應用全民健保資料庫比較不同合併症指標對中風病人醫療資源耗用之預測表現

目的：中風為影響人類健康及造成財政負擔的重大疾病，因此有關中風醫療費用的相關研究眾多，而在我國運用全民健保申報資料已成為醫療相關研究的重要資料來源。綜觀國內外中風相關研究文獻，均會於研究設計中採用合併症指標作為校正因子之一。然而合併症指標眾多，如何依研究族群合併症分佈情形，或是對欲探討的醫療照護結果進行分析時，可選擇出較合適的合併症指標(Comorbidity Measurements)，以利對研究對象做最佳化校正，是為重要議題。但國內目前針對不同學者所發展之合併症指標適用本國中風病人及醫療費用預測力表現的實證研究則闕如。方法：本研究採回溯性世代研究，使用全民健康保險2005年承保抽樣歸人檔(LHID2005，資料期間1996年-2010年)，選擇2002、2005及2008三個年度的首次中風住院病人為研究對象，採用四種均以ICD-9-CM疾病編碼系統的合併症指標（Deyo's CI、Romano's CI、D'Hoore's CI及Elixhauser's CI）比較其對中風病人出院後一年後產生高醫療費用情形之預測力。控制變項包括病人性別、年齡、首次中風住院有無手術及首次中風住院天數。將僅包含控制變項的模式作為基本模式，檢視增加不同合併症指標後對模式預測力的影響。其中並探討影響合併症指標預測力的因子，各別為合併症指標資料分析類型（權重及類別）、判斷合併症的資料來源（門診及住院）與期間（當次住院及併前1年門診與住院資料），以及時間趨勢。所使用的主要統計方法為邏輯斯迴歸分析，利用c統計量來呈現四種合併症指標預測力及相對性的比較結果。結果：四種合併症指標對於基本模式的預測力均有增進效果（c值增加至0.677至0.762），且大多數達顯著差異（p < 0.05）。以類別模式分析之預測力優於權重模式。合併症指標在「權重模式」下，預測力主要以Romano's CI表現較佳(c值0.663-0.746)；在「類別模式」下，預測力以Elixhauser's CI(c值 0.676-0.762)表現較佳。合併症指標模式以不同資料來源與期間進行計算時，整體而言預測力有提升情形，但具有顯著效果者不多，預測力以「當次住院併前1年門住診」最高。各合併症指標模式預測力隨著年度增加而有增加趨勢，主要以2002年至2005年間預測力提升有顯著差異（c值增加最大比例為12.3%，p < 0.05）。結論：基於本研究結果發現，合併症指標對於預測中風病人出院後產生高醫療費用情形，為重要校正因子之一。且以「類別模式」分析時之預測力較「權重模式」時為佳，但若考量計算便利性因素而採用權重模式分析時，建議可選用Romano's CI作為校正工具；若要以類別模式分析，建議可選用Elixhauser's CI，但此方式不適用於樣本數過小之研究。增加不同資料來源與期間後對預測力的增進效果相當有限，故建議可以視資料取得難易度、研究對象合併症分佈及欲探討之醫療照護結果做適當選擇。衛生主管機關擬定關於中風病人醫療費用相關政策或給付制度時，建議應考量病人合併症負擔做不同權重調整，以及考量衛生政策重大性及影響力，選用不同分析模式及表現較佳的合併症指標。

關鍵字

合併症； CCI ；次級資料；行政申報資料；中風；醫療費用

並列摘要

Purpose: Stroke represents a major public health problem, and placing a economic burden on family and community. A lot of about stroke medical expenditure study previously, and National Health Insurance databases have become an important resource for health services’ studies in our country. Stoke related studies that use administrative data to investigate population-based health outcomes often adopt risk-adjustment models that include comorbidities, conditions that coexist with the index disease. Several methods to measure comorbidities using administrative health care data exist. Although each of these comorbidity measures has been shown to perform acceptably in the general adult population and has been used in health services research, it is not clear which provides optimal measures of comorbidities in stroke populations. It is important to find a comorbidity measure with better performance for use with administrative data. However, there has been rare study to investigate the performance of the various available claims-based indices of comorbidity in Taiwan. Our objective was to compare the performance of comorbidity indices for predicting medical expenditure in patients with stroke. Methods: Data are collected from the National Health Insurance inpatient claims data. The study conducted three retrospective cohort study (2002, 2005 and 2008) in Taiwan. Four comorbidity measures were considered: Deyo’s Comorbidity Index (CI), Romnao’s CI , D`Hoore’s CI, and Elixhauser’s CI, which based on the International Classification of Disease, 9th Revision, Clinical Modification codes in the claims data, and then compared all comorbidity measures models predicting medical expenditure for patients with stroke 12 months after hospital discharge. The baseline model included age, gender, whether the patient received surgery or not, and the length of stay which were compared adjustment for different comorbidity models. Several implementation strategies, defined by altering data analysis type (indices and categories), two data sources (inpatient and outpatient vsist), two time frame (including the index hospitalization as well as the index and prior 1-year inpatient and outpatient claims.) and time trend, were assessed for each instrument. Discrimination was assessed with the c-statistic of multiple logistic regression were used to compare performance. Results: Compared with the baseline model, all four comorbidity indices improved the model prediction for having higher c-statistic values that ranged between 0.677 and 0.762, and most of the comorbidity indices significantly improved the model prediction by a statistically significant change in the c-statistic. The measures of comorbidity were implemented as categories (the presence or absence of the comorbid illness) have batter prediction than as indices (weighted sum of comorbidity indicators). When measures of comorbidity were used as indices, better discrimination was achieved with the Romano’s CI (c = 0.663-0.746). Comorbidity measures models with categories of comorbid illness identified using the Elixhauser’s CI achieved substantially higher levels of discrimination than those using the other method (c = 0.676-0.762). Although most comorbidity measure performed most favorably when adding information both inpatient and outpatient claims and when comorbidities were considered during the preperiod, varying data source and time frame had trivial effects on model performance. Most of the comorbidity measures prediction increased over the period 2002-2005, and the largest percentage increase in predictive discrimination for any model was the 12.3%. Conclusions: The comorbidity measures are important factors for predicting medical expenditure in patients with stroke. The measures of comorbidity were implemented as categories have batter prediction than as indices. Consider conveniently used and counted, the Romano’s CI has better discrimination used as index. When the sample size is large enough, the Elixhauser’s CI could be implemented as the categories components. While information drawn from both prior inpatient and outpatient records can be used to identify more patients with comorbid illness and more comorbid illness among patients, including this additional information does little to improve the statistical performance of risk adjustment models using either method. Investigators should choose among these comoebidity measures based on dataset availability, comfort with the methodology, comorbidity illness distribution and outcomes of interest. The department of health can consider comorbidity burden and application to set up more detail payment for stroke patients.