Emergency Department Visit Classification Using the NYU Algorithm | Page 2
Published Online: April 21, 2014
Sabina Ohri Gandhi, PhD; and Lindsay Sabik, PhD
To estimate the association of emergent and intermediate ED visit classification based on the EDA with mortality and hospitalization, we estimated logistic regression models, using NHAMCS complex survey weights. The 2 main outcome variables of interest were dichotomous indicators for whether an individual died (either on arrival to the ED, in the ED, or after being hospitalized) and whether the patient was admitted to the hospital from the ED. All models controlled for individual characteristics, including age, gender, race (white, Black, other race); payment source (private, Medicare, Medicaid/SCHIP, self-pay/charity, other); socioeconomic indicators at the patient zip code level, including low median household income and urban location; and survey year. The covariates in the model were selected because they are known to be associated with hospitalization and health outcomes.15,16
Table 1 presents the breakdown of ED visits as classified by the EDA. We were able to classify 73,076 of 111,970 total ED visits into the emergent, intermediate, and nonemergent categories. This represented a population of about 260 million ED visits. The remaining ED visits (38,894) were classified into injury (32,149), drug (304), alcohol (1718), and mental health (4723). We were not able to classify about 16% of ED visits using the EDA, which is similar to the percentage of diagnosis codes that are uncommon and could not be classified as reported by the developers of the EDA.13 We limited our analysis sample to patients with diagnosis codes that could be classified into emergent, intermediate, and nonemergent. We hypothesized that patients with ED visits classified as emergent would have a higher probability of mortality and hospitalization. Like Ballard et al, we excluded visits categorized as drug, alcohol, or mental health, since they may have been emergent but were less clearly associated with mortality or hospitalization.9 In addition, we excluded visits categorized as injury since they may have been justifiably emergent, but had a low probability of mortality or hospitalization. Within the sample that could be classified into 1 of the 3 levels, 24.1% of visits were categorized as emergent, 1.9% as intermediate, and 74.0% as nonemergent, using the 50% probability threshold.
Table 2 presents summary statistics for the analysis sample. Patients with ED visits in the sample had an average age of 37 years and were more likely to be female, white, have Medicaid or SCHIP as a primary payer, and live in an urban area. About 15% of patients were either admitted or died on arrival to the ED, in the ED, or in the hospital after an ED-associated visit.
Analysis of Emergent ED Visits Using the EDA
Table 3 presents logistic regression results estimating the association between EDA visit classification and the outcome variables, mortality, and hospital admission. We found that emergent ED visit classification was positively and significantly associated with mortality (odds ratio [OR]: 3.79, P <.01) and with ED-related hospitalization (OR: 5.28, P <.01), relative to nonemergent classification. Intermediate classification was also positively associated with mortality (OR: 2.16, P = .12) and with ED-related hospitalization (OR: 4.14, P <.01), relative to nonemergent status. The results on the control variables were as expected. Older patients, males, and patients with public insurance or self-pay/charity as the primary payer (vs private insurance) were associated with significantly higher probabilities of ED-associated death and hospitalization. The findings regarding publicly insured and self-pay ED visits were consistent with other research that has shown these patient populations have poorer health and higher inpatient mortality than the privately insured.17-19
We performed 2 specification checks to test the sensitivity of our results (full results available upon request). First, we estimated conditional logit models with year-specific ED fixed effects in order to control for any unobserved time-invariant ED-specific differences (such as staffing and resources) and to examine within-ED differences in hospitalization and mortality. Results from these models were very similar to our baseline models for both mortality (OR: 3.52, 95% confidence interval [CI]: 2.58-4.80) and hospital admission (OR: 5.22, 95% CI, 4.96-5.50). Second, following Ballard et al, we repeated our analyses using different probability thresholds (75% and 90%) for the ED measures in order to check the sensitivity of our results.9 In these models, visits were categorized as intermediate if the probability they were emergent was between 25% (10%) and 75% (90%). The results for our emergent variable represent the comparison of visits with a predicted probability of being emergent of 75% (90%) or greater with visits with a predicted probability of being emergent of 25% (10%) or less, thus we expect the ORs to increase, but be in the same direction as our baseline results. We found that this is generally the case, with ORs for our mortality outcome of 6.11 (95% CI, 3.73-10.02) and 4.90 (95% CI, 2.16-11.15) for the 75% and 90% cutoffs, respectively. Similarly, ORs for our admission outcome are 8.35 (95% CI, 7.60-9.17) and 11.26 (95% CI, 9.83-12.92) for the 75% and 90% cutoffs, respectively. All specification checks support our main finding that emergent visit status is significantly associated with hospital admission and death.
This analysis evaluated methods for classifying ED visits as emergent, using the EDA in a nationally representative sample of ED visits, and found that measures based on the EDA were significantly associated with mortality and EDassociated hospitalizations. The magnitude of association of emergent classification with hospitalization and mortality were similar to, but larger than, those from Ballard et al, who used data from privately insured individuals in a single integrated delivery system.9 The difference may be due to the fact that our sample included both the Medicaid and uninsured populations, and was not restricted to privately insured and Medicare patients in an integrated delivery system, who are likely to have better access to care. In addition, our sample excluded visits categorized as injury, which may also explain our larger result. Further, our results suggest that the EDA can be used to categorize visits as emergent and nonemergent by researchers when information is available on diagnosis codes but not on triage time (eg, claims data) or when a full chart review is not possible.
PDF is available on the last page.