<p>Data streams have emerged as a significant data form in the big data era, with concept drift being one of the most critical research challenges in data stream mining. Currently, concept drift detection methods typically assume that labels for all samples are known-an assumption that is unrealistic in real-world scenarios. Additionally, the imbalanced nature of data streams further complicates the task of concept drift detection. To address these issues, this study proposes a concept drift detection method for imbalanced data streams with limited labels (CDD-IDSLL). Specifically, a sample’s prediction certainty metric is defined based on the maximum prediction probability of samples and the sum of differences between adjacent prediction probabilities, enabling the sample query strategy to adapt to data streams with a relatively large number of classes. A weighted average method is introduced to compute a dynamic sample’s prediction certainty threshold vector, which distinguishes the varying difficulty levels of classifier predictions across different classes. Furthermore, a weighted accuracy calculation method is defined based on sample arrival time and class imbalance ratio, effectively evaluating classification accuracy in imbalanced data streams. Additionally, a dynamic adjustment method for concept drift thresholds is proposed, leveraging the average distance between adjacent misclassified samples to enhance concept drift detection performance. Comparative experiments on nine synthetic data streams and three real-world data streams demonstrate that the proposed method outperforms eight existing state-of-the-art concept drift detection methods in terms of detection performance.</p>