基于数据的稀疏规则学习方法在全方向M型心动图系统辅助诊断中的应用研究

在统计学习领域，过去的几十年里开发了很多"高性能"的算法，用这些算法设计的预测模型大都具有很好的分类性能。但这些模型通常是"黑盒子"运算，这使得它很难为用户去解释所预测结果的产生原因和过程。在很多实际应用中，特别是医学领域，对于疾病诊断的结果来说，解释性和准确度同样重要。本课题要研究一种能够实现在预测性能和解释性之间达到"最佳"权衡效果的新算法。课题组将从基于规则的预测系统入手：1)研究利用随机森林和稀疏规则学习方法来自动抽取小规模决策规则，希望通过稀疏规则选择小规模决策规则和特征子集同时保持良好的预测性能。2)研究该算法的可靠性和鲁棒性，特别是学习过程中随机性问题的产生以及训练数据噪声对决策规则的影响。3)探索该算法的可视化问题，特别是决策过程和决策结果包括决策规则的可视化。4)探索在大数据集下算法的规则动态获取和更新机制。最后用该算法开发全方向M型心动图系统辅助诊断模块并进行临床测试。

In statistical learning, there are many "high performance" learning algorithms that have been developed in the past decades. These methods typically produce prediction models that have high generalization performance. Nevertheless, the models are usually "black boxes" that make it difficult for a user to interpret why and how does a prediction occur. In many practical applications, especially those in the medical domain, interpretability is as important as prediction accuracy. In this study, the research team will develop new learning algorithms to find an "optimal" trade-off between the prediction accuracy and model interpretability. In particular, the team will focus on rule-based prediction systems and investigate the following novel research problems.1）We will use random forest (RF) and sparsity regularized learning to automatically extract a small set of decision rules. Random forest is a very competitive predication model but lacks of interpretability due to the large number of decision rules that the random forest defines. We propose to use sparsity regularization to select a small number of decision rules and a subset of features from a random forest and to maintain the prediction performance comparable.2）We will investigate stability and robustness of the proposed learning method. Specifically, we will study how the randomness introduced by the RF learning process and how the noise in the training data would affect the selected decision rules.3）We will also explore model visualization issues. Specifically, the visualization of the decision making/learning process and the decision rules.4）We will also explore the dynamic accessing and updating mechanism to the rules of the algorithm based on big data set.Finally, the team will apply the results of the above research to Omni-directional M-mode echocardiography data, which are mainly used in clinical echocardiography analysis.