Predicting psychosis-onset through online assessment of speech organisation

精神病影响了大约0.7%的英国人口（NICE 2016），这是一种精神健康状况，患者对世界的感知和解释如此不同，以至于他们与现实失去了联系，并可能出现幻觉、妄想和认知困难等症状。这显著影响了个人的生活质量和日常功能，失业的可能性是一般人群的6-7倍（NICE 2015），近三分之一的人在某种程度上无家可归（Bebbington et al., 2005）。超过一半的首发精神病（FEP）患者会在三年内复发（Alvarez-Jimenez et al., 2012）。据估计，仅英格兰每年就需要花费118亿英镑（使用2012年的数据），其中包括NHS护理等直接成本和无偿护理等间接成本（Ride et al., 2020）。精神病的发病通常发生在成年早期，然而，精神病的检测延迟越长，因此延迟开始治疗，长期结果就越负面（Oliver等人，2018）。因此，越来越多的人关注早期发现精神病，这导致确定精神病群体的临床高风险（chrp），以识别潜在的精神病前驱期的人（Fusar-Poli, 2017）。这为预防或延迟精神病的发作提供了一个关键机会。然而，chrp组之间的临床结果存在很大的异质性，该组中约有20%的人后来发展为精神障碍，很难预测该组中谁会过渡（Fusar-Poli等人，2012年和2016年）。改善这一状况的一种方法是基于已知的精神疾病潜在机制开发预后标记物，以帮助对这些chrp组进行分层，以便更好地指导早期干预和治疗分配给那些最受益的人。形式思维障碍（FTD）是一种精神病的认知症状，人们的思维方式混乱，已被证明在前驱期以减弱的形式出现（Morgan et al., 2021）。由于混乱的思维会导致可以分析的异常语言，言语可能是一个潜在的预后标记。Mota et al.（2012）已经证明语音可以基于图论进行量化，其中每个词是一个节点，其边缘对应于语义和语法连接。例如，这个图结构可以通过它的“连通性”来分析。使用这种技术，Mota等人能够在FEP患者的小样本中提前6个月预测精神分裂症的诊断（2017年）。在此基础上，该项目的主管已经表明，在研究语义网络（Nettekoven等人，2023）和非语义语音网络（Spencer等人，2021）时，可以区分chrp， FEP和健康对照，并可用于预测临床结果（Morgan等人，2021）。重要的是，语音采样是非侵入性的，可以使用移动电话等廉价设备进行远程测量。这样就可以收集到大量的数据集，并且考虑到大多数目标人群已经拥有手机，这将有助于减少抽样偏差。

Psychosis affects approximately 0.7% of the UK population (NICE 2016), a mental health condition where an individual perceives and interprets the world so differently they have lost touch with reality and may experience symptoms such as hallucinations, delusions, and cognitive difficulties. This significantly impacts quality of life and everyday functioning with individuals being 6-7 times more likely to be unemployed than the general population (NICE 2015) and nearly a third being homeless at some point (Bebbington et al., 2005). Over half of individuals experiencing first-episode psychosis (FEP) will later relapse within three years (Alvarez-Jimenez et al., 2012). This is estimated to cost England alone £11.8 billion a year (using 2012 data) in direct costs such as NHS care and indirect costs such as unpaid care (Ride et al., 2020).Psychosis-onset typically occurs in early adulthood, however the longer the delay in detection of psychosis, and thus the delay beginning treatment, the more negative the long-term outcomes (Oliver et al., 2018). Therefore, there is increasing focus on detecting psychosis earlier, which has led to identifying clinical high-risk for psychosis groups (CHR-P), to recognise people potentially in the prodromal phase of psychosis (Fusar-Poli, 2017). This allows a key opportunity to prevent or delay the onset of psychosis. However, there is large heterogeneity in clinical outcomes amongst CHR-P groups, with around 20% of this group later developing a psychotic disorder and it is difficult to predict who within this group will transition (Fusar-Poli et al., 2012 and 2016). One way to improve this is to develop prognostic markers based upon mechanisms known to underlie psychosis, to help stratify these CHR-P groups to better direct early intervention and treatment allocation for those who will most benefit.Formal thought disorder (FTD), a cognitive symptom of psychosis where people have a disorganised way of thinking, has been shown to appear in an attenuated form in the prodromal phase (Morgan et al., 2021). Since disorganised thinking leads to abnormal language that can be analysed, speech could be a potential prognostic marker. Mota et al. (2012) have shown that speech can be quantified based upon graph theory, whereby each word is a node and their edges correspond to the semantic and grammatic connections. This graph structure can be analysed for example, by its "connectedness". Using this technique, Mota et al. were able to predict a schizophrenia diagnosis up to 6 months in advance in a small sample of FEP patients (2017). Building upon this, the supervisors of this project have shown that it is possible to differentiate between CHR-P, FEP and healthy controls when studying semantic networks (Nettekoven et al, 2023) and non-semantic networks of speech (Spencer et al., 2021), and can be used to predict clinical outcomes (Morgan et al., 2021).Importantly speech sampling is non-invasive and could be measured remotely using inexpensive devices such as mobile phones. This enables large datasets to be collected and given that most of the target population already owns mobile phones, this would help to reduce sampling bias.