Abstract: In consideration of redundancy in stroke screening datasets and poor effect of the traditional classification algorithm, in order to realize efficient diagnosis and prediction of stroke screening data, this paper establishes an optimization model of classification and prediction based on deep reinforcement learning with mixed feature dimension reduction. Firstly, an improved feature selection algorithm for CFS combined with PCA is proposed to reduce the feature dimension of the original stroke screening datasets. Secondly, a deep reinforcement learning classification prediction model is constructed based on Double DQN and Dueling DQN algorithms, and a more robust loss function is introduced to optimize the model, which improves the classification effect of the model. Finally, compared with the experimental results of Na?ve Bayes, J48, SVM, KNN and DQN models in public datasets and stroke screening datasets, the proposed model exhibits superiority in feature dimension reduction and classification prediction, and its classification accuracy is higher than that of the comparison algorithm in stroke screening datasets. It can be used to provide suggestions for auxiliary diagnosis of cerebral apoplexy.