Applying Natural Language Processing to real-world patient data to optimise cancer care

每年有近40万人被诊断出患有癌症，造成超过16.7万人死亡。发病率和死亡率与社会经济因素密切相关，约有19，000例额外的癌症死亡归因于贫困。如何最好地治疗癌症患者的决定是基于通常通过使用临床试验产生的证据。然而，只有一小部分患者参加了这些研究，许多患者群体，如体弱者，有多种医疗问题的人和少数民族，代表性不足。这意味着有很大一部分人口，特别是贫困人口（他们患癌症的比例不成比例），现有证据可能不适用，从而使健康不平等现象长期存在。常规的“真实世界”患者数据，作为正常治疗的一部分收集的关于每个患者的数据，提供了一个机会，在临床试验数据不存在或不存在的情况下提供证据。该项目的愿景是从每一位接受治疗的患者身上学习。使用真实世界数据的人工智能方法可用于了解癌症诊断和治疗的模式，并对医疗创新的影响进行前瞻性评估，但需要对数据进行结构化处理。现代电子医疗记录（EHR）可以以所需格式收集数据。然而，患者信息的历史数据和许多当前来源（例如，门诊患者信件、放射学报告）通常仅作为自由文本医疗记录存在，并且因此需要首先被编码，即结构化。在这个项目中，我们将开发和应用自然语言处理（NLP）技术来从医疗记录中恢复结构化数据。然后，我们将使用这些数据来验证和改进模型，以预测癌症患者的临床结果，并查看患者的癌症治疗经验是否与其结果的临床评估一致，并可能提供头颈癌治疗相关毒性的早期预警。该项目是癌症科学部和计算机科学系/艾伦图灵研究所之间令人兴奋的合作，因此，该项目的学生将受益于接近克里斯蒂NHS基金会信托的临床团队，欧洲最大的单一站点癌症中心，以及大学的数据科学专业知识。

Nearly 400,000 people are diagnosed with cancer each year, causing more than 167,000 deaths. Incidence and mortality are strongly associated with socioeconomic factors, with around 19,000 extra cancer deaths attributed to deprivation. Decisions on how best to treat cancer patients are based on evidence which is usually generated through the use of clinical trials. However, only a small fraction of patients participate in these studies, and many patient groups such as the frail, those with multiple medical problems, and ethnic minorities, are under-represented. This means there are large sections of the population, particularly the deprived (who suffer disproportionately from cancer), where the available evidence might not apply, perpetuating health inequalities. Routine 'real-world' patient data, collected about every patient as part of their normal treatment, offers an opportunity to provide evidence where clinical trial data doesn't or will not exist. The vision of this project is to learn from every patient treated.Artificial Intelligence approaches using real-world data can be used to understand patterns in cancer diagnosis and treatment, and to provide prospective assessment of the impact of healthcare innovations, but need the data to be structured to enable its processing. Modern Electronic Healthcare Records (EHRs) can collect data in the required format. However, historical data and many current sources of patient information (e.g. out-patient letters, radiology reports) often exist only as free-text medical notes, and therefore needs to be coded, i.e. structured, first. In this project, we will develop and apply Natural Language Processing (NLP) technologies to recover structured data from medical notes. We will then use these data to validate and improve models to predict cancer patients' clinical outcomes, and to see if patients' experience of their cancer treatment agrees with clinical assessments of their outcome and might provide early warning of evolving treatment related toxicity in head and neck cancer. This project is an exciting collaboration between the Division of Cancer Sciences and the Department of Computer Science/Alan Turing Institute, and as such the project student will benefit from close proximity to the clinical teams at The Christie NHS Foundation Trust, the largest single site cancer centre in Europe, and data science expertise at the University.