CCI was originally developed based on data derived from medical records. However, many researchers have proposed that the sole dependence on medical records results in a limitation in cases for which a prompt risk assessment is required by hospitals, insurers and Health Management Organizations (HMOs) . Therefore, CCI tools have been developed for use with claims data coded using ICD-9-CM (International Classification of Diseases, 9th revision, Clinical Modification) [19, 20] and ICD-10 (Switzerland, Australia, Canada versions) [12, 21, 22].
However, the consistency between medical records and claims data is not excellent. For 485 patients who underwent prostatectomy, the comorbidities based on the medical records and claims data were compared. The kappa value was greater than 0.61 for five diseases (uncomplicated diabetes mellitus, primary solid tumor, moderate hepatic disease, connective tissue diseases, leukemia), but the remaining diseases had poor agreement . Also Newschaffer et al. compared CCI scores between the medical records and claims data in 404 patients with breast cancer. This comparison revealed that the kappa value was 0.36, corresponding to 'fair agreement' .
In a Korean study that targeted patients who underwent surgery for the treatment of gastric cancer, the kappa value of CCI comorbidities between the data resources was fair for five diseases; however, excluding these comorbidities, the kappa value was less than 0.2. Especially the prevalence of peptic ulcer disease was 41.6% according to the claims data and 3.5% according to the medical records data . Another study, which focused on patients who underwent hip joint arthroplasty, reported that the kappa value of comorbidities for two data resources was 0.8 for metastatic solid tumors and 0.51 for uncomplicated diabetes mellitus. In other comorbidities, the kappa value was smaller than 0.29 . In this study, there were discrepancies between the two data sources, supporting the results from previous studies in Korea and other countries.
One possible explanation for this disagreement is that there is an underestimation of CCI comorbidities based on medical records data. Medical records based on CCI scores were retrospectively obtained from the one medical institution, whereas claims data based on CCI scores were gathered from administrative data, which includes all medical institutions' claims data on the selected patient. In addition, poor agreement between medical record data and claims data may result from the differing motivations for data collection between medical records and claims data.
Inconsistencies have also been observed in previous studies investigating whether CCI scores obtained from claims data could be used to predict the length of hospital stay. According to a study of 20,138 patients who underwent surgery for radical urological cancer, the length of stay was prolonged by approximately 1-2 days in the group in whom CCI scores were at least 1 compared to the groups with 0 points . In another study, 1,216 patients who visited an outpatient clinic with a chief complaint of acute chest pain, compared with the group in which CCI scores were calculated to be 0-1 points from the medical records, the length of stay was delayed by 14.4-fold (95%CI 3.9-25.9) in the group in which CCI scores were calculated to be 2-3 points and by 25.3-fold (95%CI 2.4-25.5) in the group in which CCI scores were calculated to be < 4 points . However, in 1,945 patients who underwent carotid endarterectomy, CCI scores based on claims data did not correlate with the length of stay, though CCI scores based on medical records were associated with an increased length of stay . In the current study, according to a length of stay prediction model, CCI scores calculated using medical records and claims data were not shown to be prognostic factors. This may be due to the fact that CCI was originally developed to predict mortality, and it is simply not an appropriate tool for predicting length of stay . Also, the postoperative length of stay may be dependent on the severity of the procedure rather than comorbidities .
Moreover, there were disagreements in the predictability of CCI with regard to reimbursement cost. In a medical record-based study of dialysis patients, the mean medical expense per year was $54,000 in cases in which CCI scores were calculated to be less than 4 points, $108,000 in those where CCI scores were calculated to be 4-5 points, $247,000 in those where CCI scores were calculated to be 6-7 points and $407,000 in those where CCI scores were calculated to be 8 or more points. These differences were found to be statistically significant (p < 0.0001) . However, CCI scores did not correlate with medical expense in head-and-neck cancer patients . In our study, according to a reimbursement cost prediction model that was established based on the medical records and claims data as related by CCI, CCI was not selected as a prognostic factor for either model. This could be attributed to the fact that the prediction model was developed for death hazards .
This study has several limitations. First, medical records based on the CCI scores of selected patients were retrospectively obtained from the records of the National Cancer Center. However, claims data based on the CCI scores of selected patients were gathered from HIRA, which includes all medical institutions' claims data on the selected patients. But this could be regarded as strength of claims data in terms of accessibility and efficiency. Second, we considered the reimbursement cost of claims data only with respect to the availability of data. As a result, other medical expenses, such as non-covered services, were excluded. Also, the reimbursement cost was calculated as the sum of the medical costs from the date of surgery to one month after the date of discharge. Therefore, it is possible that outpatient visits except follow-up cancer treatment within a one-month period after discharge may have been included. Also though we considered some prognostic factor such as pathologic staging, there are other prognostic factors such as smoking or adjuvant chemotherapy. More information about such factor could affect the health outcomes and the predictability of CCI.
In this study, there was poor agreement between medical records and claims data. In addition, CCIs based on both data sets were not suitable for predicting length of stay or medical expenses. Given how easy it is to calculate CCIs based on claims data, especially in a social health insurance system, there should be further studies to improve methods for calculating CCI scores to predict health outcome.