Living Evidence onFirst-EpiSodeof psychosis Treatments and Outcomes (EFESTO): a living individual participant data network meta-analysis and learning

为了克服目前精神病风险检测策略的有限能力，IoPPN的EPIC实验室最近利用了精准医学方法，开发了临床预测模型，包括“跨诊断风险计算器”。该模型对二级精神卫生保健中转变为精神病的概率（即参与者随时间发展为第一种精神病性障碍的风险）进行了索引（1）。该模型基于预先选择的基于临床的变量，这些变量在电子健康记录（EHR）中广泛可用：ICD-10/高危精神状态（ARMS）诊断指数、年龄、性别、性别年龄和种族/民族。该模型被称为“transdiagnosis”，因为它不仅可以预测CHR-P阶段的精神病发作，还可以预测其他非精神病性精神障碍，如情感性精神障碍。事实上，越来越多的证据表明，大多数CHR-P个体也符合共病精神障碍（主要是抑郁和焦虑）的二级诊断标准（2）。此外，双相情感障碍也可能早于精神障碍的发作（1）。跨诊断风险计算器在SLaM Lewisham和Croydon自治市（n= 54，716）的大型成人临床登记队列中进行了外部验证，并显示出良好的外部准确性（Harrell's C=0.79）。进一步的复制证明了它的可移植性。具体而言，跨诊断风险计算器已在不同人群特征的三个独立EHR数据集中进行了外部验证，证明了足够的预后性能。首先，使用卡姆登和伊斯灵顿NHS基金会信托基金（n= 13，702，Harrell's C=0.73）（3），该基金会包括相对于用于开发计算器的人群的老年人群。第二，与牛津健康NHS基金会信托（n= 33，710，Harrell's C=0.79），其中包含城市和农村背景的混合（4）。第三，计算器在IBM商业数据库中进行了外部国际验证（n= 2，430，333，Harrell's C=0.68），尽管在该重复中缺乏种族数据，但仍保持了高于机会的准确性（5），这代表了精神病学临床预测模型的最大规模外部复制。在进一步的步骤中，这个临床预测模型已经通过自然语言处理应用程序进行了改进，以自动筛选临床笔记和信件的文本。这使得能够包括更细粒度的预测因子，例如症状/物质滥用预测因子（流泪、食欲不振、体重减轻、失眠、大麻、可卡因、内疚、易怒、妄想、绝望、睡眠障碍、洞察力差、激动和偏执）。NLP方法基于数据挖掘人工智能，代表了IoPPN提供的定制分析技术。该改进进一步将模型显著改善至Harrell C=0.85（95% CI 0.84-0.86），较原始模型增加0.06。上述研究结果表明，风险计算器提供了一种潜在的工具，可以提高对精神病风险个体的检测。然而，在精神病学中缺乏实施科学，只有不到1%的已发表的临床预测模型最终在临床实践中实施。尽管如此，在IoPPN进行的一项试点研究中测试了跨诊断风险计算器的实际实施情况（7），约77%的临床医生对风险计算器发出的提示做出了回应，表明依从性良好。

To overcome the limited power of current detection strategies for psychosis-risk, the EPIC lab at IoPPN has recently leveraged precision medicine methods and developed clinical prediction models, including the 'transdiagnostic risk calculator'. This model indexes the probability of transition to psychosis (i.e. the risk of the participant developing a first psychotic disorder over time) within secondary mental health care (1). This model was based on preselected clinically based variables which are widely available in electronic health records (EHRs): index ICD-10/At-Risk Mental State (ARMS) diagnosis, age, sex, age by sex, and race/ethnicity. The model was termed "transdiagnostic" because it predicts the onset of psychosis not only from a CHR-P stage but also from other non-psychotic mental disorders such as affective mental disorders. Indeed, there is accumulating evidence showing that the majority of CHR-P individuals would also meet criteria for a secondary diagnosis of co-morbid mental disorder, mostly depression and anxiety (2). Moreover, bipolar mood disorders can also predate the onset of psychotic disorders (1).The transdiagnostic risk calculator was externally validated in a large clinical registry cohort of adults in the SLaM boroughs of Lewisham and Croydon (n=54,716) and showed good external accuracy (Harrell's C=0.79). Further replications have demonstrated its transportability. Specifically, the transdiagnostic risk calculator has been externally validated in three independent EHR datasets of different population profiles, demonstrating adequate prognostic performance. Firstly, with the Camden and Islington NHS Foundation Trust (n=13,702, Harrell's C=0.73) (3), which comprised an older population relative to the one used to develop the calculator. Secondly, with the Oxford Health NHS Foundation Trust (n=33,710, Harrell's C=0.79), whichcontained a mix of both urban and rural backgrounds (4).Thirdly, the calculator was externally validated internationally in the IBM Commercial database (n=2,430,333, Harrell's C=0.68), retaining an accuracy above chance despite the lack of ethnicity data in this replication (5), which represents the largest external replication of clinical prediction models in psychiatry. In a further step, this clinical prediction model has been refined with Natural Language Processing applications to automatically screen the text of clinical notes and letters. This enables the inclusion of more fine-grained predictors such as symptom/substance misuse predictors (tearfulness, poorappetite, weight loss, insomnia, cannabis, cocaine, guilt, irritability, delusions, hopelessness, disturbed sleep, poor insight, agitation, and paranoia). NLP methods are based on data-mining artificial intelligence and represent bespoke analytic techniques available at the IoPPN. This refinement further improved the model significantly to Harrell's C=0.85 (95% CI 0.84-0.86), an increase of 0.06 from the original model. The above findings show that the risk calculator provides a potential tool that could improve detection of individuals at risk of psychosis. However there is a dearth of implementation science in psychiatry and only less than 1% of published clinical prediction models are eventually implemented in clinical practice (6). Despite this, the real-world implementation of the transdiagnostic risk calculator was tested in a pilot study conducted at the IoPPN (7), with about 77% of clinicians respondingto prompts issued by the risk calculator, indicating a good level of adherence.