基于概率最小二乘支持向量机的智能医疗决策系统研究

Somatosensory evoked potential (SEP) is a very effective noninvasive monitoring technique, which is widely used in spinal surgery. By comparing the changes of the SEP signal with the SEP baseline, the spinal cord function is monitored in the operation. It is found that SEP baseline was not constant during surgery, and was affected by non operative factors (such as blood pressure, anesthesia, etc.). Therefore, the traditional static SEP baseline monitoring method is prone to false alarm. In clinical practice, the observer dynamically adjusts the SEP baseline to prevent false positives based on personal knowledge and experience. However, the clinical practice of SEP monitoring is limited by the those factors, such as the high cost of training,fatigue and mood of the observer. An effective and reliable decision system is needed for SEP monitoring. In this project, a Probabilistic Least Squares Support Vector Machine based Intelligent Clinical Decision Making System is proposed, including the work of online probabilistic least squares support vector machine model, multimodal based hybrid decision model, the confidence intervals and prediction interval estimation of probabilistic least squares support vector machine. An accurate, fast, reliable and intelligent SEP monitoring system is offered by this project.

体感诱发电位（SEP）是一种非常有效的无创伤性的监护技术，广泛用于脊柱外科手术中。对比手术中SEP信号与基准SEP的幅值变化来监控脊髓功能状况。研究发现SEP基线在手术期间并非恒定，受非手术因素（如血压，麻醉等）影响发生变化。因此，传统静态SEP基线监护方法容易产生误报警现象。在临床实践中，监护人员依据丰富的知识和经验动态调整SEP基线来防止误报。而专业人员的高昂培养费用、疲劳和情绪等各种原因引起的主观性偏差，阻碍了术中SEP脊髓监护方法在临床中的有效普及。急需一种有效可靠的决策系统来对SEP异常进行监控。本项目提出的基于概率最小二乘支持向量机的智能医疗决策系统，包括基于在线概率最小二乘支持向量机的动态时空模型、基于多模型的混合决策模型、概率最小二乘支持向量机的置信区间和预测区间估计三方面的来实现精确、快速、可靠和智能的SEP监护系统。

临床医疗决策问题中，常常需要考虑影响因素的多样性及各因素之间的错综复杂的联系。特别是在患者，手术人员，环境等不确定的因素影响下，建立一个通用的医疗辅助决策模型，是理论者和实践者面临的一个迫切需要解决的问题。本项目采用数据驱动建模方法对临床手术信号监控展开研究。主要研究成果如下：（1）将手术数据的不确定性融入模型当中，考虑到手术监控中的随机特性，构建了一种在线概率最小二乘支持向量机的时空模型用于SEP监控；（2）研究了单台手术数据和整体手术数据之间的关系，通过多模型架构，给出了一种概率权重最小二乘支持向量机预测模型；（3）针对模型预测值的概率特性，通过对期望和方差的合理估计，给出了概率最小二乘支持向量机预测值的置信区间估计。最后，通过实证研究，验证了本项目所提出模型的有效性。