School of Population Health, The University of Western Australia, 35 Stirling Highway, Crawley, WA 6009, Australia.
BMC Med Res Methodol. 2013 Oct 1;13:121. doi: 10.1186/1471-2288-13-121.
Administrative data are a valuable source of estimates of diabetes prevalence for groups such as coronary heart disease (CHD) patients. The primary aim of this study was to measure concordance between medical records and linked administrative health data for recording diabetes in CHD patients, and to assess temporal differences in concordance. Secondary aims were to determine the optimal lookback period for identifying diabetes in this patient group, whether concordance differed for Indigenous people, and to identify predictors of false positives and negatives in administrative data.
A population representative sample of 3943 CHD patients hospitalized in Western Australia in 1998 and 2002-04 were selected, and designated according to the International Classification of Diseases (ICD) version in use at the time (ICD-9 and ICD-10 respectively). Crude prevalence and concordance were compared for the two samples. Concordance measures were estimated from administrative data comparing diabetes status recorded on the selected CHD admission ('index admission') and on any hospitalization in the previous 1, 2, 5, 10 or 15 years, against hospital medical records. Potential modifiers of agreement were determined using chi-square tests and multivariable logistic regression models.
Identification of diabetes on the index CHD admission was underestimated more in the ICD-10 than ICD-9 sample (sensitivity 81.5% versus 91.1%, underestimation 15.1% versus 4.4% respectively). Sensitivity increased to 89.6% in the ICD-10 period using at least 10 years of hospitalization history. Sensitivity was higher and specificity lower in Indigenous patients, and followed a similar pattern of improving concordance with increasing lookback period. Characteristics associated with false negatives for diabetes on the index CHD hospital admission were elective admission, in-hospital death, principal diagnosis, and in the ICD-10 period only, fewer recorded comorbidities.
The accuracy of identifying diabetes status in CHD patients is improved in linked administrative health data by using at least 10 years of hospitalization history. Use of this method would reduce bias when measuring temporal trends in diabetes prevalence in this patient group. Concordance measures are as reliable in Indigenous as non-Indigenous patients.
行政数据是评估冠心病（CHD）等患者群体糖尿病患病率的有价值的数据源。本研究的主要目的是衡量 CHD 患者的病历与相关行政健康数据记录糖尿病的一致性，并评估一致性的时间差异。次要目的是确定用于识别该患者群体中糖尿病的最佳回溯期，土著患者的一致性是否存在差异，并确定行政数据中假阳性和假阴性的预测因素。
选择了 1998 年和 2002-04 年在西澳大利亚住院的 3943 例代表性 CHD 患者样本，并根据当时使用的国际疾病分类（ICD）版本（分别为 ICD-9 和 ICD-10）进行指定。对两个样本进行了粗患病率和一致性比较。使用行政数据比较选定的 CHD 入院（“索引入院”）和之前 1、2、5、10 或 15 年的任何住院期间记录的糖尿病状况，根据医院病历来估计一致性衡量标准。使用卡方检验和多变量逻辑回归模型确定一致性的潜在修饰符。
在 ICD-10 样本中，索引 CHD 入院时糖尿病的识别率低于 ICD-9 样本（敏感性分别为 81.5%和 91.1%，低估率分别为 15.1%和 4.4%）。在 ICD-10 期间，使用至少 10 年的住院记录，敏感性增加至 89.6%。土著患者的敏感性较高，特异性较低，且随着回溯期的增加，一致性也随之提高。与索引 CHD 入院时糖尿病假阴性相关的特征是择期入院、院内死亡、主要诊断，以及仅在 ICD-10 期间，记录的合并症较少。
通过使用至少 10 年的住院记录，在相关行政健康数据中，识别 CHD 患者的糖尿病状态的准确性得到了提高。当在该患者群体中测量糖尿病患病率的时间趋势时，使用该方法可以减少偏倚。在土著患者和非土著患者中，一致性衡量标准同样可靠。
