Sundararajan Vijaya, Quan Hude, Halfon Patricia, Fushimi Kiyohide, Luthi Jean-Christophe, Burnand Bernard, Ghali William A
Victorian Department of Human Servicest, Royal Melbourne Hospital, Australia.
Med Care. 2007 Dec;45(12):1210-5. doi: 10.1097/MLR.0b013e3181484347.
The Charlson comorbidity index has been widely used for risk adjustment in outcome studies using administrative health data. Recently, 3 International Statistical Classification of Diseases, Tenth Revision (ICD-10) translations have been published for the Charlson comorbidities. This study was conducted to compare the predictive performance of these versions (the Halfon, Sundararajan, and Quan versions) of the ICD-10 coding algorithms using data from 4 countries.
Data from Australia (N = 2000-2001, max 25 diagnosis codes), Canada (N = 2002-2003, max 16 diagnosis codes), Switzerland (N = 1999-2001, unlimited number of diagnosis codes), and Japan (N = 2003, max 11 diagnosis codes) were analyzed. Only the first admission for patients age 18 years and older, with a length of stay of >/=2 days was included. For each algorithm, 2 logistic regression models were fitted with hospital mortality as the outcome and the Charlson individual comorbidities or the Charlson index score as independent variables. The c-statistic (representing the area under the receiver operating characteristic curve) and its 95% probability bootstrap distribution were employed to evaluate model performance.
Overall, within each population's data, the distribution of comorbidity level categories was similar across the 3 translations. The Quan version produced slightly higher median c-statistics than the Halfon or Sundararajan versions in all datasets. For example, in Japanese data, the median c-statistics were 0.712 (Quan), 0.709 (Sundararajan), and 0.694 (Halfon) using individual comorbidity coefficients. In general, the probability distributions between the Quan and the Sundararajan versions overlapped, whereas those between the Quan and the Halfon version did not.
Our analyses show that all of the ICD-10 versions of the Charlson algorithm performed satisfactorily (c-statistics 0.70-0.86), with the Quan version showing a trend toward outperforming the other versions in all data sets.
查尔森合并症指数已广泛用于利用行政卫生数据进行的结局研究中的风险调整。最近，已发布了查尔森合并症的3种国际疾病分类第十版（ICD - 10）翻译版本。本研究旨在使用来自4个国家的数据比较这些ICD - 10编码算法版本（哈尔丰版、桑达拉扬版和泉版）的预测性能。
分析了来自澳大利亚（N = 2000 - 2001，最多25个诊断代码）、加拿大（N = 2002 - 2003，最多16个诊断代码）、瑞士（N = 1999 - 2001，诊断代码数量不限）和日本（N = 2003，最多11个诊断代码）的数据。仅纳入年龄18岁及以上、住院时间≥2天的患者的首次入院情况。对于每种算法，拟合了2个逻辑回归模型，以医院死亡率为结局，以查尔森个体合并症或查尔森指数评分作为自变量。采用c统计量（代表受试者工作特征曲线下的面积）及其95%概率的自助抽样分布来评估模型性能。
总体而言，在每个国家的数据中，3种翻译版本的合并症水平类别分布相似。在所有数据集中，泉版产生的中位数c统计量略高于哈尔丰版或桑达拉扬版。例如，在日本数据中，使用个体合并症系数时，中位数c统计量分别为0.712（泉版）、0.709（桑达拉扬版）和0.694（哈尔丰版）。一般来说，泉版和桑达拉扬版之间的概率分布重叠，而泉版和哈尔丰版之间的概率分布不重叠。
我们的分析表明，查尔森算法的所有ICD - 10版本表现令人满意（c统计量为0.70 - 0.86），泉版在所有数据集中均表现出优于其他版本的趋势。
