Accurate diabetes diagnosis and readmission prediction are crucial for effective patient health management. In this paper, we developed an Adaptive Multi-Channel Fusion Network (AMCFN). Specifically, we have defined a feature enhancement module that combines an attention mechanism to dynamically assign relevant weights to input data, thereby enabling the model to focus on processing task-relevant inputs. Meanwhile, we designed a multi-channel fusion network that utilizes various network architectures to simultaneously extract diverse deep features from the input data and concatenate these features. Finally, we use multiple kernel functions to map the fused features into different feature spaces, and combine their outputs through weighted sum into a new matrix for the final prediction task. Extensive experiments were conducted on the Pima Indian Diabetes Dataset (PIDD), the Early-Stage Diabetes Risk Prediction Dataset (ESDRPD) and the Diabetes 130-US Hospitals for Years 1999–2008 Dataset (D130-US). Our model outperforms existing baseline models in diabetes diagnosis and readmission prediction. The ablation experiment also demonstrates the effectiveness of the design of each module. Our research provided improved clinical decision support for diabetes diagnosis and readmission prediction, with the potential to reduce the waste of medical resources.