功能磁共振成像（functional magnetic resonance imaging，fMRI）技术被广泛用于揭示大脑认知加工过程的神经机制。近年来，各种机器学习技术被用于fMRI数据上进行神经疾病的预测研究。
在机器学习方法中，多任务学习方法通过同时学习多个任务共享任务间的信息，提升算法的泛化性能，弥补训练数据不足所造成的模型与真实情况差异过大的问题。由于采集大样本的fMRI数据尤其是病人的fMRI数据具有一定的困难，将多任务学习方法应用在病人的静息态fMRI中将有助于提升fMRI数据的利用率，获得更好的疾病预测准确率，并帮助我们更好的理解疾病的神经机理影响。当前尚未有研究探讨如何利用多任务学习方法对具有轻度认知障碍的脑小血管病（cerebral small vessel disease with mild cognitive impairment，CSVD-MCI）患者的静息态fMRI数据建立多任务分类器，实现对CSVD-MCI的预测。而在多任务学习中，根据数据的特性选择合适的任务是非常关键的。因此本研究将围绕多任务学习分类方法深入探讨如何充分针对静息态fMRI数据的特性，选择合适的任务建立多任务分类器实现对CSVD-MCI预测。我们将不同的静息态脑网络以及大脑功能连接网络的不同拓扑属性分别作为任务学习对象开展了以下两方面研究：
研究一，以不同的静息态脑网络为任务对象，构建多任务稀疏表示分类器进行CSVD-MCI的判别。由于不同的静息态脑网络之间空间激活脑区(spatial map，SM)不同，特征共享方法选择出的体素不适用于所有脑网络，因此本研究采用在样本空间进行学习的多任务稀疏表示分类器。该方法能够利用不同的任务共享被试样本的特性提高分类器的准确率和泛化能力。实验结果表明，以静息态脑网络为任务对象时多任务稀疏表示分类方法相比单任务稀疏表示分类方法可以获得更好的CSVD-MCI分类性能。并且在默认网络、右额顶网络、执行控制网络、视觉网络几个静息态脑网络中，右额顶网络和默认网络对多任务分类准确率有较大的影响。在后默认网络中，MCI病人的左侧角回的同步性降低，且其同步性与行为测试呈现正相关，该区域的变化可能导致了被试认知能力的降低。
研究二，以静息态的大脑功能连接网络中节点不同的拓扑属性为任务对象，构建多任务逻辑回归分类器进行CSVD-MCI的判别，从而利用不同拓扑属性间的互补信息来提高分类准确率。实验结果表明，以拓扑属性为任务对象时，多任务逻辑回归分类器的准确率显著高于单任务逻辑回归分类器，将多个拓扑属性进行联合学习可以有效提高分类效果。其中度属性有着最高的准确率，表明拓扑属性的度具有很好的判别CSVD-MCI的特性。此外，研究还发现了CSVD-MCI病人的额叶的节点中心性增强，其可能是对其他大脑区域认知功能丧失的补偿。
综合两个研究，我们发现利用静息态功能连接网络节点的拓扑属性构建的多任务学习分类器解码性能要高于利用静息态脑网络构建多任务学习分类器，这可能表明静息态功能连接网络节点的拓扑属性比纯粹的静息态网络具有更好的分辨CSVD-MCI的能力。此外两个研究分别发现右额顶网络和额叶脑区的重要性，这可能表明额叶对于区分CSVD-MCI具有重要的作用。

Functional magnetic resonance imaging (fMRI) technology is widely used to reveal the neural mechanism of cognitive processing in the brain. Various machine-learning techniques have been used to predict neurological diseases on fMRI data.
In the machine learning method, multi-task learning improves the generalization performance of the algorithm by learning multiple tasks at the same time, and makes up for the under-fitting of the model caused by insufficient training data. Because it is difficult to collect fMRI data of large sample size, especially from patients, applying multi-task learning methods to the patient’s resting-fMRI will improve the utilization of fMRI data, obtain better disease prediction accuracy, and help us better understand the neurological effects of disease. At present, there is no research on how to use multi-task learning method to establish a discriminant classifier of Mild Cognitive Impairment (MCI) for cerebral small-vessel disease (CSVD) resting fMRI data. In multi-task learning, it is very critical to choose the right task according to the characteristics of the data. Therefore, this study will deeply explore how to fully target the characteristics of resting state fMRI data, select an appropriate task and establish a multi-task classifier to achieve prediction of MCI in patients with CSVD. We carried out the following two aspects of research on different resting brain networks and different topological attributes of brain network functional connections as task learning objects:
Firstly, we construct a multi-task sparse representation classifier to discriminate CSVD-MCI by using the resting brain network as the task object. Due to the low correlation between different resting brain networks, this study uses a multi-task sparse representation classifier that learns in the sample space. This method can use different tasks to share the characteristics of the sample to improve the accuracy and generalization ability of the classifier. Experimental results show that the multi-task sparse representation classification method can obtain better CSVD-MCI classification performance when using the resting brain network as the task object. And among the rest brain networks of Default Mode Network (DMN), right Frontal Parietal Network (rFPN), Executive Control Network (ECN) and Visual Network (VN), the rFPN and DMN have greater impact on the accuracy of multitask classification. In the post-DMN, the activation of angular gyrus of the MCI patient is reduced, and its activation is positively correlated with the behavior test. Changes in this area may lead to a decrease in the cognitive ability of the subject. 
Secondly, we construct multi-task logistic regression classifier to discriminate CSVD-MCI by using the topological attributes of resting brain network functional connection as the task object, and use the complementary information between different topological attributes to improve the classification performance. The experimental results show that the accuracy of the multi-task logistic regression classifier is significantly higher than that of the single-task logistic regression classifier. Joint learning of multiple topological attributes can effectively improve the classification accuracy. The simplest degree attribute has the highest accuracy, indicating that the degree of the network attribute has good characteristics for distinguishing CSVD-MCI. In addition, the study found that MCI patients showed increased nodal centrality in the brain areas of the frontal lobe, which may be a compensation for the loss of cognitive function in other brain regions.
Combining the two studies, we found that the decoding performance of the multi-task learning classifier by using the topological properties of the resting state network nodes is higher than that of using different resting state networks, which may indicate the topological properties of the system have a better ability to distinguish CSVD-MCI than the pure resting state network. In addition, two studies found the importance of the right frontal parietal network and frontal lobe brain area, which may indicate that the frontal lobe has an important role in distinguishing CSVD-MCI.