生存時間研究における調整型ランダムフォーレスト法

摘要

決定木を基本学習器として用いたアンサンブル学習法のなかで，とくにRandom Forest法は，有用性が広範に用いられている方法の一つである．他方，癌臨床研究の有効性の統計的評価に用いられる生存時間解析の分野において，アンサンブル学習法の研究はあまり行われていない．本報告では，生存時間研究におけるRandomForest法の有効性を検討した．さらに，予測確度の向上を意図して，個々の基本学習器を縮小回帰推定に基づいて調整する，調整型生存時間RandomForest法を提案した．調整型生存時間RandomForest法の性能は，文献事例により検討した．その結果，既存の生存時間RandomForest法に比べて良好な結果を示した．%One of the important themes in survival analysis is to explore prognoses factors that influence survival time. Recently, the tree-structured method has been applied to evaluate covariates (e.g., Crowley, 2004); however, it is well known that this method has provides poor prediction model. This problem could be improved by modeling many trees in a linear combination, namely, ensemble learning. The ensemble learning method is actively studied in machine learning and statistics. The random forest is popular method that is applied in many fields (e.g., bioin-formatics, environmentrics, and so on). In this presentation, we extended the random forest method to analyze survival data. Our proposed model has weight parameters, which are estimated by lasso (Tibshirani, 1996), for each tree. We call this method regurarized random survival forest method. Therefore, in regurarized random survival forest method, the trees (base learner) that strongly influence survival time will have large estimated parameter values; the parameters of trees that lack influence will be estimated as zero (pruning). Evaluation of regurarized random forest method, using simulated and real data sets, indicated that regurarized random survival forest method performs better than the ordinaly random survival forest method.