Stroke screening is a crucial measure for reducing stroke occurrence, disability, and mortality. However there are numerous risk factors and the limited number of high-risk stroke groups in all screening populations, the screening data is redundant and imbalanced. We propose an improved feature selection algorithm to identify stroke key risk factors and an oversampling MRF-SMOTE algorithm to balance data. Then, a deep reinforcement learning classification model based on the Dueling DQN (Deep Q-network) algorithm is constructed for stroke classification, with the optimized loss function. Benchmark models include KNN, SVM, random forest, and Dueling DQN. Stroke screening data is pre-processed by selecting key risk factors and oversampled by MRF-SMOTE. Experiments show that the optimized Dueling DQN model can classify the pre-processed stroke screening data in terms of accuracy, AUC, precision, and F1-measure, which are 0.8982, 0.96, 0.899, and 0.8981 and improved by 4.37 %, 6.92 %, 3.78 % and 4.04 %, respectively, compared with the existing Dueling DQN.

卒中筛查是减少卒中发生、残疾和死亡率的重要措施。但由于危险因素众多，且所有筛查人群中脑卒中高危人群数量有限，筛查数据冗余且不平衡。我们提出了一种改进的特征选择算法来识别中风关键风险因素，并提出了一种过采样 MRF-SMOTE 算法来平衡数据。然后，构建了基于Dueling DQN（Deep Q-network）算法的深度强化学习分类模型，用于笔画分类，并优化了损失函数。基准模型包括 KNN、 SVM、随机森林和 Dueling DQN。通过选择关键风险因素对中风筛查数据进行预处理，并通过 MRF-SMOTE 进行过采样。实验表明，优化后的Dueling DQN模型可以对预处理后的中风筛查数据进行分类，准确率、AUC、精度和F1-measure分别为0.8982、0.96、0.899和0.8981，分别提高了4.37%、6.92%。与现有的 Dueling DQN 相比，分别提高了 3.78 % 和 4.04 %。