- 學位論文
利用分類樹進行胰臟癌病患的存活分群
Utilizing Classification and Regression Tree for Prognostic Grouping of Patients with Pancreatic Cancer
摘要
背景：不論在全球或台灣，胰臟癌皆是死亡率極高的癌症，組織型態上以腺癌(Pancreatic ductal adenocarcinoma, PDAC)為主要亞型，胰臟癌五年存活率僅約6%，其主因為發現罹癌時多為晚期，且胰臟位於人體後腹膜腔，手術治療困難，目前僅兩成病患發現時可以接受手術治療，其餘的八成皆須仰賴化療。本研究欲使用臨床變數及組織型態研究胰臟癌病人存活之間的異質性，主要欲回答在美國癌症聯合委員會(AJCC)第八版中佔有最多數的第三期病人(84%)的存活情形是否具有組內差異，及胰臟癌病患存活是否可以不同組織型態進一步分類。由於精確的病患分類是增進醫療效果的重要步驟，本計畫預期針對此二目標進行研究，進而提升治療效果。 方法：本研究使用台灣癌症登記資料庫研究，研究納入該資料庫中長表自2013年1月1日至2017年12月31日期間被診斷胰臟癌的病患之期別資訊，以及納入年報檔自1995年1月1日至2017年12月31日之數據以探討組織型態之存活差異。西方驗證資料方面則使用2000年1月1日至20016年12月31日之美國癌症登記資料 (The Surveillance, Epidemiology, and End Results, SEER)，以胰臟癌的初診日為研究基準日，再探討臨床變項例如期別、腫瘤分級分化(grade)等之預測效益。在臨床變項之存活分析上，本研究使用對數秩檢驗檢定組別之差異及Cox比例風險模式評估計算其風險比值，而進一步將存活差異進行分類時則採用決策樹進行分群。 結果：胰臟癌有手術的病患存活狀況比沒手術的病患顯著改善(p值小於0.0001)，因此是否進行手術為胰臟癌病患存活最重要的分層因子，進而才會是腫瘤本身特性之影響。在接受手術的胰臟癌第三期病患中，本研究透過決策樹挑選出兩個重要性最高的變數，為淋巴結擴散程度及腫瘤分級分化。研究結果經調整干擾因子後，針對淋巴結擴散程度，N1的死亡風險為N0的1.52倍(95%信賴區間為[1.18-1.94], p值小於0.0001)，N2的死亡風險為N0的2.30倍(95%信賴區間為[1.71-3.08], p值小於0.0001)；在腫瘤分級分化下，分化中度的死亡風險為分化良好的1.70倍(95%信賴區間為[1.03-2.79], p值為0.004)，分化不良組的死亡風險為分化良好的3.80倍(95%信賴區間為[2.25-6.39], p值小於0.0001)。透過交互作用項進一步進行病患之分類，依照病人存活狀況，我們使用決策樹可分類為四個組別，預後較高組的平均存活時間為23.5個月，而中高組、中低組及最低組的中位存活時間分別為18.4個月、14.5個月及9個月，四組間皆有顯著差異(p值小於0.0001)。胰臟癌病患的組織型態上，經由分組後可將病患之存活曲線可分為四個的存活區塊，第一組為最好的組織亞型，其四年存活率為90%，而第二組、第三組及四組的中位存活時間分別為41.0個月、13.5個月及5.4個月。而第二組、第三組及第四組的死亡風險分別為第一組的5.52倍、9.66倍及20.32倍，對數秩檢定結果為四組有顯著差異(p值小於0.0001)，本分群方式驗證在美國癌症登記資料庫亦有相同的結果。 結論：在本研究中，淋巴結擴散程度及腫瘤分級分化皆為第三期胰臟癌病患之重要存活預測因子，未來可提供臨床上治療決策之輔助，同時由於此結果能適用美國的胰臟癌病患，顯示此結果具有外推性，不僅限於台灣病患。組織型態之結果以固狀偽乳頭狀腫瘤預後狀況最佳，其餘組織型態則各有差異，代表在未來治療上應將組織型態納入治療考量。
並列摘要
Background: Pancreatic cancer is one of the most malignant cancer types with poor survival outcomes. Currently, physicians use the American Joint Committee on Cancer (AJCC) staging system to classify patients into different subgroups with distinct outcomes. However, the AJCC staging system classifies the majority of pancreatic cancer patients into stage III and dramatic heterogeneity is observed among the patients. The aim of this study is to provide precise classification systems by utilizing the clinical variables and histological subtypes of pancreatic cancer in order to help clinicians to make a better treatment plan. Method: We retrieved the data from the Taiwan Cancer Registry (TCR) and the US-based Surveillance, Epidemiology, and End Results (SEER). Stage-specific and histological overall survival were analyzed by the Kaplan-Meier survival curves and log-rank tests. Cox proportional hazard regression model was applied to determine the predictive effects of clinical variables. Patients were further divided into subgroups with distinct survival outcomes by using the classification and regression trees (CART). All results were validated by the SEER data. Results: Pancreatic cancer patients who underwent surgery showed significantly better survival outcomes than those patients without surgery, suggesting that whether a patient receives surgery is the most important factor. Subsequently, two clinical variables including lymph nodes involvement and cancer grade showed significant associations. Regarding lymph nodes involvement, patients with N2 had a higher risk of mortality (HR=2.30, 95%CI, 1.71-3.08, P<.0001) than patients with N0. Based on grades, patients with poorly-differentiated had a higher risk of mortality (HR=3.80, 95%CI, 2.25-6.39, P<.0001) than patients with well-differentiated. In accordance with the survival outcomes of patients, we could divide them into different groups by interaction terms with CART algorithm. The median survival times of the four groups were 23.5, 18.4, 14.5, and 9.0 months for ‘High’, ‘Moderately High’, ‘Moderately Low’, and ‘Low’ group, respectively (P<.0001). Histological subtypes were analyzed by previously described pipelines. Eight histological subtypes of pancreatic cancer patients could be classified into four different groups with distinct survival outcomes from high (Group1) to low (Group4). The 4-year survival rates of Group1 was 90%. The median survival of Group2, Group3, and Group4 were 41.0, 13.5, and 5.4 months (P<.0001). Similar results were observed by using the SEER data. Conclusion: Among the stage III pancreatic cancer patients, our results showed that lymph nodes involvement and cancer grades are important predicting factors for the survival outcomes. Further classification of pancreatic cancer patients based on their histological information is another important predictor of the survival outcomes, and the differences among these 4 groups are dramatically significant. In conclusion, taking the clinical and histological information into consideration may help physicians to make better therapeutic strategies.