Acute myocardial infarction (AMI) outcomes are studied frequently in health service research with hospital discharge administrative data [1–8]. Risk adjustment is an important tool used in health service research to account for differences in AMI patient's characteristics. To develop such a tool for AMI patients, Tu et al. [3] initially selected 43 comorbidities that were used as potential risk factors for AMI in United States hospital report cards. The nine comorbidities were chosen from the 43 comorbidities based on their clinical plausibility and statistical significance. Along with adjustments for sex and age, these 9 comorbidities were found to substantially predict AMI mortality in Ontario provincial database, and were validated in 2 external databases from Manitoba, Canada and California, the United States. Subsequently, many studies have assessed its validity and employed the AMI prediction method [4, 5, 9–21]. Grunau et al. compared Tu et al's disease specific prediction rule to many other methods [20]. These authors found that Tu's method performed better than other general methods that summarized comorbidities as a score, or included comorbidities that are less relevant for AMI outcomes. Hence, Canada has accepted Tu's risk adjustment method to generate public reports of AMI outcomes [21].

With the introduction of tenth revision of the International Classification of Disease (ICD-10) coding system by the World Health Organization (WHO) [22] in 1992, this risk adjustment tool may no longer be valid. ICD-10 is significantly different from ICD-9 coding system, differing in the richness of clinical information and use of alphanumeric codes relative to numeric codes in ICD-9. Hence, previously validated [23] ICD-9 coding algorithms used for the nine comorbidities cannot be applied to the new ICD-10 data. Without valid ICD-10 coding algorithms, administrative data can not be used to measure and control for patients' comorbidities when studying their clinical outcomes. Further, future policy and quality decisions using ICD-10 data require the development and validation of ICD-10 coding algorithms for these comorbidities.

The objective of this study was to develop ICD-10 coding algorithms for the nine AMI comorbidities. The validity of these newly developed ICD-10 algorithms was determined by reviewing medical charts in recording AMI comorbidities. We then applied the ICD-10 and past ICD-9 coding algorithms to administrative data from a large Canadian province. Comparisons between the two coding algorithms were made by comparing their performance in predicting in-hospital, 30-day, and 1-year AMI mortality.

We developed the ICD-10 coding algorithms to define nine comorbidities used to adjust risk factors in AMI outcome studies. The newly developed ICD-10 coding algorithms were comparable with previously published and widely used ICD-9 coding algorithms in detection of cases and predicting mortality. We also know that the ICD-9 and 10 data had similar validity in recording the nine comorbidities when compared to chart review data.

The similar sensitivity, specificity, PPV and NPV found for the ICD-9 and ICD-10 coding algorithms when referenced to the chart review data suggests that both are similarly accurate in recording AMI comorbidities. However, both ICD-9 and ICD-10 datasets matched poorly to the chart review data for pulmonary edema, cardiac dysrhythmias, and cerebrovascular disease. The low sensitivities for pulmonary edema and cardiac dysrhythmias were likely due to the ambiguity of their diagnoses and/or being symptoms of underlying conditions [1, 28–30]. In contrast to comorbidities that are well documented, such as cancer, physicians may document non-specific conditions or symptoms poorly in charts. Therefore, it is difficult for coders to make definitive decisions on the presence of these types of conditions based on the chart information, which can result in coding errors. The low positive predictive value for cerebrovascular disease is partly due to the low prevalence. Further, Benesch et al. [31] found a third of asymptomatic patients coded in ICD-9-CM as having cerebrovascular diseases, but found no evidence of occlusion or stenosis of the precerebral arteries [31]. Kokotailo and Hill [32] found that there was a range of coding errors for cerebrovascular disease, where transient ischemic attack (TIA) was coded for acute ischemic stroke. The difficulty with coding TIA was that the symptoms were non-specific, arising from many other conditions besides a temporary disruption in blood flow to a part of the brain [33]. Nevertheless, the similarity in misclassification by ICD-9 and ICD-10 coding algorithms for these three comorbidities suggests that risk-adjustment with the AMI nine comorbidities will not be affected by the change in coding system from ICD-9 to ICD-10.

When we applied the ICD-10 coding algorithms to the hospital discharge data, we found similarities across frequencies, mortality rates, and overall model performance to ICD-9 coding algorithms. This result suggests that the ICD-10 coding algorithms for these nine comorbidities performed similarly to past ICD-9 coding algorithms. This inference was further supported by the similar results found in our study and the study conducted by Tu et al. [3] Both studies found that the two most prevalent comorbidities were congestive heart failure and cardiac dysrhythmias, and the least prevalent comorbidity was pulmonary edema [3]. This similarity between the two studies was also shown by the overall model performances. Tu et al's 0.78 and 0.79 C-statistics for their 30-day and 1-year mortality logistic regressions, respectively matched closely to the 0.81 C-statistic we found in predicting for both mortality rates for our ICD-10 data [3]. Comparing our odds ratios to Tu et al.'s findings, we both found that patients with shock and acute renal failure tended to have the highest likelihoods of dying [3].

This study had some limitations. The first limitation was low frequencies for comorbidities in the chart review data, making it difficult to draw inferences to a broader population. Secondly, having chart review data from only teaching hospitals might generate a non-random sample, as it might have fewer healthy patients than a sample that included patients from non-teaching hospitals. Evidence of this result can be seen by the higher frequencies in the comorbidities in the Alberta chart review data relative to British Columbia discharge abstract data. Thirdly, gathering chart review data from the first year ICD-10 was implemented might result in a sample with high coding error. Fourthly, possible provincial and temporal changes in the occurrence of AMI, mortality, and comorbidities were not controlled. In the 10 year period, the criteria for AMI diagnosis were revised and efficient treatment to chronic disease was produced. For example, biomarkers of cardiac troponin and creatinine kinase-MB mass were included in diagnosis of myocardial infarction to increase sensitivity and specificity [34]. The advances in disease treatment and management can partly explain the slight variation between ICD-9 and ICD-10 models that can not be accounted for in our analysis. Fifthly, we excluded patients who were discharged on the day of admission or the following day. The limitation of this exclusion is that some true AMI cases were excluded. We re-analyzed our data after including deaths within 1 day. Including the 3,049 deaths within one day produced a decrease in the C-statistics for all our models. This suggests that the C-statistics reported in our study are slightly overestimated. Nevertheless, there remains a small difference between the predictive ability of ICD-9 and ICD-10 models. Lastly, the proportion of AMI patients that could not be linked with vital death registry or had moved out of the BC province without notice to the population registry during our study period were lost to follow-up. According to Li et al. (35), the accuracy of linkage between the registry file and inpatient discharge dataset was about 98%. We also suspect that only a few AMI patients moved out of the BC province given that our longest period of observing mortality was one year. Hence, our results were less likely affected by this small proportion of AMI patients lost to follow-up.

The ICD-10 coding algorithms for AMI comorbidities produced in this study generated comparable results to past ICD-9 coding algorithms. This result was evident when ICD-9 and ICD-10 comorbidities were similar in frequencies and model performance. A step we took to ensure that not only did ICD-10 coding definitions matched well to past ICD-9 coding definitions, but also to a criterion standard, chart review data. The results from this step demonstrated comparable sensitivity, specificity, PPV and NPV between ICD-9 and ICD-10 coding algorithms. Evidence that the ICD-10 coding algorithms performed similar to past coding algorithms will allow future policy and quality judgements to be made on AMI outcomes with accurate ICD-10 comorbidities.