Using electronic medical record data to shorten diagnostic odysseys for rare genetic disorders in children and adults in two New York City health care settings

罕见的遗传病影响3.5%-6%的人口，并与诊断的奥德赛有关，这种疾病可以 可持续数十年。作为缩短婴幼儿诊断过程的第一步，我们 开发了基于规则和自然语言处理(NLP)的算法来识别婴儿和 0-3岁的儿童，他们是典型的疾病。我们的算法在识别非典型疾病患者方面是准确的 这些年龄来自电子健康记录(EHR)。这样确定的队列对患者来说是非常丰富的 接受过基因检测的人。为从未经历过这种非典型疾病的患者进行手动电子病历审查 一种罕见的遗传病的评估显示，52%的人可以适当地转介进行这样的评估。 在UG3阶段，我们将创建一个新的门诊诊所，即西奈山遗传外展(GO)，配备工作人员 与曾接受过儿科和内科培训的医学遗传学家一起评估我们的 EHR表型算法。在一项试点研究中，我们将部署基于规则和NLP的算法来识别200 年龄在0-12岁之间的儿童有50%的风险有未确诊的罕见遗传特征。我们将调查 五家诊所的儿科医生提供了关于诊断奥德赛和基因测试的基线知识 关于主题的教育，然后研究我们的算法部署的影响。对于被转介的患者 西奈山，我们将决定临床遗传评估和诊断测试的结果， 包括对后续医疗保健的影响。为了改进现有的算法，我们开发了一种 自动摘要引擎，可识别164例罕见遗传病患者，识别率为83% 精确度。我们将把它扩展到更多的特征，并使用他们的EHR数据来改进我们的儿科EHR表型 算法。目标是在不降低精度低于50%的情况下提高灵敏度，目前为~25%。 在UH3阶段，我们将在非学术领域部署我们优化的罕见病检测算法 医疗保健环境，西奈山南拿骚医院，一家没有 现场医疗遗传服务。我们的模型将利用大流行加速的远程医疗专业知识来 促进未得到充分服务的人群获得遗传学服务。我们的目标将是实现类似的敏感性 与我们的儿科算法和相对成功的转诊机制相结合的精确度。另外，我们 将扩展我们的基于临床规则的和NLP算法来检测青少年和成年患者可能患有 罕见的遗传疾病，并评估我们的方法对诊断奥德赛的影响。我们将改变我们的 基于儿科规则的算法，首先适用于12-21岁的患者，然后适用于年轻人。我们将利用 我们针对罕见遗传病的自动抽象引擎，用于迭代改进。对于成年人，我们将上课 按器官系统分类的特征，以提高队列规模/统计能力。最后，我们将集合在一起学习 关于诊断奥德赛本身的信息，包括我们的算法在缩短它们方面的影响。

Rare genetic diseases affect 3.5-6% of the population and are associated with diagnostic odysseys that can last up to decades. As first steps towards shortening diagnostic odysseys for infants and toddlers, we developed rules-based and natural language processing- (NLP-) based algorithms to identify infants and children aged 0–3 years who were typically ill. Our algorithms were accurate for identify atypical ill patients at these ages from electronic health records (EHRs). Cohorts so identified were strongly enriched for patients who had undergone genetic testing. Manual EHR review for such atypically ill patient who had never been evaluated for a rare genetic disease revealed that 52% could appropriately be referred for such an evaluation. During the UG3 phase, we will create a novel outpatient clinic, Mount Sinai Genetics Outreach (GO), staffed with medical geneticists with prior pediatric and internal medicine training, to evaluate patients identified by our EHR phenotyping algorithms. In a pilot study, we will deploy rules- and NLP-based algorithms to identify 200 children aged 0-12 years with >50% risk of having an undiagnosed rare genetic trait. We will survey pediatricians at five practices for baseline knowledge about diagnostic odysseys and genetic testing, provide education about the topic and then study the impact of our algorithm deployment. For patients referred to Mount Sinai GO, we will determine the outcomes of clinical genetic evaluations and diagnostic testing, including impact on subsequent health care. In order to improve our existing algorithms, we developed an automated abstraction engine that identifies patients diagnosed with 164 rare genetic disorders with 83% accuracy. We will expand this to more traits and use their EHR data to improve our pediatric EHR phenotyping algorithms. The goal is to increase sensitivity, currently at ~25%, without dropping precision below 50%. During the UH3 phase, we will deploy our optimized rare disease-detecting algorithms in a non-academic health care setting, Mount Sinai South Nassau Hospital, a non-academic community hospital setting without onsite medical genetic services. Our model will leverage pandemic-accelerated expertise in telehealth to facilitate access of underserved populations to genetics services. Our goal will be to achieve similar sensitivity and precision with our pediatric algorithms as well as a comparably successful referral mechanism. Also, we will extend our clinical rule-based and NLP algorithms to detect adolescent and adult patients likely to have rare genetic disorders and assess the impact of our approach on diagnostic odysseys. We will alter our pediatric rules-based algorithm, first to patients aged 12-21 years and then to younger adults. We will leverage our automated abstraction engine for rare genetic disease for iterative improvements. For adults, we will class traits by organ system in order to improve cohort size/statistical power. Finally, we will assemble and study information about diagnostic odysseys per se, including the impact of our algorithms in shortening them.