正常成人大脑通过左偏侧化的语言网络感知语音。然而，偏侧化的语言网络及其功能如何在早期听觉经验中产生仍然是语言神经科学领域亟待解决的重要科学问题。正常人群的早期听觉经验开始于孕中期，导致其对应的脑功能活动及功能网络的发育难以测量。人工耳蜗植入的先天聋儿为研究早期听觉脑功能发育提供了一个独特的研究窗口。通过神经影像技术采集的脑功能活动及其网络发育数据具有高维性，使用传统单变量分析方法研究其相互间的关系，及与行为发展之间的关系具有一定局限性。为此，本论文以人工耳蜗植入聋儿为研究对象，使用可以有效处理高维数据的机器学习方法，探究生命早期听觉及语言的脑功能发育机制。

本论文使用近红外功能成像技术对人工耳蜗植入聋儿的脑功能发育进行了纵向追踪，在开机后约一年内完成至少两次重复测量，使用机器学习方法对脑、听觉及语言行为发展的关系进行分类分析。首先，研究发现人工耳蜗植入聋儿术后约一年内，双侧语言网络的功能活动发育，特别是左半球颞叶前端的功能活动发育，可以较为准确地区分高行为发展组和低行为发展组儿童。此外，语言网络的功能连接发育也具有类似的行为发展分类能力，其中半球间非同位区连接有显著的分类贡献。进一步，我们首次使用机器学习方法探究了功能网络连接与其功能活动发育之间的关系。研究发现，静息及倾听状态下的语言网络发育，可预测左半球功能活动发育，特别是静息语言网络可预测腹侧通路的噪音内语音功能活动发育，以及背侧通路的语音功能活动发育。不对称半球间连接在这种预测中都起到了重要作用。综合上述研究，不对称的半球间信息交流在语言网络及其听觉功能的早期发育中具有重要作用，这揭示了语言网络的早期发育机制，为探究具有左半球优势的颞额网络在语言发育早期的形成过程提供了新的信息。这一结果也为探究不同神经机制发育过程中的关系提供了基于机器学习的方法示范。

此外，本论文还比较了使用不同特征选择分析框架进行脑发育研究的差异。研究发现，相比于嵌套交叉验证框架，基于整个数据集进行先验特征选择的方式会导致对分类准确率的过高估计，使分类结果存在一定的偏差。这为使用机器学习方法研究脑发育提供了特征选择分析框架选择方面的指导。

The mature human brain perceives speech through a bilateral frontotemporal network with left-hemisphere dominance. However, how the lateralized language network and its functions emerge during early auditory experiences remains a key question of language neuroscience. Auditory experiences in normal hearing populations begin in mid-pregnancy, making it difficult to measure functional activity and network development. Congenitally deaf children with cochlear implants (CI) provide a unique opportunity to examine brain function development with early auditory experiences. The high dimensionality of neuroimaging data also poses significant challenges to the statistical methods conventionally used to analyze neuro-behavioral relationships. Therefore, this study employed machine learning methods effective in handling high-dimensional data to explore the behavioral and brain functional development of CI children.

Functional near-infrared spectroscopy was used to longitudinally track the brain functional development of CI children, with at least two valid repeated measurements within about one year of CI activation. Machine learning classification methods were used to examine the relationship between brain and behavioral development. First, the study found that the development of functional activity in the bilateral language network within about one year after surgery can distinguish children with low and high behavioral improvement, with prominent contribution from the left anterior temporal lobe. Functional connectivity development in the language network demonstrates similar behavior classification ability, with significant contributions from asymmetrical inter-hemispheric connections. Furthermore, we applied machine learning methods to exploring the relationship between the development of functional network connectivity and its cortical processing. The study found that the development of the language network at resting state and passive listening can classify development of auditory cortical processing in the left hemisphere, particularly speech-in-noise processing along the ventral auditory stream and phonological processing along the dorsal pathway, to which asymmetric inter-hemispheric connectivity contributed significantly. Overall, the results suggest that asymmetrical inter-hemispheric communication drives early development of the language network and its functions, shedding light to the formation process of the lateralized temporo-frontal network with early auditory experiences. The study also provides a first and successful demonstration of using machine learning methods to untangle the developmental relationship among different neural mechanisms.

In addition, this study also compared different feature selection procedures using the same dataset, finding that compared to nested cross-validation blind to the test set, a priori feature selection performed on the entire dataset can result in overestimation of classification accuracy. This result provides reference for feature selection procedure in application of machine learning methods to brain imaging data.