使用癌症登記系統資料預測台灣卵巢癌病患的存活情形

Abstract

背景：卵巢癌不論在台灣亦或是全球女性中，癌症死亡率排名都在前十名內，不同國家間診斷出卵巢癌的期別有所差異，而期別對於卵巢癌的存活率有顯著的影響，目前卵巢癌主要以發生率最高的上皮型卵巢癌(Epithelial ovarian cancer)為主，其中又可細分為漿液型(Serous)、黏液型(Mucinous)、子宮內膜型(Endometrioid)、明亮細胞型(clean cell)卵巢癌，不同亞型的卵巢癌發生率與存活率皆有差異，並且也具有種族間的差異，因此本研究期望藉由台灣癌症登記中心所蒐集的卵巢癌病人資料建立屬於台灣人的卵巢癌存活預測模型，再使用美國的癌症登記數據庫The Surveillance, Epidemiology, and End Results (SEER) 的資料來進行外部驗證及評估種族差異的影響，且希望藉由本研究成果能對於醫生診治不同卵巢癌亞型病人時能給予治療方案上的輔助。 方法：本研究主要使用台灣癌症登記系統(The Taiwan Cancer Registry, TCR)之長表資料，選取於2009年1月1日至2015年12月31日間診斷為卵巢癌的病人，追蹤至2017年12月31日，建立兩種存活預測模型，模型一納入常見臨床變項，模型二新增癌症特定因子變項，以期增加模型準確度；使用美國癌症研究所的SEER資料庫進行外部驗證，納入同樣的罹病年份與追蹤年數，可區分為白人、黑人、亞洲人的資料進行外部驗證及外推性研究。 結果：以所有死亡為終點的分析中，在模型一中達到顯著的變項有年齡(18－39, HR=1；40－49,HR=1.48, P=0.07；50－59, HR=1.53, P=0.054；60+, HR=2.59, P<0.001)；腫瘤亞型(Serous, HR=1；Clear cell, HR=1.85, P<0.001；Endometrioid, HR=0.9, P=0.438；Mucinous, HR=1.63, P=0.004)、腫瘤分級分化程度(Grade low, HR=1；Grade high, HR=2.12, P=0.002)、病理T(Pathological T1, HR=1；Pathological T2, HR=3.06, P<0.001；Pathological T3, HR=6.18, P<0.001)、病理M(HR=3.25，P<0.001)、是否進行化療(HR=0.69，P=0.027)與淋巴結侵犯比例(HR=2.34，P<0.001)，加入的交互作用項有：年齡*腫瘤分級分化程度、腫瘤亞型*病理N、腫瘤分級分化程度*病理M、病理N*病理M；在模型二達到顯著的變項有年齡(18－39, HR=1；40－49,HR=1.19, P=0.514；50－59, HR=1.01, P=0.78；60+, HR=1.96, P=0.009)；腫瘤亞型(Serous, HR=1；Clear cell, HR=2.93, P<0.001；Endometrioid, HR=1.68, P=0.002；Mucinous, HR=3.32, P<0.001)、腫瘤分級分化程度(Grade low, HR=1；Grade high, HR=2.05, P=0.003)、病理T(Pathological T1, HR=1；Pathological T2, HR=3.67, P<0.001；Pathological T3, HR=6.4, P<0.001)、病理M(HR=1.63，P<0.001)、治療後CA125數值(0-35 ug/ml, HR=1；35-100 ug/ml, HR=2.31, P<0.001；100+, HR=3.92, P<0.001)、是否殘存腫瘤(HR=2, P=0.005)，加入的交互作用項有：年齡*病理N、腫瘤分級分化程度*治療後是否殘存腫瘤；兩種模型的C-index數值不論在台灣或是美國資料集皆在0.7以上，且於存活預測的表現良好，存活與觀察間的差異大多數不超過5%。 結論：本研究所建立的兩種模型在預測台灣卵巢癌病人的存活上具有良好的表現，且在美國SEER資料中也同樣具有良好表現，基於種族的資料可證實本研究之模型具有外推性，期望可藉由本研究對於醫師在臨床治療決策的評估及醫病溝通上可達到助益。

Background: Ovarian cancer (OC) is the top 10 cause of cancer death not only in Taiwan but also in the whole world. The diagnosed stage information of OC varies from country to country, and it is a significant survival predictor of OC. Currently, epithelial ovarian cancer, which has the highest incidence, is the most common type of OC, and can be further divided into four subtypes, including serous, mucinous, endometrioid, and clear cell ovarian cancer. Unsurprisingly, different subtypes of OC have distinct incidence and survival rates, and these data are also specific to populations with distinct genetic backgrounds. To solve these issues, this thesis aims to develop a survival prediction model for OC in Taiwan by using the data collected from the Taiwan Cancer Registry. Next, the prediction model will be validated by using the data from the SEER dataset, a national cancer registry in the United States, in order to evaluate whether racial differences exist. Method: In this study, patients diagnosed with ovarian cancer between January 1, 2009 and December 31, 2015 were analyzed by using the data from the TCR, and all patients were followed until December 31, 2017. Two survival prediction models were developed by using different combinations of variables, respectively. The model 1 included the clinical variables in common to the TCR and the SEER datasets whereas the model 2 added cancer-specific variables from the TCR to improve the accuracy. The SEER dataset was used for external validation, and the analyzed patients had the same study period as what we had from the TCR data. That is patients with the same year of diagnosis and the year of follow-up time were analyzed. Lastly, the patients and can be further divided into three different groups according to their genetic ancestry, including white, black, and Asian. Results: In the analysis using all death as the endpoint, the variables that reached significance in the model 1 included age(18－39, HR=1;40－49,HR=1.48, P=0.07;50－59, HR=1.53, P=0.054; 60+, HR=2.59, P<0.001); histology subtype (Serous, HR=1; Clear cell, HR=1.85, P<0.001; Endometrioid, HR=0.9, P=0.438; Mucinous, HR=1.63, P=0.004),Tumor grade(Grade low, HR=1；Grade high, HR=2.12, P=0.002), Pathological T(Pathological T1, HR=1; Pathological T2, HR=3.06, P<0.001; Pathological T3, HR=6.18, P<0.001), Pathological M(HR=3.25，P<0.001), chemotherapy(HR=0.69，P=0.027), lymph node stage (HR=1.41, P=0.011), metastasis (HR=1.68, P<0.001) and lymph node ratio (HR=2.34，P<0.001).The interaction variables included age*Tumor grade, histology subtype* Pathological M, Tumor grade* Pathological M, Pathological N* Pathological M. The significant variables obtained from the model 2 included age (18－39, HR=1；40－49,HR=1.19, P=0.514；50－59, HR=1.01, P=0.78；60+, HR=1.96, P=0.009), histology subtype (Serous, HR=1；Clear cell, HR=2.93, P<0.001；Endometrioid, HR=1.68, P=0.002；Mucinous, HR=3.32, P<0.001),Tumor grade(Grade low, HR=1；Grade high, HR=2.05, P=0.003), Pathological T (Pathological T1, HR=1；Pathological T2, HR=3.67, P<0.001；Pathological T3, HR=6.4, P<0.001), Pathological M(HR=1.63，P<0.001), Carbohydrate Antigen 125 lab value after treatment (0-35 ug/ml, HR=1；35-100 ug/ml, HR=2.31, P<0.001；100+, HR=3.92, P<0.001), Residual Tumor(HR=2, P=0.005). The interaction variables included age* Pathological N, Tumor grade* Residual Tumor . The C-index values of these two models were above 0.7 in both TCR and SEER datasets. The models had good performance in survival prediction based on the calibration data, and the proportional difference between prediction and observation were mostly less than 5%. Conclusion: The two models developed in this study showed good performance in predicting survival of ovarian cancer patients in Taiwan and also in the SEER dataset. Notably, no huge racial differences were observed in predicting survival outcomes in OC patients, and thus we believe these two models can be useful for clinical treatment decision making and communication between patients and their physicians in the future.