Figures 1A, C analyze the prevalence of thyroid cancer among women and men in China from 1990 to 2021. In 1990 ( Figure 1A ), the crude prevalence rate of thyroid cancer in China increased with age, and the prevalence rate peaked at the age of 55-59. However, in 2021 ( Figure 1C ), the peak prevalence rate for women shifted to the age of 60-64, while that for men remained at 55-59, showing a significant decline at the age of 60-64. Since 1990, the overall prevalence rate has shown a marked increase. For women, the prevalence rate is higher than that of men in most age groups. Moreover, the development trend of the disease among women is largely consistent, with a decline occurring between the ages of 45 and 49.

Between 1990 and 2021, cases per 100,000 people rose from 0.87 (95%UI 0.69, 1.0) to 3.88 (95%UI 3.12, 4.88). Among males, the number of affected cases increased from 0.19 (95%UI 0.15, 0.26) to 1.55 (95%UI 1.11, 2.08). The number of female patients increased from 0.68 (95%UI 0.49, 0.84) to 2.33 (95%UI 1.69, 3.46).

The ASPR of the entire population increased from 8.1 (6.41, 9.66) in 1990 to 20.01 (16.14, 25.23). The proportion of males increased from 3.52 (2.77, 4.73) to 15.87 (11.35, 20.96). The ASPR of females increased from 13 (9.36, 16.16) to 24.24 (17.62, 36.4). This indicates that the prevalence rates of both men and women have increased, and the ASPR of women has always been higher than that of men ( Table 1 ).

In 1990 ( Figure 1B ), the crude mortality rate of the age group under 50 in China was close to 0. It began to change significantly at the age of 50. Before age 80, women had a higher mortality rate than men. After the age of 80, the mortality rate of men increased rapidly and reached its peak between the ages of 90 and 94. By 2021 ( Figure 1D ), the crude death rate for the under-50 age group remained close to zero. The mortality gap between men and women remained small until age 75. However, the mortality rate for men increased rapidly after the age of 75, reaching its peak at 90-94 years old. Furthermore, compared with 1990, the mortality rate of men is increasing while that of women is decreasing. Despite these changes, the overall mortality trends of both sexes remain similar.

The AMPR of the entire population decreased from 0.47 (0.4, 0.55) in 1990 to 0.39 (0.31, 0.47). The AMPR of men increased from 0.38 (0.31, 0.51) to 0.49 (0.33, 0.64), while that of women decreased from 0.57 (0.43, 0.69) to 0.32 (0.24, 0.45) ( Table 1 ).

Based on the data analysis of GBD 2021, the risk of death from thyroid cancer caused by high body mass index (BMI) in China showed a significant increasing trend from 1990 to 2021.As shown in Figure 2 , the proportion of deaths from thyroid cancer due to high BMI in males increased from 6.89% to 10.40% (an increase of 50.97%), and in females it increased from 7.84% to 12% (an increase of 53.06%).

It is worth noting that metabolic risks associated with female obesity may be more sensitive to the prognosis of thyroid cancer. Overall, high BMI remains a key modifiable risk factor, with a more pronounced impact on women.

Using BMI as the exposure variable and thyroid cancer as the outcome variable, select SNPs related to BMI as instrumental variables, and meet the following conditions: (1) Association hypothesis: Select SNP loci that are strongly correlated with exposure factors to ensure that the selected SNPs accurately reflect the genetic characteristics of thyroid cancer. (2) Independence hypothesis: SNP loci associated with confounding factors were excluded to eliminate the influence of external interference on the research results. (3) Exclusivity assumption: Only SNP loci that have an effect on the junction through exposure factors are retained to ensure the uniqueness of the causal relationship (32, 33).

When choosing tools, set the significance threshold at P<5×10-6 to preliminarily screen SNPS that are significantly related to BMI (34). Then, set the linkage disequilibrium parameter r2 = 0.001 and the genetic distance to 10,000 kb to avoid the bias caused by the linkage disequilibrium among SNPS. To conform to the exclusivity hypothesis, P<5×10-6 was set to eliminate SNPS related to the outcome (35). To meet the independence hypothesis, SNPS related to confounding factors need to be eliminated, and PheoScanner is used to query SNP-related models.

Through MR Analysis, there is a positive causal relationship between thyroid cancer and BMI(MR-Egger: OR=1.822, 95%CI: 0.1414-23.519, P=0.0646; IVW: OR=0.953, 95%CI: 0.358-2.2535, P=0.00242, Weighted mode: OR=0.403,95%CI: 0.048-3.395, P=0.0404), as shown in Table 2 .

The IVW method is the most commonly used and statistically effective method in MR analysis. If the P value is less than 0.05, it is considered that there is statistical evidence supporting the causal effect of exposure on outcomes. However, the IVW method is highly sensitive to horizontal pleiotropy, and the estimation results of IVW may be biased. If the P-values of IVW and weighted median method are both significant and consistent in direction, this will enhance the conclusion of causal inference. The results showed that the P-values of IVW and Weighted mode were both less than 0.05, and the effect directions were consistent, indicating a causal relationship between BMI and thyroid cancer. This suggests that BMI may be an important factor in increasing the risk of thyroid cancer.

In order to better predict the future prevalence trend of thyroid cancer, we used ARIMA to predict the prevalence and mortality of thyroid cancer in the next 10 years. The prevalence is projected to rise from 20.01% in 2021 to 23.58% by 2031. The mortality rate is expected to remain stable, around 0.41%, during the same period. As shown in Figure 3 .

Through the predictive analysis of GBD data, it was found that the prevalence of thyroid cancer still shows an increasing trend. To gain deeper insights into the current status and future progression of the disease, we built several machine learning models using hospital-based data. A total of 6,565 patients were included in this study, including 4,631 patients with malignant thyroid cancer and 1,934 patients with benign thyroid cancer. The data were randomly divided into the training set and the test set. The training set had a total of 5252 samples, and the test set had a total of 1313 samples.

For cases of missing data, Hot Deck Imputation was employed. Find the complete sample that is most similar to the missing value sample from the dataset, and use the value of this complete sample to fill in the missing value. The advantage of this method lies in its ability to preserve the original distribution of data, as it is filled based on actual observed values.

This study adopted a feature selection strategy that comprehensively considers clinical relevance, data availability and completeness, and model interpretability requirements when constructing a thyroid cancer benign and malignant prediction model. We carefully selected 9 clinical features closely related to thyroid disease diagnosis and patient basic information from hospital electronic medical records as model inputs.

Before conducting the model establishment analysis, the normality test of each index should be carried out first (36). When the P value of the normality test is greater than 0.05, it indicates that the index follows a normal distribution. The specific data are shown in Table 3 . The indicators that follow a normal distribution include age, BMI, TG,FT3 etc.

Six machine learning prediction models were constructed to predict the prevalence of thyroid cancer patients. During the model parameter tuning process, we utilized the grid search method. The aim is to find the optimal combination of hyperparameters through algorithms rather than manual trial and error, in order to achieve the best performance of the model. Predefine the discrete value range or list for each hyperparameter, and then iterate through all possible combinations of hyperparameters. For each combination, the model undergoes training and evaluation, followed by recording of performance metrics.

For the evaluation of the model, we utilized metrics such as AUC, accuracy, sensitivity, and specificity. The specific calculation formulas are as follows:

True Positive (TP): The number of samples that are actually positive and also predicted as positive. True Negative (TN): The number of samples that are actually negative and also predicted as negative. False Positive (FP): The number of samples that are actually negative but are predicted as positive. False Negative (FN): The number of samples that are actually positive but are predicted as negative. Table 4 shows the performance evaluation indicators of the six machine learning models constructed in this study (XGBoost, RandomForest, DecisionTree, GNB, MLP, SVM) in predicting the benign and malignant nature of thyroid cancer. The XGBoost model exhibits the best overall predictive performance. Its AUC value is as high as 0.999, accuracy is 0.984, sensitivity is 0.973, specificity is 0.988, positive predictive value is 0.968, negative predictive value is 0.990, and F1 score is 0.971. This indicates that XGBoost has extremely high accuracy and robustness in distinguishing benign and malignant thyroid cancer. In contrast, the performance of MLP and SVM is relatively poor. The AUC of MLP is only 0.661, the accuracy is 0.715, the sensitivity is only 0.469, and the F1 score is 0.480, indicating that its performance in this thyroid cancer benign and malignant prediction task is not ideal. Therefore, the XGBoost model based on gradient boosting demonstrated excellent performance in the benign and malignant prediction task of thyroid cancer in this study, far surpassing other traditional machine learning methods.

The importance of features and the interpretability of the model are key aspects in evaluating performance. In the XGBoost model, BMI has the greatest influence among the nine common clinical features, as shown in Figure 4 . The SHAP graph illustrates how each feature contributes to the model by displaying the distribution of SHAP values. The horizontal axis represents the SHAP value of the XGBoost model, and the color of the points indicates the size of the feature value. If the SHAP is positive, it indicates that the feature has a positive contribution.

From Figure 4 , it can be clearly seen that the SHAP value distribution range of BMI is the widest, and its points are mainly concentrated in the larger positive areas. This indicates that BMI is the most important clinical feature affecting the prediction of benign and malignant thyroid cancer. A high BMI value is associated with a higher SHAP value, indicating that a high BMI significantly increases the likelihood of the model predicting malignancy. This is consistent with the conclusion drawn from Mendelian randomization analysis in the study that there is a positive causal relationship between BMI and thyroid cancer, further emphasizing the crucial role of BMI in the occurrence and development of thyroid cancer.