Decision-tree Algorithm Optimize Hematopoietic Progenitor Cell-based Prediction in Peripheral Blood Stem Cell Mobilization

利用決定樹流程改良以造血前驅細胞為基礎對周邊血液幹細胞動員之預測

李俊毅(Chun-Yi Li) ； 余垣斌(Yuan-Bin Yu)

《童綜合醫學雜誌》 11卷1期 (2017/06) Pp. 10-19

摘要

Background and Objectives : Enumeration of hematopoietic progenitor cells (HPC) using an automated hematology analyzer provides rapid, inexpensive, and less technically dependent prediction of peripheral blood stem cell (PBSC) mobilization. This study aimed to incorporate HPC enumeration along with other predictors for optimizing a successful harvest. Materials and Methods : Between 2007 and 2012, 189 consecutive patients who proceeded to PBSC harvesting with a preharvest HPC ≥ 20 × 10^6/L were recruited. A failed PBSC mobilization was defined as ＜ 2 × 10^6 CD34^+ cells/kg. Variables predicting a successful harvest identified by multivariate logistic regression and correlation analysis were subjected to classification and regression tree (CART) analysis. Results : A total of 154 (81.5%) patients successfully achieved mobilization of CD34^+ cells (median 8.18 × 10^6 CD34^+ cells/kg). Five independent host predictors including age ≥ 60, a diagnosis of solid tumor, prior chemotherapy cycles ≥ 5, prior radiotherapy, and mobilization with G-CSF alone or high-dose cyclophosphamide, as well as laboratory markers including HPC and mononuclear cell (MNC) counts, were used for CART analysis. The number of host predictors with a cutoff at two, HPC cutoff at 28 x 10^6/L and MNC cutoff at 3.5 x 10^9/L were best discriminative for successful prediction. In the decision tree algorithm, patients predicted as good mobilizers (0 to 2 risk factors) had a higher success rate (150/169, 88.8%) than that (4/20, 20.0%) of those predicted as poor mobilizers (3-5 risk factors). Moreover, patients predicted as good mobilizers and further with a HPC enumeration ≥ 28 x 10^6/L had a high probability of achieving successful mobilization (138/148, 93.2%). Conclusion : Our CART algorithm incorporating host predictors, HPC enumeration and MNC count may improve prediction and thus increase the success of PBSC mobilization. Further prospective validation is necessary.

關鍵字

hematopoietic progenitor cell ； mononuclear cell ； peripheral blood stem cell harvest ； autologous stem cell transplantation ； classification and regression tree analysis

並列摘要

背景：以自動血液分析儀來計算造血前驅細胞（HPCs）的數目以預測周邊血液幹細胞（PBSCs）的動員程度是一個快速、較不昂貴、較不需技術導向的方式。本研究在探討如何整合HPCs 及其他預測指標以成功的收集幹細胞。方法：我們分析了自2007 至2012 年189 位收集PBSCs 的患者，患者在收集前HPCs 均大於20x10^6/L。收集失敗的定義是每公斤體重少於2x10^6 個CD34^+ 的細胞。我們以多變數邏輯式回歸和相關性分析找出預測成功收集的變數，最後以分類和回歸樹（CART）分析。結果：總共有154 位患者成功收集到足量的CD34^+ 細胞（中位數為8.18x10^6/ 每公斤）。五個獨立的宿主因素包括年紀大於60 、固態腫瘤診斷、先前化學治療超過五次、先前曾經接受放射線治療及單獨使用白血球生長因子或高劑量cyclophosphamide 作為動員處方，加上實驗室方面的HPCs 及單核球數這兩個指標，進入最後的CART 分析。兩個宿主危險因子、HPCs 數28x10^6/L 及單核球數3.5x10^9/L 是最能用以成功區別收集成不成功的分界值。在這個決定樹的流程中，有0 至2 個宿主危險因子的患者（預測其為良好動員者）其收集成功率遠高於3 至5 個危險因子的患者（88.8% vs. 20%）。在這些良好動員者中，若其HPCs 數大於28x10^6/L，則有更高的機會（93.2%）可以收集成功。結論：這個CART 流程整合了宿主因素、HPCs 及單核球數可望改善以往的預測模型而提高成功收集PBSCs 的機會，未來仍需前瞻性的研究加以驗證。

並列關鍵字

造血前驅細胞 ； 單核細胞 ； 周邊血液幹細胞收集 ； 自體造血幹細胞移植 ； 分類和回歸樹分析