使用交叉驗證風險評分方法在雙臨床試驗評估指標中檢測高效能次群組病患之探討

Abstract

CVRS2方法可以同時考慮臨床試驗的兩種評估指標，有效分群潛在受益於治療的受試者。透過兩個反應變數與高維度的臨床資料，計算受試者若接受治療，對兩個評估指標的風險分數。再依據二維的風險分數使用kmeans分群方法將受試者分群，其中包括只有單個評估指標受益於治療的受試者次群、兩個評估指標皆受益於治療的受試者次群以及無受益於治療的受試者次群。然而，我們發現CVRS2方法有些潛在的問題，像是結果不太穩定，共變量之間共線性過大可能會影響到分群表現等。對此我們替換了部分的流程，在事前使用elastic net或因素分析降低資料維度、重複執行10次CVRS2方法、使用model-based或階層式分群方法以及嘗試不同的風險分數計算方式，希望能讓最終的分群結果更理想。透過模擬資料，我們發現重複執行10次CVRS2方法、事前使用因素分析轉換資料等方法組合的表現皆優於原先CVRS2方法。使用勝算比風險分數的計算方法也能更明顯的分出只有單個評估指標受益於治療的受試者次群。最後，我們也應用了這些方法在一筆和牙周病孕婦患者的資料，兩個評估指標為嬰兒是否早產與嬰兒出生時體重是否過輕。修改後的CVRS2方法也成功的找到了對兩個反應變數都有顯著治療效果的受試者次群。

The existing CVRS2 method allows for simultaneous consideration of two trial endpoints, effectively clustering patients who have the potential to benefit from the treatment. By using two response variables and high-dimensional clinical data, the method calculates bivariate risk scores for subjects if they were to receive treatment with respect to two endpoints. Then, the patients are clustered using the kmeans clustering method based on the bivariate risk scores, resulting in 4 subgroups of patients: those benefiting from treatment based on a single endpoint, those benefiting based on both endpoints, and those not benefiting from treatment. However, we have noticed some potential problems with the CVRS2 method, including unstable results and when high collinearity exists among covariates, which may affect clustering performance. To address these problems, we have made several modifications to the process. These modifications involve reducing covariates dimension using elastic net or factor analysis prior to the CVRS2 method, repeating the CVRS2 method for 10 times, using model-based or hierarchical clustering method, and experimenting with another risk score calculation approach. These modifications aim to achieve more desirable results. Through simulation studies, we found that repeating the CVRS2 method for 10 times and using factor analysis to reduce the dimension of covariates before the analysis outperformed the CVRS2 method. Additionally, using the odds ratio risk score calculation approach helped distinguish the subgroup of patients benefiting from treatment based on a single endpoint. Finally, we applied these modified methods to a dataset of pregnant patients with periodontal disease. Preterm birth and low birth weight are the two trial endpoints. The modified CVRS2 method successfully identified a subgroup of patients exhibiting significant treatment effects in both response variables.