Staging discrepancies between Hospital-Based Cancer Registry and Diagnosis Procedure Combination data

Monitoring current trends in cancer care is essential for improving the quality of care (1). Although such a monitoring system has not yet been established in Japan, several existing databases have the potential to be used for this purpose. Two of the best candidates for use in such monitoring are the national Hospital-Based Cancer Registry (HBCR) and the Diagnosis Procedure Combination (DPC) databases (2). In Japan, HBCRs are maintained in designated cancer care hospitals (DCCH) as a requirement for designation by the government (3). DCCHs serve as hubs for managing cancer care in their respective regions or communities. Hospitals are expected to use their HBCR in evaluating the care they provide and to submit annual data to the national database managed by the National Cancer Center. In 2013, 397 DCCHs were designated to provide care for ~70% of all new cancer patients (1). At each DCCH, tumor registrars enter patients’ information into the HBCR following nationally standardized coding rules. Tumor registrars are required to complete a training course provided by the National Cancer Center. In this course, trainees learn the staging systems along with the detailed coding rules of the HBCR. After completing this course, they must pass a certification examination. The DPC is a case-mix classification system for determining the daily amount of reimbursement from the national insurance to hospitals’ inpatient services (4). Hospitals that opt to receive reimbursement based on the DPC system must submit health care service data equivalent to fee-for-service claims to the national DPC database. These data are used by governmental committees to examine the effect of the DPC system on the cost and care (2). Along with health care services, the DPC data also contain patients’ demographic information and medical services provided as well as administrative data created when patients are discharged from the hospital. The administrative data contain cancer staging information for that hospitalization. Because both HBCR and DPC data include cancer staging information, a valuable opportunity exists to assess the accuracy of the information. Registrars for the two databases enter information on cancer patients independently. In the USA, previous studies have shown some discrepancies in metastasis coding data between Medicare claims and the Surveillance, Epidemiology and End Results (SEER) Program (5,6). To our knowledge, no study has evaluated the concordance of HBCR data with DPC data in Japan. Our study aims to assess the accuracy of tumor–node–metastasis (TNM) classification concordance between the HBCR data and the DPC data among voluntarily participating hospitals. We compared HBCR and DPC data for the same patients by linking the two data sets using the patient identifiers within the participating hospitals. The same data that are submitted to the national databases were provided to the research team by four voluntarily participating DCCHs (Hospitals A–D, two university hospitals and two general hospitals) in Japan. The HBCR collects information on all malignant tumors, including intraepithelial tumors. The definition of malignancy is based on the behavioral codes of 2 or 3 in the International Classification of Diseases for Oncology, third edition (ICD-O-3). Tumor characteristics, including the topology (site) and morphology (histology) codes of ICD-O-3 for all cancers and stages (i.e. clinical [c-] stage and pathological [p-] stage), are required by the coding rules for the five major cancers in Japan (breast, colorectal, liver, lung and stomach). When the staging information is unknown because a patient finished initial treatment in another hospital, the HBCR codes it as unknown, as required by the registration rules. For cases in 2011, stages were coded according to the sixth edition of the Union for International Cancer Control (UICC). The DPC collects information on each hospitalization. Stages prior to inpatient treatment were recorded for those cases in which the cancer was primary and the first-course treatment was performed during hospitalization. In 2011, the DPC also used the TNM classification codes of the sixth edition of the TNM Classification of Malignant Tumours. However, the overall stages (i.e. Stages I–IV) were based on the Japanese classification, not the TNM classification. We extracted HBCR and DPC data on all patients diagnosed in 2011 with one of the five major cancers in Japan (stomach, colorectal, lung, breast and liver) from the databases of the participating hospitals. We compared the TNM classifications in the HBCR data with those in the DPC and then calculated concordance rate for patients with the five major cancers. The unit of record in the HBCR was ‘tumor’, whereas the unit of record in the DPC was ‘hospitalization’. For each patient, the HBCR data generally contained two types of staging information, the clinical and pathological stages, while the DPC data contained staging information entered ‘each time’ patients were hospitalized. Due to the differences in the way staging information was entered between the two systems, we could not pair the TNM classification in HBCR and that in DPC for determining the one-to-one concordance. For research purposes, we chose a conservative (i.e. less strict) criterion, considering a case as concordant if the TNM classification in the DPC data of at least one hospitalization was the same as either the clinical or pathological stage in the HBCR. We excluded cases in which the TNM classifications were unknown in both the HBCR (c- and p- stage) and DPC data for all hospitalizations for that patient. We also compared concordance rate across the four hospitals. All statistical analyses were performed using Stata 13 (Stata Corporation, College Station, TX, USA). This study was approved by the Ethics Committee at the National Cancer Center, Japan. Overall, HBCRs of the four hospitals contained 2015 cases with any of the five major cancers. Of these, 282 cases were excluded because the corresponding DPC data were lacking and patients were treated only in outpatient settings (e.g. for resection of colon polyps later found to contain cancers). After limiting the cases to patients who underwent first-time treatment, the study sample contained 1462 cases. The sample included 366 patients with stomach cancer, 330 with colorectal cancer, 437 with lung cancer, 117 with liver cancer and 212 with breast cancer. More than half of the sample were men (841, 57%), and the mean age at diagnosis was 68 years (SD = 12). Data for 319 (22%) patients were obtained from Hospital A, 396 (27%) from B, 151 (10%) from C and 596 (41%) from D. For 77% of the patients (95% confidence interval [CI]: 75–79%), the T, N, and M classifications were concordant between the two data sets. For respective TNM classifications, concordance was achieved in 87% of patients (95% CI: 85–89%) for the T stage, 90% (95% CI: 88–91%) for the N stage, and 93% (95% CI: 92–94%) for the M stage (Table 1). Among the TNM classifications for the patients with stomach, colorectal or lung cancer, the M stages tended to have the highest concordance rate (93–95%). For liver cancer, concordance rates were the highest for N stages (96%, 95% CI: 92-99%). For all four of these cancers, T stage generally showed a lower concordance rate (83–89%). For patients with breast cancer, the concordance rate for TNM classification was similar: 89% (95% CI: 85–93%) for the T classification, 87% (95% CI: 83–92%) for the N classification and 85% (95% CI: 80–90%) for the M classification. The concordance rate by hospitals and cancer types Note: The denominators are different from the total number of patients if information was not available in either the Hospital-Based Cancer Registry or Diagnosis Procedure Combination system. The concordance rate by hospitals and cancer types Note: The denominators are different from the total number of patients if information was not available in either the Hospital-Based Cancer Registry or Diagnosis Procedure Combination system. For the T stage, the concordance rates for lung cancer (89%, 95% CI: 87–92%) and breast cancer (89%, 95% CI: 85–93%) were relatively higher, while the concordance rate for colorectal cancer (83%, 95% CI: 79–86%) was relatively lower. For the N stage, the concordance rates for liver cancer (96%, 95% CI: 92–99%) and lung cancer (93%, 95% CI: 90–95%) were relatively higher than for other cancers (87–89%). For the M stage, concordance rate was more than 93%, except in cases of breast cancer (85%, 95% CI: 80–90%). The TNM classification concordance rate varied across hospitals (Table 1). The overall concordance rate for the M stage for all five cancers was slightly higher than those for the T or N stages in Hospitals B, C and D. The concordance rate of N and M stages at Hospital A was lower than those at the other hospitals in total. The trend varied slightly across the cancer sites. If TNM information was unavailable in one of the data sets, we considered the cases discordant. Figure 1 describes the patterns of missing data of the discordant cases. Less than 10% of the discordant cases were due to data missing from the HBCR for T (6%), N (3%) and M (2%) stages. More than 15% of the discordant cases were due to missing data in the DPC system for T (16%), N (30%) and M (61%) stages. Furthermore, a check of the consistency between TNM and overall stages showed that overall stages in the HBCR data were perfectly (100%) consistent with T, N and M stage combinations. The same analyses could not be done with DPC data except for breast cancer because DPC used the Japanese domestic staging system for overall stages and the UICC system for TNM stages. For breast cancer, which was staged identically by the UICC and Japanese systems for 2011 cases, the overall stages in DPC were consistent with TNM in 95.1% of the cases. Data availability among the discordant cases. Note: Numbers in parenthesis indicate the number of discordant cases. This study investigated the concordance rate for the TNM classification between HBCR and DPC data and found that classifications were concordant for only 77% (95% CI: 75–79%) of patients. Although this study is preliminary, our results indicate the need for better quality control of TNM staging information. We cannot be certain as to whether HBCR or DPC data are more accurate until the data are compared with medical records. However, we have several reasons to anticipate that the HBCR data are more reliable. For one, the HBCR has several mechanisms for quality control, such as tumor registrars’ training and site visits by National Cancer Center staff, while the DPC does not have similar mechanisms. Another reason for having greater confidence in the reliability of HBCR data is the greater record completeness we found in the HBCR database in this study. For example, more than half of the discordant M classification data were due to missing DPC data. Finally, DPC data may have different staging information between hospitalizations, which complicates data analyses. A possible explanation for the discrepancy lies with variances in data collection and quality control of data at the institution and individual levels. Further analysis stratified by both patient and systematic factors is needed to clarify the underlying reasons for the discordance. The low concordance rate for TNM staging as a whole (77%) may be an overestimate because any match between HBCR (c- or p- stage) and DPC (at any hospital stay) was labeled as concordant. Since the DPC records staging data for each hospitalization and the staging information entered may change from one hospital stay to another, we cannot analyze the pattern of discordance. Furthermore, we found variations in the concordance rate across the four hospitals, potentially due to different quality controls in data recording in each hospital. These discrepancies between systems make the comparison of care/practice patterns across hospitals challenging. This study has several limitations. First, we used data from only four DCCHs in Japan. These hospitals were motivated to participate in this study and were not necessarily representative of all DCCHs. However, these hospitals did not differ much in the distribution of the cancer stages from that of the national report (3). Second, we did not evaluate the concordance rate of the overall stage because the HBCR and DPC data are based on different types of classifications (UICC versus Japanese classification for each cancer). Third, we did not assess the accuracy of the TNM classification through comparison with patient medical records. Although we have some evidence suggesting that HBCR is more accurate, we cannot draw definitive conclusions. Despite these limitations, because this study is the first investigation of discrepancies in the TNM classification between HBCR and DPC data, it provides insight into potential issues in using these databases to evaluate clinical practices and survival. In conclusion, we found that the concordance rate for TNM classification between HBCR and DPC data was suboptimal. The TNM classification discrepancies may lead to inaccurate results in evaluations of cancer practices and survival rate. More attention should be paid to data accuracy in large databases. This study was supported by the Cancer Research and Development Fund of National Cancer Center, Japan. Ayako Okuyama works for the Hospital-based Cancer Registry section at the National Cancer Center. Takahiro Higashi does extensive research using hospital-based cancer registry.