Evaluation of falls detected by natural language processing algorithm and not coded external cause of morbidity

Area of Science:

• Gerontology
• Public Health
• Medical Informatics

Background:

• Falls are a major cause of death and disability in older adults.
• Current methods using diagnosis codes underestimate the true incidence of falls.
• Accurate identification of fall-related emergency department (ED) visits is crucial for public health surveillance and patient care.

Purpose of the Study:

• To apply a natural language processing (NLP) algorithm to identify fall-related ED visits among older adults.
• To compare the characteristics of patients identified by NLP with those identified by traditional diagnosis codes.
• To assess the impact of different identification strategies on understanding fall prevalence and patient outcomes.

Main Methods:

• A cross-sectional study analyzed ED encounter data for older adults from December 2016 to 2020.
• An NLP algorithm processed provider notes to detect fall-related keywords, excluding negated or spurious mentions.
• International Classification of Diseases (ICD) codes were used as a traditional comparison method.

Main Results:

• The NLP algorithm identified 14,604 fall-related ED encounters out of 50,153.
• Nearly half (49%) of NLP-identified cases were missed by ICD code-based methods.
• Patients identified solely by NLP had higher comorbidity scores and increased 30-day mortality, often linked to severe conditions like sepsis or kidney disease.

Conclusions:

• NLP algorithms significantly enhance the detection of fall-related ED visits compared to traditional methods.
• This improved identification highlights a vulnerable patient subgroup often missed by standard surveillance.
• Careful consideration of causal links between falls and underlying illnesses is necessary when interpreting NLP findings.