Leveraging EHR Information to Measure Pressure Ulcer Risk in Veterans with SCI

描述（由申请人提供）： 压疮（PrU）是脊髓损伤（SCI）退伍军人在生活质量和护理成本方面最重要的并发症之一。目前使用的风险评估工具，布雷登量表，遭受了一些限制。(1)Braden量表不是SCI特异性的，并且在SCI人群中具有严重的天花板效应。2008年至2010年在22个VA SCI/D中心进行的VA外部同行评审计划（EPRP）调查发现，平均95.3%的患者在不耐烦入院后24小时内接受了Braden评分，其中91.3%的患者被确定为高风险。(2)一些有研究证据或强有力的临床支持的风险变量在现有的评估工具中没有得到很好的体现。(3)目前的风险评估工具主要是为住院患者开发的。然而，在急性损伤后，大多数SCI患者在医院外获得PrU。根据EPRP在2008年至2010年期间的调查，2010年入住VHA SCI/D中心的退伍军人中，只有不到2%的人发展了医院获得性PrU。这项研究的目的是：1）开发自然语言处理（NLP）程序，以识别PrU的发生; 2）根据现有的结构化数据开发PrU发生的预测模型，以便对PrU风险评估产生早期影响; 3）开发NLP程序，以可靠地从临床记录中的文本中提取有关潜在预测因素的信息; 4）开发NLP程序，以识别PrU的发生。4）结合通过结构化和文本提取的NLP数据获得的联合收割机风险信息，并开发预测PrU的稳健风险评估。项目方法：这是一项对脊髓损伤退伍军人的回顾性队列研究。队列的开始包括2009财年在VHA接受治疗的所有SCI退伍军人，他们在过去12个月内没有压疮记录。将由专家小组审查文献中确定的潜在风险因素（例如人口统计学、疾病状态、合并症、健康行为、心理社会因素、家庭护理）的逻辑一致性、完整性和临床相关性。将对EHR进行审查，以确定识别的风险因素是否存在于结构化数据（在数据库/表格中编码）或叙述性数据（临床记录中的文本）中。将通过VA信息学和计算基础设施（VINCI）获得目标队列或2009 - 2013财年（预期最新可用数据）的所有结构化和叙述性数据。在目标1中，我们将使用2X2（卡方检验）频率表比较基于NLP的PrU发生率与基于结构化（ICD-9-CM）数据的PrU发生率。在目标2中，将开发基于现有结构化数据的PrU发生预测模型，并与基于Braden量表的预测进行比较。在目标3中，将开发NLP系统，以从EHR文本中提取风险因素。在基于新风险模型的Aim 4预测中，将结构化数据与NLP数据相结合的预测与仅基于结构化数据和Braden量表的预测进行比较。将使用多变量逻辑回归模型开发预测模型。

DESCRIPTION (provided by applicant): Pressure ulcers (PrU) are among the most significant complications in Veterans with spinal cord impairment (SCI) in terms of quality of life and cost of care. The currently used risk assessment tool, the Braden Scale, suffers from a number of limitations. (1) The Braden scale is not SCI-specific and has severe ceiling effects in the SCI population. VA External Peer Review Program (EPRP) surveys, conducted between 2008 and 2010 at 22 VA SCI/D Centers, found on average 95.3% of all patients received a Braden score within 24 hours of an impatient admission with 91.3% of those measured identified as being at high risk. (2) Some risk variables for which there is research evidence or strong clinical support are not well represented in existing assessment tool. (3) The current risk assessment tool was primarily developed for use in the inpatient setting. However, after the acute post-injury period, most individuals with SCI acquire PrUs outside the hospital. Based on the EPRP survey between 2008 and 2010, less than 2% of Veterans admitted to VHA SCI/D Centers in 2010 developed hospital acquired PrUs. The Aims for this study are: 1) Develop natural language processing (NLP) programs to identify the occurrence of PrUs; 2) Develop predictive models of occurrence of PrUs based on available structured data for early impact on PrU risk assessment; 3) Develop NLP programs to reliably extract information about potential predictors from text in clinical notes; 4) Combine risk information obtained through structured and text- extracted NLP data, and develop robust risk assessment predictive of PrUs. Project Methods: This is a retrospective cohort study of Veterans with SCI. The inception of the cohort includes all Veterans with SCI cared for in the VHA in FY 2009 that had no record of a pressure ulcer in the previous 12 months. Potential risk factors (e.g. demographics, diseases status, co-morbidities, health behaviors, psychosocial factors, home care) identified in the literature will be reviewed by an expert panel for logical consistency, completeness and clinical relevance. Review of the EHR will be conducted to determine if the identified risk factors are found in structured (coded in database/table) or in narrative data (text in clinical notes). All structured and narrative data for the targeted cohort or FY 2009-2013 (anticipated most recent data available) will be obtained through the VA Informatics and Computing Infrastructure (VINCI). In Aim 1 we will use 2X2 (Chi-square test) frequency tables to compare the rates of PrU occurrence based on NLP with those based on structured (ICD-9-CM) data. In Aim 2 predictive models of PrU occurrence based on available structured data alone will be developed and compared with the predictions based on the Braden Scale. In Aim 3 NLP systems will be developed to extract risk factors from the EHR text. In Aim 4 predictions based on the new risk models, combining structured data with NLP data will be compared with predictions based on structured data alone and the Braden Scale. Prediction models will be developed with multivariable logistic regression models.