Efficient prediction of landslide dam stability is crucial for emergency response and damage reduction. In this study, a comprehensive analysis is conducted on eight landslide dam characteristics. Four machine learning (ML) algorithms, namely Support Vector Machine (SVM), Random Forest (RF), Artificial Neural Networks (ANN) and Logistic Regression (LR), are then applied to predict the stability of landslide dams. This prediction is based on two stability definitions: the dam's ability to endure for over a year and its collapse status at the time of the study. The results derived from the test set distinctly demonstrate that the RF model outperforms the other three ones in terms of its effectiveness. By employing the Synthetic Minority Over-sampling Technique (SMOTE), the issue of the RF model being biased towards predicting unstable dams due to imbalanced samples has been effectively alleviated. This approach resulted in overall accuracies of 76.19 % and 82.35 %, with biases of 0.8 % and 11.6 % and Classification Efficiency Index (CEI) values of 1.024 and 1.046, respectively, under the two stability definitions. Through Principal Component Analysis (PCA), it is further determined that the largest 5 % of particles constitute the primary materials influencing the stability of landslide dams. Additionally, a novel index termed the dam composition index (DCI) has been proposed to characterize the gradation of landslide dams. The proposed prediction method for landslide dam stability demonstrates outstanding performance and contributes to more effective emergency planning.