Methodology to support the development of prognostic models that incorporate and inform the observation processes within electronic health records

人们对利用患者电子健康记录（EHRs）提供的大量数据来开发临床预测模型（cpm）非常感兴趣，该模型可以根据我们现在对患者的了解来预测一个人在未来某个时间点是否会发生某种事件[1,2]。cpm可以支持基于证据的决策和最近对精准医疗的推动，但在电子病历中建模时仍然存在挑战。具体来说，电子病历通常包含患者的纵向信息，通过与卫生服务机构的反复接触，但观察时间和测量频率在患者内部和患者之间会有很大差异。例如，假设CPM使用实验室结果和血压来预测患者被转移到重症监护病房（ICU）的实时风险，从而促进定期医院就诊期间的早期预警。虽然患者的电子病历可能包含常规的血压测量，但他们的实验室结果可能不太频繁地被观察到。该项目将开发方法，使CPM能够在没有风险因素的情况下进行实时预测（例如，仅使用血压进行预测），然后可能“要求”某些患者提供额外信息，而不是试图在给定时间点推断缺失的风险因素（例如，实验室结果）。这种“互动测量”将有助于在那些有高危不良结果（如病情恶化转至ICU）的患者中有针对性地收集高频数据。同样，如果附加信息（例如实验室结果）是先验的，那么这将反映临床医生对患者的看法，因为观察频率包含了关于患者潜在健康状况的信息——即所谓的信息性观察[3]。在cpm中纳入观察过程和结果过程在cpm中很少被考虑，但可以使他们从临床判断中学习[4,5]。因此，本项目将：(i)探索目前在CPM开发中纳入观察过程的方法，（ii）开发允许CPM处理异构风险因素测量频率的方法，同时为交互式测量提供信息，以及（iii）研究交互式测量和信息观察之间的关系。对于临床例子，我们打算关注：1)重症监护（医院）环境下的出院与护理升级，以及2)初级保健中疾病发病率的预测。

There is substantial interest in using the abundance of data available through patients' electronic health records (EHRs) to develop clinical prediction models (CPMs) that predict whether a person will have an event at some future time point based on what we know about them now [1,2]. CPMs can support both evidence-based decision-making and the recent drive for precision medicine, but challenges remain when modelling within EHRs. Specifically, EHRs usually contain longitudinal information about a patient, through repeated contact with health services, but the observation times and frequency of measurement will vary considerably within and across patients. For example, suppose a CPM uses lab results and blood pressure to predict the real-time risk of a patient being transferred to an intensive care unit (ICU), thereby facilitating early warning during regular hospital visits. While a patients' EHR might contain regular blood pressure measurements, their lab results might be observed less frequently. Rather than trying to impute missing risk factors (e.g. lab results) at a given time point, this project will develop methods that allow the CPM to make real-time predictions in their absence (e.g. predict using only blood pressure), and then potentially 'request' additional information for certain patients. Such 'interactive measurement' would facilitate targeted high-frequency data collection in those at high-risk of adverse outcome (e.g. deterioration towards ICU transfer). Likewise, if the additional information (e.g. lab results) were available a-priori, then this would reflect the clinician's beliefs about a patient since the observation frequency contains information on the patient's underlying health status - so-called informative observation [3]. Incorporating the observation process and the outcome process within CPMs is rarely considered within CPMs, but could allow them to learn from clinical judgements [4,5]. Therefore, this project will: (i) explore the current approaches of incorporating observation processes within CPM development, (ii) develop methods that allow CPMs to handle heterogeneous risk factor measurement frequency, while concurrently informing interactive measurements, and (iii) study the relationships between interactive measurement and informative observation.For clinical examples, we intend to focus on: 1) discharge versus care escalation in a critical care (hospital) context, and 2) prediction of disease incidence in primary care.