Finding psychosis subtypes using machine learning, clinical, genetic and multimodal imaging data

这个项目将解决精神神经科学中的一个关键挑战：我们能否改变诊断界限，以反映精神疾病的潜在神经生物学亚型？最近，《精神疾病诊断和统计手册》(DSM-V)--精神疾病的诊断标准清单--受到了许多负面宣传，许多人批评精神疾病的诊断过于专断，只基于症状，而不是潜在的神经生物学或其他标记物。尽管这些说法有一定的道理，但目前还不清楚应该用什么来取代DSM-V。这项研究将使用机器学习技术(算法可以在非常复杂的-例如成像-数据中找到相似性，并将类似的受试者分组成簇)，以在由诊断为精神分裂症、分裂情感障碍和双相情感障碍的受试者组成的大型数据集中寻找精神障碍的亚型。数据本身来自大脑成像(功能磁共振成像和脑电)实验，以及心理和症状测量。尽管许多研究小组目前正在使用机器学习试图找到精神疾病的新亚型，但这种方法的不利之处可能是，这些算法用来区分疾病亚型的复杂数据特征无法从生物学意义上解释。例如，仅仅通过观察脑电(EEG)数据的不同模式是不可能知道它们是否是由特定类型的神经元或感受器的功能障碍引起的。我们可以尝试通过使用神经元及其受体如何产生大脑成像数据的模型来解决这个复杂的问题：从这些模型和我们拥有的成像数据中，我们可以估计特定细胞类型和受体的功能或功能障碍。通过在我们所有的受试者中这样做，并将这些数据添加到机器学习分类器中，我们希望获得在生物学意义上可以解释的精神障碍的亚型。拥有可以用生物学术语理解的疾病亚型的好处是显而易见的：它们应该导致更好地研究每种疾病亚型的不同原因，并更好地开发针对亚型的治疗方法。

This project will address a key challenge in psychiatric neuroscience: Can we change diagnostic boundaries to reflect underlying neurobiological subtypes of mental illnesses? Lately, the Diagnostic and Statistical Manual of Mental Disorders (DSM-V) - the list of diagnostic criteria for psychiatric disorders - has received a lot of negative publicity, with many criticising psychiatric diagnoses for being too categorical and only based on symptoms, and not on underlying neurobiology or other markers. Whilst there is some truth to these claims, it remains very unclear what should replace the DSM-V. This study will use machine learning techniques (algorithms that can find similarities in very complex - e.g. imaging - data and group similar subjects into clusters) to find subtypes of psychotic disorders in large datasets comprised of subjects with diagnoses of schizophrenia, schizoaffective and bipolar disorders. The data themselves come from brain imaging (functional magnetic resonance imaging and electroencephalography) experiments, and psychological and symptom measures. Although numerous research groups are currently using machine learning to try to find new subtypes of psychiatric disorders, the downside of this approach can be that the complex data features that these algorithms use to differentiate between illness subtypes are not interpretable in a biological sense. For example, it would be impossible to know just from looking at distinct patterns of electroencephalography (EEG) data whether they were caused by a dysfunction in a specific type of neuron or receptor. We can try to solve this complex problem by using models of how neurons and their receptors generate brain imaging data: from these models, and the imaging data we have, we can estimate the function or dysfunction of specific cell types and receptors. By doing this in all our subjects, and adding this data to the machine learning classifier, we hope to obtain subtypes of psychotic disorders that are interpretable in a biological sense. The advantages of having illness subtypes that can be understood in biological terms are clear: they ought to lead to better research into the distinct causes of each illness subtype, and better development of subtype-specific treatments.