| Sumario: | The global incidence of intracranial aneurysms is increasing annually, and their rupture is associated with a high mortality rate. Many community residents often unknowingly develop intracranial aneurysms and are at risk of rupturing. To solve this problem, we conduct an innovative approach using machine learning to predict both the occurrence and rupture of intracranial aneurysms in patients with brain diseases and analyze the essential features derived from residents' health data at various stages of clinical admission. Specifically, we design an ensemble classifier candidate pool model for the initial two stages of admission diagnosis and a deep fusion network model that integrates textual and structured data for the detailed screening stage. Also, the feature importance is explored by the Shapley value and word frequency. The proposed deep fusion neural network achieves the highest predictive performance, with a precision of 0.787, sensitivity of 0.785, specificity of 0.870, F1 score of 0.785, and AUC of 0.871. In addition, text features contribute most significantly to model output, and word frequency varies across different disease types in patient medical records.
|
|---|