Study design
This was a multicenter, retrospective cohort study using a structured medical record review for validation of a prospectively derived ICD-9 coding algorithm. We obtained ethics approval with a waiver of informed consent from the Partners Human Research Committee and the Beth Israel Deaconess Medical Center Committee on Clinical Investigation.
Study setting and population
This study was conducted at three urban, academic EDs, which are active participants in the Emergency Medicine Network [15]. The EDs have a combined annual visit volume of 175,000 and are staffed by emergency medicine, internal medicine, and surgery residents, and patient care is supervised by attending emergency physicians 24 hours/day.
Study Protocol
Using the electronic medical records system at each site, we searched the following ICD-9-CM codes to identify possible visits for hypoglycemia: 250.3 (diabetes with other coma), 250.8 (diabetes with other specified manifestations) 251.0 (hypoglycemic coma), 251.1 (other specified hypoglycemia), 251.2 (hypoglycemia, unspecified), 270.3 (leucine-induced hypoglycemia), 775.0 (hypoglycemia in an infant born to a diabetic mother), 775.6 (neonatal hypoglycemia), and 962.3 (poisoning by insulins and antidiabetic agents).
Given the diversity of potential ICD-9-CM codes, we searched this broad range of codes and in all diagnosis fields (up to ten listed) in an attempt to capture all possible ED hypoglycemia visits. For admitted patients, we examined only ED-based codes, to avoid inclusion of incident hypoglycemia that occurred during inpatient hospitalization. In cases where multiple candidate codes were present, we recorded only the first-listed code. The exception to this was for the more ambiguous codes 250.3 and 250.8, for which we preferentially recorded any of the other candidate codes if present. We based this strategy on detailed examination of the ICD-9-CM coding manual [9], review of the experience from previously reported approaches [10–14], and discussion with coding experts.
The code 250.8 may be used for other specific diabetes-associated complications in addition to hypoglycemia, including: 259.8 (secondary diabetic glycogenosis), 272.7 (diabetic lipidosis), 707.xx (ulcers of the lower extremity), 709.3 (Oppenheim-Urbach syndrome), and 730.0–730.2, 731.8 (osteomyelitis). Based on discussion with coding experts, we determined that 681.xx (cellulitis of fingers/toes), 682.xx (other cellulitis), and 686.9x (local skin infection) may also be utilized as a co-diagnoses for 250.8, although not specifically mentioned in the manual. We prospectively proposed the coding algorithm displayed in Figure 1 and validated its accuracy through chart review.
We identified all ED visits with candidate ICD-9-CM codes between July 1, 2005 and June 30, 2006 at each site, and obtained written ED charts. For patients with multiple ED visits during the data collection period, we requested only the first visit to avoid overrepresentation by certain patients. Trained research staff abstracted all charts using a standardized form, and the research group met weekly to discuss data collection and resolve abstraction issues. Additionally, two reviewers independently abstracted 10% of charts to evaluate inter-rater agreement in data collection. To enhance the reliability of our chart review, we abstracted only charts with complete ED triage assessment, nursing notes, and emergency physician notes and considered all other charts incomplete.
Key Outcome Measures
We considered this chart validation as the gold standard for confirmation of true hypoglycemia visits. In reviewing the ED chart, we confirmed cases of hypoglycemia based on the following criteria: 1) Any documented pre-hospital or ED glucose value (serum or capillary) 3.9 mmol/l, or 2) Charted physician discharge diagnosis of hypoglycemia. We based the glucose threshold on the consensus recommendation defined by the American Diabetes Association Workgroup on Hypoglycemia [16]. We included physician diagnosis of hypoglycemia to capture cases in which hypoglycemia resolved prior to first glucose level, i.e. patients receiving glucose for symptoms consistent with hypoglycemia prior to blood glucose determination. Additionally, we collected patient disposition (discharge or hospital admission) to evaluate for differences in coding accuracy based on this factor.
Data analysis
We performed statistical analyses using Stata 9.0 (College Station, TX). We summarized data using basic descriptive statistics. We determined accuracy of specific codes and coding algorithm by calculating PPV with 95% confidence intervals (CI). Although we included a broad range of possible ICD-9-CM codes, the numbers of missed hypoglycemia visits, not captured by the candidate codes, were unknown. Under the ideal assumption that no cases were missed, we calculated estimated point estimates for sensitivity, specificity, and negative predictive values (NPV). Additionally, we performed a sensitivity analysis by increasing the presumed number of missed cases by 10%.
We determined inter-rater agreement for chart abstraction by calculating the kappa statistic for the subgroup of charts abstracted by two investigators. Additionally, we compared accuracy of the algorithm stratified by candidate ICD-9-CM code position, study site and ED disposition using chi-squared test and with two-tailed p < 0.05 considered statistically significant. Finally, we evaluated the accuracy of all ICD-9-CM codes to identify potential improvements and calculated the accuracy of the revised algorithm including the proposed modifications.