Chronic obstructive pulmonary disease (COPD) is a chronic respiratory disease that seriously endangers human health and has high incidence and mortality worldwide. Therefore, an effective predictive model is required for COPD diagnosis. Given the limited data samples available in current COPD studies, we propose a method for diagnosing COPD based on transfer learning called balanced probability distribution (BPD) algorithm; this algorithm integrates instance-and feature-based transfers to improve the prediction accuracy of the model. First, instance-based cascaded transfer learning was used to initialize the weight distribution of the training data and obtain instances closer to the target domain. Second, the crossdomain feature filtering algorithm was adopted to filter irrelevant features, eliminate redundant features, and obtain the co-occurrence features of the source and target domains. Moreover, the remaining features were assigned different weights and transformed into the same space to reduce the distribution difference between the domains. Third, the BPD algorithm was used to balance the examples and the co-occurrence features from multiple disease source domains and construct a more suitable classification model of the target domain. Finally, the elastic network was used to further improve the generalization performance of the model. The experimental results show that the prediction effect of the BPD model is better than that of state-of-the-art methods and has strong generalization ability and robustness. We proved that our proposed BPD method works well in the COPD prediction model when the sample size is small. INDEX TERMS Balanced probability distribution (BPD) algorithm, chronic obstructive pulmonary disease (COPD), feature extraction, few-shot learning, transfer learning.