Excessive expansion of tandem repeats (TRs) increases the risk of various diseases, including cancer, autism, and neurodegenerative disorders. Accurately predicting the pathogenicity of TR expansions is challenging due to their dependence on repeat length, motif repeat threshold, and the dynamic nature of repeat changes. The existing method primarily uses association information between TR loci and genes, neglecting sequence characteristics that contribute to TR instability. To address these challenges, we introduce TREPP, a machine learning-based framework designed to predict high-risk pathogenic TR loci. To improve the predictive performance, TREPP integrates multi-dimensional features, including flanking sequence composition, GC content, and the density of surrounding TR regions. At the same time, the lack of validated pathogenic TRs results in a severe imbalance between positive and negative samples. To solve this problem, we employ a multi-sampling strategy and train multiple CatBoost models through ensemble learning to mitigate data imbalance. Test on a manually curated pathogenic TR dataset, TREPP achieves an AUPRC of 91.57%, increases 4.17% over the state-of-the-art method RExPRT. Additionally, in the OMIM and GeneTrek disease gene databases, TREPP identifies more TR-associated pathogenic genes than RExPRT, demonstrating its comprehensive predictive capability.