular, is shown to have excellent prediction performance in previous studies (Senders et al., 2020; Jiang et al., 2021).

The Surveillance, Epidemiology, and End Results (SEER) program is a population-based cancer registry system sponsored by the United States’ National Cancer Institute (NCI) that currently covers about 28% of the population in 18 registered states (Noone et al., 2016). Using the considerations listed above, a prognostic modeling analysis of patients with DTCDM was conducted using SEER data. This study assessed the ability of clinical and non-clinical factors to predict DTCDM using the XGB model. The XGB model was also built to predict the 10-years OS rate of TC patients with DM. The performance of the XGBoost model using logistic regression (LR), random forests (RF), and support vector machine (SVM) models were compared.

Survival prediction is very difficult in malignancy but important for treatment planning and patient management (Cheon et al., 2016). Compared to the empirical predictions of clinicians, the XGBoost model provides a more reliable option for the 10-years survival status of patients with distant metastases from thyroid cancer. The current study found that XGB had a good predictive value and could help clinicians to develop a rational individualized treatment and management plan. Although thyroid cancer is a relatively slow-growing cancer, once distant metastasis has occurred, the tumor grows exponentially at the location of metastasis, explaining why patients with DM have a worse prognosis (Rajan et al., 2020). While DTC generally has a favorable prognosis, the clinical course can be poor (See et al., 2017). Assessing the survival of patients with DTCDM is thus of great clinical importance.

This study used survival analysis to screen for factors that may affect the OS of patients with DTCDM. Although, the multiple Cox model we developed showed that race is not a prognostic factor for DTCDM patients. One study showed that race does affect the survival of patients with differentiated thyroid cancer and that treatment options need to be specific to different racial groups (Tang et al., 2018). However, Crepeau et al. (Crepeau et al., 2021) found that thyroid cancer prognosis and recurrence did not differ by race following the same surgical approach, especially when all patients receive the same quality of care. It is likely that differences in the prognosis of TC patients by race are the result of social and economic differences between racial groups.

The XGB algorithm is a new type of AI algorithm that is easy to use, efficient and accurate. This algorithm is becoming increasingly popular in the medical field and is now widely used for disease prediction and early diagnosis. The current study used the prognostic variables obtained from survival analysis to develop a prognostic model for DTCDM OS using the XGB algorithm. The model was validated by combining the clinical and non-clinical variables and was shown to be highly effective. Additional factors that may affect OS and provide more clinical information about DTCDM, were also reported. Traditional ML LR also performed well, possibly because of the study data or because LR performs just as well as ML in clinical prediction models (Christodoulou et al., 2019).

Age was the highest-scoring, and thus most significant, variable in the XGB model, Patients <51 years of age had significantly higher OS than patients >80 years of age. The role for age in predicting TC has been confirmed by other studies, with older patients having a poorer prognosis than younger patients (Sampson et al., 2007; Nixon et al., 2012). Compared with the 8th edition of the TNM staging system with a cut-off value of 55 years, the optimal cut-off values for age in this study were 50 and 77 (Amin et al., 2017). This implies that the cut-off value for age may need to be studied in depth for patients with distant metastatic thyroid cancer. In the current study, there was also a significant score for the 51–79 age group. These findings suggest that treatment should be tailored to patients in different age groups.

The results of one study confirm that patients with PTC with distant metastases have a good prognosis after treatment (Sampson et al., 2007). Another study found that sex is a prognostic factor for DTCDM, likely because estrogen production limits cancer progression (Suteau et al., 2021). The current study found that tumor size was an important factor affecting DTCDM OS, with relatively significant scores for both 1–27 mm and 28–65 mm, a finding supported by Nguyen et al. (Nguyen et al., 2018). Han et al. reported that 15–20 mm tumors do not affect the OS of TCDM patients (Han et al., 2017). Unexamined metastases and those localized in the lymph nodes also scored highly. The number of lymph node metastases correlates strongly with the presence of DM while the risk of DM can be assessed based on the number of lymph nodes (Jeon et al., 2016). These findings may help to resolve controversy over the indication of lateral lymph node dissection (Fan et al., 2018). The regional metastatic status of the lymph nodes should be assessed in all patients with DTCDM.

Radiotherapy had a relatively significant score in the OS prognostic model of patients with DTCDM. Studies indicate that radioactive iodine (RAI) treatment is very effective in patients with small metastases, indicating that early diagnosis improves outcomes (Durante et al., 2006; Sampson et al., 2007). Diagnosis of DM and initiation of RAI therapy before overt metastases appear, especially in children and adolescents for whom selective treatment is more appropriate (Sugino et al., 2020). While RAI treatment is beneficial for TC survival, however, high-quality RAI accumulation may increase the risk of secondary tumor mutations and more aggressive disease, thus negatively impacting patient survival (Su et al., 2015; Pasqual et al., 2022). Nies et al. concluded that repeated RAI treatment is unlikely to benefit TC patients and may do more harm than good over their lifespan (Nies et al., 2021). In summary, studies differ on whether and how to treat DTC patients with RAI (Jeon et al., 2016; Lin et al., 2018).

Total thyroidectomy and other surgical procedures were important in the prognostic model. Sampson et al. concluded that a total thyroidectomy should be performed alongside RAI treatment (Sampson et al., 2007). It is also suggested that, where possible, local curative surgery with RAI and thyroid hormone suppression should be performed in patients with DTCDM (Ito et al., 2010). However, the survival benefit of thyroid cancer surgery may vary depending on the site of metastasis (Besic et al., 2016). The current study found that chemotherapy was a strong predictor of OS in patients with DTCDM and a risk factor for OS in survival analysis. Chemotherapy is often administered to patients with large tumors who are no longer candidates for surgery or show iodine resistance. However, a study indicates that chemotherapy is highly toxic and is associated with a poor response rate (Schmidbauer et al., 2017). The high mortality rate of chemotherapy patients may be due to the relative severity of the disease in this group of patients, in addition to the toxic impact of the treatment. Recent studies have shown that targeted therapies such as tyrosine kinase inhibitors (TKIs) offer high survival rates and that patients may have a better outcome if targeted treatments are combined with chemotherapy (Kraeber-Bodéré et al., 2010; Carling and Udelsman, 2014; Viola et al., 2016; Lorusso et al., 2021).

Although the predictive model used in this study had a good performance, there were some limitations. First, the study relied on regression data and some samples with missing information were removed, which may have biased the model. Second, outcome data for individuals receiving targeted therapies were not included in the sample, which may have made the prediction model less comprehensive. Finally, more work needs to be done to explain the predictive efficacy of ML versus traditional statistical methods.

This study analyzed the clinical characteristics and prognosis of patients with DTCDM and constructed prognostic models using four machine learning methods. The XGB model was effective at predicting the 10-years OS of patients with DTCDM and may help clinicians to make more accurate and personalized clinical decisions. This is particularly important to improve the long-term prognosis of high-risk patients.