Unlocking mental health records at scale using few-shot AI

**概要**Akrivia正在使用人工智能（AI）研究领域的一项新创新来释放电子健康记录（EHR）数据的潜力。Akrivia管理着世界上最大的精神病学电子病历数据库，代表16家NHS医疗机构（hco）安全地管理着400多万名患者的未识别数据。Akrivia的目标是利用这一独特的资源来改变精神疾病和痴呆症的研究，推动治疗发现并降低试验成本。然而，约85%的英国精神病学电子病历的可操作数据存储在非结构化的“自由文本”笔记中，由于其个人信息水平，这些笔记难以大规模分析，非nhs研究人员完全无法访问。“自然语言处理”（NLP）可以为这种不可访问性提供解决方案，Akrivia开发了一种基于人工智能的NLP系统，可以提取有关药物、症状、诊断等的新结构化数据。像Akrivia这样的NLP模型可以达到很高的准确性，但传统上受限于需要大量人工注释的训练数据。创建这些训练数据需要很长时间，这意味着Akrivia目前的NLP解决方案无法满足其用户群的需求。作为回应，Akrivia开发了一种替代的原型NLP解决方案，使用了2021年发表的一种新颖的“少量”训练方法。这种模型只需要很少的训练数据就能达到较高的单任务性能。Akrivia已经使用他们的原型模型取代了标准NLP开发管道中的人类注释器，以更短的生产时间（每个概念约4周，而不是约6-9个月）实现了相同（或更好）的任务准确性。通过这个项目，Akrivia将创建工具，以比以前更快的速度扩展其NLP开发，同时显着减少敏感患者数据的暴露。该工具包将允许没有技术人工智能专业知识的临床医生直接开发模型，创建一个“研究人员在循环”的解决方案，以确保Akrivia的NLP库嵌入专家领域知识。该工具包还将在定制的NLP解决方案中开辟一条潜在的新服务线。Akrivia希望为研究精神疾病和痴呆症的研究人员和临床医生提供尽可能广泛和深入的数据集。这些疾病复杂、昂贵，而且历来缺乏资金和治疗选择。具有深入、广泛的疾病状态描述的大规模患者数据对于肿瘤等其他领域的药物开发和有效治疗提供至关重要。通过这个项目，Akrivia将开发工具，使精神疾病和痴呆症的可比数据在几个月内成为现实。

**Summary**Akrivia is using a new innovation from the field of artificial intelligence (AI) research to unlock the potential of electronic health record (EHR) data. Akrivia curates the world's largest database of psychiatric EHRs, with 4 million+ patients' deidentified data managed securely on behalf of 16 NHS healthcare organisations (HCOs).Akrivia's goal is to use this unique resource to transform mental illness and dementias research, driving treatment discovery and reducing trials costs. However, ~85% of UK psychiatric EHRs' actionable data is stored in unstructured, 'free-text' notes, which are difficult to analyse at scale and completely inaccessible to non-NHS researchers due to their level of personal information.'Natural language processing' (NLP) can provide a solution to this inaccessibility, and Akrivia has developed an AI-based NLP system to extract new structured data on medications, symptoms, diagnoses etc. NLP models like Akrivia's can achieve high accuracy, but are traditionally limited by the need for large amounts of human-annotated training data. Creating this training data takes a long time, meaning that Akrivia's current NLP solution is not scalable enough meet the demands of their user base.In response, Akrivia has developed an alternative, prototype NLP solution using a novel 'few-shot' training method published in 2021\. This few-shot model requires very little training data to achieve high single task performance. Akrivia has used their prototype model to replace human annotators in their standard NLP development pipeline, achieving equal (or better) task accuracy with far shorter time-to-production (~4 weeks per concept versus ~6-9 months).**Vision**With this project, Akrivia will create the tools to scale their NLP development far faster than previously possible, with significantly less exposure of sensitive patient data. The toolkit will allow clinicians without technical AI expertise to develop models directly, creating a 'researcher-in-the-loop' solution to ensure Akrivia's NLP library embeds expert domain knowledge. The toolkit will also open a potential new service line in bespoke NLP solutions.Akrivia wants to provide researchers and clinicians working on mental illness and dementias with as broad and deep a dataset as possible. These diseases are complex, costly, and historically lacking in funding and treatment options. Large scale patient data with deep, broad descriptions of disease states has been critical to drug development and effective therapy provision in other areas like oncology. Through this project, Akrivia will develop the tools to make comparable data for mental illness and dementias a reality within a matter of months.