Chronic diseases are major challenges in the global public health field. Accurate prediction of chronic diseases is crucial for timely intervention and management, significantly enhancing patient care and reducing treatment costs. However, category imbalance in medical data seriously restricts the clinical applicability of prediction models. This study proposes a chronic disease prediction method based on adaptive soft selection stacking, combined with the DAC-MLPNet meta-model, effectively predicting chronic diseases through two innovative techniques. First, the SMOTENC algorithm, an extension of SMOTE that supports mixed categorical and numerical features, is employed for data resampling, incorporating local density checks to ensure sample quality and balance. Second, adaptive soft selection stacking is applied to effectively combine predictions from multiple models, enhancing overall predictive accuracy and robustness. The processed data is then fed into the DAC-MLPNet meta-model, it utilizes a deep convolutional multilayer perceptron network for high-dimensional feature extraction and prediction. Experimental results show that the proposed model achieves 97.1% accuracy on the diabetes dataset and outperforms traditional machine learning methods on the heart disease and stroke datasets. This not only improves the accuracy and robustness of predictions but also provides valuable insights for medical decision making.