运用机器学习预测首发精神分裂症长期结局：聚焦阴性症状维度

Schizophrenia is a common and severe mental disease with long-term outcome varied in different individuals. The variables predicted long-term outcome of schizophrenia has multi-class, random and fluctuant characteristics. It is a hotspot to predict the outcome of disease at early stage with clinical, biological characteristics and other factors. The method based on traditional mathematical model has a lower predictive accuracy, and provides the information limited to binary classification. Multi-voxel pattern analysis (MVPA) is based on machine learning-support vector machine theory to analyze data. MVPA can integrate clinical assessments, with cognitive function and neuroimaging data to predict the outcome at individual level. In our early work, a large database on first-episode schizophrenia with longitudinal follow-up study has been built, and some relative research has done. This study is focused on the negative symptoms as the most significant clinical symptom dimension which hinders the long-term functional recovery, and their characteristic biomarkers. PyMVPA (Python MVPA) analysis is conducted to generate model with multi-aspect data at baseline. The model can be optimized according to the enduring negative symptoms and other assessment features from multi-period follow-up study. Meanwhile, sectionalization test optimal model and predictive weight map can also be generated. In the second stage, another new sample will be collected and follow-up. The outcome at individual level can be predicted by the generalization validation of optimal model from first stage. This study will provide more valuable and supportive evidence in schizophrenia translational medicine research, outcome mechanism and new therapeutic targets.

精神分裂症是精神科常见的重性精神障碍，其疾病长期结局具有较大的个体差异，具有多类别、随机性和波动性等特征。疾病初期患者的相关临床及生物学特征、影响因素等可能对疾病预后的预测具有重要的作用，是当前该领域研究的热点之一。传统数学模型预测力差，且只能给出两分类群组水平预测信息。多体素模式分析(MVPA)是基于机器学习理论-支持向量机开发的数模新技术，能整合临床、认知数据及磁共振图像，对单个患者作结局预测。依托于大样本首发精神分裂症资源库和前期研究结果，本研究拟聚焦于精神分裂症阴性症状这一重要的临床症状维度及相关特征性生物学标记，对患者基线时的多维数据整合建模，并将结合独有的、随访后获取的多时点阴性症状迁延等指标，建立分段最优预测模型，获得预测权重图。在第二阶段将新收集独立样本并随访，对前期建立的患者结局预测力予以验证。预期本研究将为精神分裂症转归机制研究提供科学线索，具有重要的临床转化意义。

精神分裂症是精神科常见的重性精神障碍，其疾病长期结局具有较大的个体差异。疾病初期患者的相关临床及生物学特征、影响因素等可能对疾病预后的预测具有重要的作用，是当前该领域研究的热点之一。本课题克服传统数学模型预测力差，且只能给出群组水平预测信息等问题，基于机器学习理论-支持向量机开发的数模新技术，整合临床、认知数据及磁共振图像，依托于大样本首发精神分裂症资源库和前期研究结果，聚焦于精神分裂症阴性症状这一重要的临床症状维度及相关特征性生物学标记，提取已有的基线神经认知功能数据，采用基于树的AdaBoost集成算法进行分类器的训练，建立预测模型，运用层次聚类分析的方法分别计算首发精神分裂症患者和正常对照的神经认知层次，较准确地预测个体是否具有精神分裂症，抑郁障碍或属于正常对照。得出精神分裂症特有的神经认知受损模式有：运动控制速度，感知灵敏度和逆反学习。整合这批患者临床、认知数据及磁共振图像信息。对精神分裂症基线时的多维数据整合建模，并结合随访后获取的多时点阴性症及结局指标，建立最优预测模型，获得预测权重图，发现脑白质结构网络(左右中央前回、左侧额中回、右侧额下回)、代表阴性症状严重度的量表Factor2因子，和一些认知因子对结局预测价值最大，提示脑白质解剖网络异常，阴性症状和认知受损影响精神分裂症结局，部分证实了“脑连接异常”病因假说。课题第二阶段新收集独立样本并随访，对前期建立的患者结局预测力模型按时间进程进行了初步验证及调整。本课题对探索在个体水平作结局预测可能性。为精神分裂症转归机制研究提供科学线索，具有临床转化意义。