RESUMO
OBJECTIVES: To examine the use of electronic medical record (EMR) data to ascertain falls and develop a fall risk prediction model in an older population. DESIGN: Retrospective longitudinal study using 10 years of EMR data (2004-2014). A series of 3-year cohorts included members continuously enrolled for a minimum of 3 years, requiring 2 years pre-fall (no previous record of a fall) and a 1-year fall risk period. SETTING: Kaiser Permanente Hawaii, an ambulatory setting. PARTICIPANTS: A total of 57 678 adults, age 60 years and older. MEASUREMENTS: Initial EMR searches were guided by current literature and geriatricians to understand coding sources of falls as our outcome. Falls were captured by two coding sources: International Classification of Diseases, Ninth Revision (ICD-9) codes (E880-889) and/or a fall listed as a "primary reason for visit." A comprehensive list of EMR predictors of falls were included into prediction models enabling statistical subset selection from many variables and modeling by logistic regression. RESULTS: Although 72% of falls in the training data set were coded as "primary reason for visit," 22% of falls were coded as ICD-9 and 6% coded as both. About 80% were reported in face-to-face encounters (eg, emergency department). A total of 2164 individuals had a fall in the risk period. Using the 13 key predictors (age, comorbidities, female sex, other mental disorder, walking issues, Parkinson's disease, urinary incontinence, depression, polypharmacy, psychotropic and anticonvulsant medications, osteoarthritis, osteoporosis) identified through LASSO regression, the final model had a sensitivity of 67%, specificity of 69%, positive predictive value of 8%, negative predictive value of 98%, and area under the curve of .74. CONCLUSION: This study demonstrated how the EMR can be used to ascertain falls and develop a fall risk prediction model with moderate sensitivity/specificity. Concurrent work with clinical providers to enhance fall documentation will improve the ability of the EMR to capture falls and consequently may improve the model to predict fall risk.