The International Classification of Diseases (ICD) coding task assigns medical codes to electronic health records (EHR), supporting healthcare services such as medical reimbursement and clinical decision-making. Current ICD automatic coding methods primarily focus on learning clinical text representations and modeling label correlations, while the rich knowledge contained in similar instances is often overlooked. To address this issue, a novel ICD automatic coding method based on debiased contrastive learning and a k-nearest neighbors (KNN) mechanism was proposed. Firstly, coding synonym attention was used to extract code-related text representations, which were then optimized through debiased contrastive learning. During the inference phase, training samples were used as references and KNN was employed to retrieve similar instances. Finally, the results from both components were combined using a weighted sum to generate the final coding prediction score. Experimental results on two subsets of the MIMIC-III dataset show that the proposed model improves the baseline in terms of Macro-F₁ and P@5 metrics by 0.6%~1.1% and 0.5%~0.7%, respectively, validating the effectiveness of the method.