Hekman Daniel J, Maru Apoorva P, Barton Hanna J, Wiegmann Douglas, Shah Manish N, Cochran Amy L, Ötleş Erkin, Patterson Brian W
BerbeeWalsh Department of Emergency Medicine, School of Medicine and Public Health, University of Wisconsin-Madison, Madison, WI 53706, United States.
Department of Industrial and Systems Engineering, College of Engineering, University of Wisconsin-Madison, Madison, WI 53706, United States.
JAMIA Open. 2025 Jun 20;8(3):ooaf047. doi: 10.1093/jamiaopen/ooaf047. eCollection 2025 Jun.
Falls are a leading cause of morbidity and mortality among older adults. Common methods for identifying fall-related ED visits within both claims and electronic health record datasets rely on diagnosis code-based definitions, which underestimate the true prevalence of falls. This study applies a natural language processing (NLP) algorithm to ED provider notes to identify patients presenting due to falls and compares the characteristics of NLP-identified cases to those identified through diagnosis codes to identify the impact of identification strategy.
This cross-sectional study analyzed ED encounter data from older adult patients who visited an ED between December 2016 and 2020. The NLP algorithm identified falls based on provider notes, searching for keywords related to falls and excluding negated and spurious matches. We also applied common ICD code methods to identify falls.
We processed 50 153 ED encounters and the NLP approach identified 14 604 encounters for patients who fell. Of those, 7086 (49%) were not identified using external cause of morbidity ICD codes. Patients identified by just the NLP algorithm exhibited higher Elixhauser comorbidity scores and increased likelihood of 30-day mortality. Patients identified by NLP algorithm but not ICD codes were more likely to have severe underlying conditions such as sepsis or acute kidney disease rather than traumatic injuries.
The NLP algorithm identifies many fall-related visits not identified by traditional methods.
If the causal relationships between falls and comorbid conditions are not considered in NLP algorithms, they can easily identify patients who fell, but the fall was a sequela of underlying medical illness.
跌倒在老年人发病和死亡原因中占主导地位。在索赔和电子健康记录数据集中识别与跌倒相关的急诊就诊的常用方法依赖于基于诊断代码的定义，这低估了跌倒的真实患病率。本研究应用自然语言处理（NLP）算法对急诊医生记录进行分析，以识别因跌倒就诊的患者，并将NLP识别病例的特征与通过诊断代码识别的病例特征进行比较，以确定识别策略的影响。
这项横断面研究分析了2016年12月至2020年期间到急诊就诊的老年患者的急诊就诊数据。NLP算法根据医生记录识别跌倒情况，搜索与跌倒相关的关键词，并排除否定和虚假匹配。我们还应用常见的国际疾病分类（ICD）代码方法来识别跌倒情况。
我们处理了50153次急诊就诊记录，NLP方法识别出14604例跌倒患者的就诊记录。其中，7086例（49%）未通过疾病外部原因ICD代码识别。仅通过NLP算法识别出的患者具有更高的埃利克斯豪泽合并症评分和30天死亡率增加的可能性。通过NLP算法但未通过ICD代码识别出的患者更有可能患有严重的基础疾病，如败血症或急性肾疾病，而非创伤性损伤。
NLP算法识别出许多传统方法未识别出的与跌倒相关的就诊情况。
如果在NLP算法中不考虑跌倒与合并症之间的因果关系，它们可以轻松识别出跌倒的患者，但跌倒可能是基础疾病的后遗症。
