Secondary use of EMRs for surgical complication surveillance

项目摘要 手术后并发症(PSCs)已经成为医院日益关注的问题，特别是在 关注更长病程的支付改革，以及针对30天再入院和不良反应的医疗保险处罚 结果包括深部或器官间隙手术部位感染(DOS-SSI)。因后再入院- 出院并发症尤其成为质量改进的目标，因为许多这样的事件 被认为是可以预防的。电子健康记录(EHR)的广泛采用导致了许多 PSCs的临床风险模型。这些建模工作主要针对外科专业 发生大量事件的领域(如结直肠手术)以及应用 复杂的统计建模/机器学习，允许丢失数据、交互和非线性。 然而，在准确性和普适性方面仍有相当大的改进空间。在我们的 在目前的资助期内，我们已经证明了临床笔记对PSCs的预测价值。然而，有一个 当前模式的明显局限性是，他们在大专院校接受的是大批量外科专业的培训 护理机构拥有高质量的临床数据和使用先进的信息学方法。这其中的推动力 建议基本上有两个方面：(I)可以为规模较小的机构和 通过转移学习和通过未经确认的结果利用更多数据(即，模仿的结果)的专业 黄金标准结果，但可靠性较低)，并对可靠性进行适当核算。(Ii)决策可以是 通过利用实验室、生命体征和临床记录等随时间变化的实时数据显著改进 使用所有可获得的信息为患者提供PSCs的当前风险。我们的目标是i)发展 并应用PSC的纵向风险模型来明确地解释一些 实时可用的信息(例如，实验室、生命体征、临床记录)，以便将其集成到 临床医生的决策；ii)开发转移学习并将其应用于PSC风险模型；iii)开发模型 允许使用更广泛可获得的未经确认结果的方法，同时明确说明 对于由于使用这种未经确认的结果而导致的额外不确定性和偏差，而不是 现有的黄金标准；以及四)制定一个广泛适用的模式评估和监测框架。 由于各种原因，模型在实践中往往不会像在研究中那样表现。这一框架将 使我们能够识别这些问题，并更高效地将这些复杂的预测模型转换和应用到 实践，使研究能够立即产生临床影响。成功的开发将开启 下一代患者监控、警报和干预的大门，适用于所有外科专科 机构。我们将公开相关的建模结果，以便较低数量的机构可以 利用这里开发的迁移学习方法，而不需要我们的实际数据。这将是 通过及早发现并发症和限制并发症，最终改善患者护理并降低总成本 由于梅奥诊所和全国其他机构的PSCs再次入院。

Project Summary Post-surgical complications (PSCs) have been an increasing concern for hospitals, particularly in light of payment reform focusing on longer episodes and Medicare penalties for 30-day readmissions and adverse outcomes including deep or organ space surgical site infections (DOS-SSIs). Readmission due to post- discharge complications in particular has become a target for quality improvement since many of these events are considered preventable. The wide adoption of electronic health records (EHRs) has led to a number of clinical risk models for PSCs. These modeling efforts have primarily been targeted at the surgical specialty areas within which a large number of events occur (such as colorectal surgery) as well as applying sophisticated statistical modeling / machine learning to allow for missing data, interactions, and nonlinearities. However, there is still considerable room for improvement both in terms of accuracy and generalizability. In our current funding period, we have demonstrated the predictive value of clinical notes for PSCs. However, one glaring limitation of current models is that they are trained on high volume surgical specialties at large tertiary care institutions with high quality clinical data and use of advanced informatics approaches. The impetus of this proposal is essentially two-fold: (i) Accurate models can be created for lower volume institutions and specialties via transfer learning and leveraging more data via unconfirmed outcomes (i.e., those that mimic gold standard outcomes, but are less reliable) with proper accounting of reliability. (ii) Decision making can be significantly improved by leveraging time varying, real-time data such as labs, vitals, and clinical notes to provide the current risk of PSCs for patients using all information as it becomes available. We aim to i) develop and apply longitudinal risk models for PSCs to explicitly account for the time varying nature of some of the information (e.g., labs, vitals, clinical notes) as it becomes available in real-time so that it can be integrated into the clinician’s decision making; ii) develop and apply transfer learning to PSC risk models; iii) develop modeling approaches that allow for the use of more widely available unconfirmed outcomes, while explicitly accounting for the additional uncertainty and bias due to the use of such unconfirmed outcomes when compared to a less available gold standard; and iv) develop a widely applicable framework for model evaluation and monitoring. Models will often not perform in practice as they do in research for a variety of reasons. This framework will allow us to identify these issues and more efficiently translate and apply these complex predictive models into practice so that the research can have an immediate clinical impact. Successful development would open the door for next generation patient monitoring, alerts, and interventions for all surgical specialties and all institutions. We will make the relevant modeling results publicly available so that lower volume institutions can leverage the transfer learning approach developed here without the need for our actual data. This will ultimately lead to improved patient care and lower overall cost by identifying complications early and limiting readmission due to PSCs at Mayo Clinic and other institutions across the nation.