Assessing discharge conditions for elderly coronary patients, especially those with serious or complex disease symptoms, is a crucial task for physicians. Clinical evaluations need to consider demographic information, physiological monitoring, diagnoses, and vital signs, which are multi-dimensional, multi-source, and heterogeneous data. While machine learning models can predict discharge risk based on these data, challenges remain in dealing with redundant features and extracting key features to enhance prediction performance. In this work, we restructured patients’ comprehensive characteristics into a data sequence according to multiple vital clinical stages. Through feature engineering, we constructed steady-state indexes (SSIs), which are generated features that track the patient’s condition changes and the stability of vital signs. Additionally, we standardized variable-length biochemical test data to a fixed-length to address the issue of inconsistent data lengths. Then, we proposed a two-stage multi-module deep fusion network for discharge risk assessment. In the first stage, the data was divided into modules and we extracted features from biochemical test data, SSIs, and comprehensive clinical data using a transformer encoder, CNN, and BiLSTM, respectively. In the second stage, we designed a dual-layer attention fusion network, where dual pooling channel attention was applied to biochemical test data to capture more relevant relationships for discharge results, and sparse attention combining local and simple global information was used on aggregated features to reduce computational complexity. Experiments were conducted on datasets collected from three local 3A hospitals, and the results demonstrated that our method outperformed other methods in the evaluated metrics. Ablation experiments further verified the benefits of segmenting different types or sources of data into different modules for clinical data analysis.