以支持向量機分析HLA分型與各紅球血型系統異體抗體的關聯性

張志昇(Jhy-Sheng Chang)

指導教授 ： 蘇家玉

共同指導教授 ： 李元綺(Yuan-Chii G. Lee)

臺北醫學大學/醫學科技學院/醫學資訊研究所/碩士(2013年)

https://doi.org/10.6831/TMU.2013.00079

摘要

目的：
人類各紅血球血型系統產生不同異體抗體的變異性，常因各種族而不同，和紅血球的抗原比率較無關係。依文獻資料顯示，產生紅血球異體抗體的免疫能力和個體的人類組織相容性抗原(Human Leukocyte Antigen , HLA)分型有極大的相關性。臺大醫院HLA資料庫及紅血球異體抗體資料庫均為經常性輸血所衍生的資料庫。因此本研究希望能利用資料探戡技術來分析這兩組資料庫，希望能找出不同HLA分型產生各種紅血球血型系統異體抗體的關聯性。

方法：
本研究根據臺大醫院人類組織抗原HLA資料庫，與相關的紅血球抗體鑑定結果資料庫，在資料處理部分進行重新編整HLA及紅血球異體抗體資料庫的格式。接下來運用資料探勘中之支持向量機（support vector machines, SVM）技術分析HLA Class I 和Class II分型與紅血球異體抗體關聯性，資料探勘過程包括資料權重選取、SVM參數調整及重要特徵選取等步驟。此外本研究也進一步利用HLA分型資料來預測腫瘤相關疾病，如肝性疾病hepatic disease、骨髓性血癌myeloid leukemia、淋巴性血癌lymphoblastic leukemia。希望能利用資料探勘工具和機器學習技術來分析HLA class I, class II和紅血球異體抗體間的關聯性。

結論與展望：
實驗結果證實利用HLA class I和class II分型來預測紅血球血型系統所產生異體抗體，以SVM資料探工具評估可達到91.8% ~ 99.7% 的準確率。此外以HLA分析預測腫瘤相關疾病之準確率亦可達55.9%~96.2%。研究中以特徵選擇演算法選擇出與紅血球異體抗體高度相關之HLA分析與臨床專業知識符合，顯示以資料探勘分析能找出有鑑別力的特徵。利用HLA資料庫與紅血球異體抗體資料庫進行資料探勘可以找出台灣地區HLA和產生各種紅血球血型系統異體抗體的相關性，進而可預測輸注紅血球產生異體抗體的機率，並提供預防性輸血(prophylactic transfusion ) 的功用。

關鍵字

紅血球異體抗體, HLA分型, 資料探勘, 支持向量機

並列摘要

Background:
Human blood group systems cause different red blood cell (RBC) alloantibodies due to different ethnic people instead of blood group antigen diversity. Many studies have shown that the immune potent of alloantibodies in blood group systems and human leukocyte antigen (HLA) phenotypes have close relationship. However, it is time-consuming and labor-intensive to manually map HLA phenotypes to red blood alloantibodies. Using data mining techniques to analyze HLA typing and predict alloantibodies or diseases has become highly desirable.

Method:
This study extracts and organizes datasets from HLA phenotypes and RBC alloantibody databank in Taiwan Blood Foundation. After that, we incorporate a support vector machine package, LibSVM, to analyze the correlation among HLA class I, class II, and RBC allo-immune. Data mining techniques are applied for weight adjustment, parameter tuning and feature selection. Moreover, we also use HLA phenotypes to predict diseases, such as hepatic disease, myeloid leukemia, lymphoblastic leukemia.

Results and conclusion:
Experiment results demonstrate that the proposed method achieves 91.8%~99.7% overall accuracy by support vector machine tools for alloantibody prediction. Moreover, it also attains overall accuracy 55.9%~96.2% in predicting related diseases. Based on data mining techniques, we can also identify discriminative features to predict RBC allo-immune alloantibody. In summary, the proposed method can be utilized to analyze correlation between HLA and blood group system alloantibody databank and provide useful insights to prophylactic transfusion in Taiwan.

並列關鍵字

RBC allo antibody, HLA typing, data mining, SVM

參考文獻

4. 陳嘉文, Analysis Antibody Identification Data from Blood Donation and Transfusion Services in Kaohsiung. J Biomed Lab Sci, 2007. 19(3): p. 99-105.

5. 張鳳娟, 馬偕紀念醫院病人異體抗體的頻率. Formosan J Med 2004. 2004(8): p. 755-759.

1. Takeshita, A., et al., Collaborative study of irregular erythrocyte antibodies in Japan: results from the Japanese study group of allo-immunity and antigen diversity in Asian populations. Transfus Apher Sci, 2010. 43(1): p. 3-8.

2. Jan, R.-H., Incidence of Alloantibodies in Transfused Patients in Eastern Taiwan,. TZU CHI MED J 2009. 21(1).

6. Yang, K.L., et al., High-resolution human leukocyte antigen (HLA) haplotypes and linkage disequilibrium of HLA-B and -C and HLA-DRB1 and -DQB1 alleles in a Taiwanese population. Hum Immunol, 2009. 70(4): p. 269-76.