Objective To develop a computer-aided diagnosis model for lung cancer based on routine health examination data for identifying individuals with a current high risk of lung cancer in health screening settings, thereby providing decision support for subsequent clinical confirmation. Methods Individuals who underwent health examinations at the Health Management Center of West China Hospital, Sichuan University, between 2010 and 2022 were enrolled. After screening, a retrospective cohort of 5257 subjects was retained, comprising 1307 patients with lung cancer and 3950 non-lung cancer controls. A three-tier feature fusion model was designed: Heterogeneous feature encoding module: a multi-layer perceptron and bidirectional encoder representations from transformers (BERT) were employed to extract feature vectors from structured data and unstructured data (medical records and imaging report texts), respectively. Heterogeneous feature fusion architecture: dimensional expansion concatenation coupled with a gated recurrent unit based gating network was implemented to achieve multi-scale feature alignment and deep interaction, thereby addressing dimensional discrepancies and information redundancy. Attention-based decision mechanism: word-level attention with weighted pooling was applied to dynamically capture key features and generate risk probability distributions. Model performance was evaluated using precision, recall, F1-score, and the area under the receiver operating characteristic curve (AUC-ROC). Results The proposed model significantly outperformed both single-data-type models and simple concatenation approaches. On the test set, the proposed model achieved a recall of 0.861, an F1-score of 0.882, and an AUC-ROC of 0.972, substantially surpassing the best-performing model trained on structured data alone (extreme gradient boosting: recall=0.630, F1-score=0.725, AUC-ROC=0.916) and the model trained on unstructured data alone (BERT coupled with a bidirectional long short-term memory network: recall=0.833, F1-score=0.846, AUC-ROC=0.944). Feature elimination experiments demonstrated minimal performance variation across different feature subsets, confirming the model’s capability to effectively identify and mitigate the impact of irrelevant features. Subgroup analyses revealed that the model performed optimally in female subjects (recall=0.835, F1-score=0.838, AUC-ROC=0.950) and individuals aged >69 years (recall=0.913, F1-score=0.875, AUC-ROC=0.911). Conclusion The proposed model based on heterogeneous health examination data can identify high-risk individuals for lung cancer among health examination populations using only routine screening data, thereby facilitating the early diagnosis of lung cancer in this population.