OBJECTIVE: Assessing the severity of coronary heart disease (CHD) is critical for clinical decision-making. Thyroid autoantibodies are associated with cardiovascular disease, but their ability to predict CHD severity remains unclear. This study aimed to systematically elucidate the predictive value of thyroid autoantibodies for CHD severity using machine learning methods. METHODS: This retrospective study included 942 patients hospitalized in the cardiovascular department between January 2024 and June 2025, comprising 590 patients with severe lesions and 352 with nonsevere lesions. Traditional statistical analysis employed correlation analysis and multivariate logistic regression. Eight machine learning models were subsequently constructed and compared: logistic regression (LR), support vector machine (SVM), K-nearest neighbors (KNN), adaptive boosting (AdaBoost), multilayer perceptron (MLP), random forest (RF), gradient boosting (GB), and XGBoost (XGB). The optimal model was determined using Shapley additive propagation (SHAP), and the robustness of the core findings was validated through hierarchical analysis and subgroup feature importance comparisons. RESULTS: Multivariate logistic regression revealed that log anti-TPO was an independent risk factor for severe coronary lesions (OR = 2.19; 95% CI: 1.87-2.57; P<0.001), whereas log anti-Tg shows a negative correlation after adjustment (OR = 0.72; 95% CI: 0.61-0.85; P<0.001). However, given the significant multicollinearity between the two variables (r = 0.55), this negative correlation strongly suggests that it is a statistical artifact. The gradient boosting tree performed best (AUROC:0.855). SHAP analysis consistently confirmed three key predictive features: log anti-TPO, log anti-Tg, and glycated hemoglobin (HbA1c). SHAP dependency plots further revealed a distinct threshold effect for log anti-TPO, while elevated log anti-Tg was associated with reduced risk, However, this negative association is likely a statistical artifact rather than an independent protective effect. Stratified analysis and sex-specific feature importance assessments confirmed that log anti-TPO demonstrated a highly robust and significant predictive value across all the subgroups, ranking as the primary predictor in both the male and female cohorts. CONCLUSION: The presence of thyroid autoantibodies represents an independent key predictor of CHD severity. Thyroid peroxidase antibodies (TPO-Ab) serve as a strong risk marker, The gradient boosting tree model demonstrated optimal predictive performance when these biomarkers were integrated.