Diagnostic modeling of voice disorders is important to improve the efficiency of patient diagnosis. However, due to sample imbalance and complexity of voice data analysis, basic acoustic features and traditional machine learning methods are not effective in predicting voice status. This letter proposes a three-layer ensemble learning model with two innovations: 1) At the feature level, acoustic features are expanded using Multilayer Perceptron and SincNet; 2) Structurally, a tree-based ensemble learning model is proposed that utilizes XGBoost for feature selection and feature transformation, with LightGBM as the classifier. Experiments have shown that the proposed method in this paper outperforms traditional machine learning models.