In this study, we demonstrated high predictive value for detection of ED hypoglycemia visits using the new ICD-9-CM coding algorithm. To our knowledge, this is the first chart validation of the accuracy of ICD-9-CM codes for hypoglycemia, an important complication in diabetes and common ED presentation.
Our algorithm differs in several important ways from prior reports on hypoglycemia. Two separate analyses of the relationship between anti-hypertensive agents and hypoglycemia in patients with diabetes, Herings et al. and Morris et al. used ICD-9-CM codes 251.0, 251.1, and 251.2 to identify possible episodes of hypoglycemia [12
]. Schorr et al. broadened candidate codes to include 250.3 and 962.3 [11
]. Notably, none of these studies included 250.8 in their search strategies, nor accounted these possible missed episodes of hypoglycemia. In our analysis, 250.8 comprised the vast majority (83%) of hypoglycemia cases, and exclusion of this code in these prior studies likely contributed to substantial underestimates in the incidence of hypoglycemia.
Johnson, et al. first reported the utility of 250.8 in their strategy to identify ED visits for hypoglycemia [10
]. In their analysis, this code also identified over half of hypoglycemia episodes and had a PPV of 73% for their case definition of hypoglycemia. Our chart validation differed in several important ways. First, Johnson et al. limited their search to first-listed ICD-9-CM codes and excluded all other diagnosis fields. We evaluated all diagnosis fields and while candidate codes in the first diagnosis field had a higher PPV, limitation to the primary diagnosis field would have excluded 27% of hypoglycemia cases. Additionally, candidate codes in the latter diagnosis fields maintained a high PPV (81%). Secondly, their validation of hypoglycemia cases involved searching of brief communications following the ED visit for symptoms consistent with hypoglycemia, which included confusion, loss of consciousness, and seizure. While these symptoms are consistent with hypoglycemia, they are nonspecific and have many other causes in addition to hypoglycemia [17
]. We performed detailed chart reviews to confirm hypoglycemia visits, which allowed for more precise validation. Finally, we confirmed the importance of 250.8 in identification of the majority of hypoglycemia visits, but additionally refined the search strategy by exclusion of visits with alternate co-diagnoses. Exclusion of these cases reliably eliminated 85% of false-positive cases coded as 250.8, while eliminating only a small number of true-positive cases.
Overall, 250.8 with exclusion of alternate co-diagnoses and 251.2 comprised >95% of hypoglycemia identified by the algorithm. Although there were insufficient data to evaluate 250.0 and 250.1 in this analysis, we recommend inclusion of these codes in any search for hypoglycemia. These codes are likely to have high specificity/PPV, and their utilization may be higher in different institutions and practice environments. As anticipated, codes 270.3, 775.0 and 775.6 were rare in the ED setting but may have utility for other settings, such as maternity and neonatal units.
We hypothesized that we would not be able to determine if diabetes with other coma (250.3) were caused by hypoglycemia or diabetic ketoacidosis (DKA). There were insufficient data to evaluate this hypothesis (only one case), and charts identified with this code should be interpreted with caution. Compared to prior 100% PPV in prior analysis [10
], we were surprised by the relatively low predictive value (54%) of 962.3. Based on this finding and our secondary analysis, we recommend exclusion of 962.3 and relying only on 250.8 to identify most of these associated visits. This proposed strategy will require validation in other institutions and settings.
The ICD-9-CM classification system is imperfect for case identification, as it was created for reimbursement rather than research purposes [18
]. Given this inherent limitation, the proposed algorithm for hypoglycemia compares favorably to those suggested for a variety of diseases. ICD-9-CM coding accuracy for upper gastrointestinal disorders (93–95% PPV) and Clostridium difficile colitis (87% PPV) were similar [19
], but were more unreliable for stroke (61–79% PPV) [21
], soft tissue disorders (64% PPV) [22
], pneumococcal pneumonia (58% PPV) [23
], traumatic brain injury (20–38% PPV) [24
], and venous thromboembolism (31% PPV) [25
]. Based on these comparisons, the proposed ICD-9-CM coding algorithm for hypoglycemia would be expected to perform as well or better than coding strategies for other diagnoses in accurate case identification from administrative data.
The current study has some potential limitations. We did not abstract the 25% of charts that were missing or incomplete. While this lowered the number of total cases, the distribution of ICD-9-CM codes was similar in abstracted and non-abstracted charts, and the likelihood of biased accuracy estimates was small. The coding of hypoglycemia visits was based on three academic EDs and may not generalize to other areas of the hospital, outpatient providers, and other geographic areas. However, there are a finite number of options for coding hypoglycemia, and although the frequency of individual code usage will likely vary based on local practice and familiarity of the coding personnel, we anticipate that accuracy will not vary substantially. We were unable to formally evaluate false-negatives in our data set. Thus, we were only able to calculate PPV with accuracy, and our estimates for sensitivity, specificity, and NPV are significantly limited. To help address this limitation, we performed a sensitivity analysis, which demonstrated a high level of accuracy even when we increased the assumption of missed cases, but inferences based on these values should be guarded. Finally, the accuracy of case confirmation by chart review depended on a retrospective evaluation for hypoglycemia, which may overestimate or underestimate the true number of cases. Standardized definitions and training of reviewer limited the potential for such bias, and high inter-rater agreement demonstrated internal reliability of the chart review.