運用統計模型以糞便潛血濃度預測大腸癌關係之研究

Statistical models for Prediction of Incident Colorectal Cancer in Association with Fecal Hemoglobin Concentration

莊鴻宇(Hung-yu Chuang)

指導教授 ： 陳秀熙

共同指導教授 ： 李永凌(Yung-ling Lee)

國立臺灣大學/公共衛生學院/流行病學與預防醫學研究所/碩士(2019年)

https://doi.org/10.6342/NTU201903985

摘要

背景:糞便免疫潛血測試是目前已知最好的非侵入性大腸癌篩檢工具，可以偵測糞便中的血紅素濃度。目前大部分大腸直腸癌的預測模型都是用來預測已經有症狀的病人，並非只是參與篩檢無症狀的族群。另外，重複參加篩檢而蒐集到的糞便免疫潛血測試結果往往並未列入預測的考量，模型如果加入糞便潛血隨著時間變化的狀態會更加準確。
目標:使用社區篩檢資料，我們希望可以達到：1.全國族群篩檢糞便潛血的濃度跟大腸直腸癌發生率的關係。2.研究考慮時間變項時，預測大腸癌發生率的準確度。3.使用有時間變項的糞便潛血濃度，建立馬可夫動態羅吉斯回歸模型預測大腸直腸癌的發生率。
材料與方法:本篩檢計畫由2004開始蒐集了50到69歲共920946個參與者的糞便免疫潛血資料，其中有9731個晚期大腸直腸癌，本研究計畫一直持續追蹤到2014。研究使用了三種方法來建立大腸直腸癌的預測模型：羅吉斯回歸模型、時間相依與非時間相依的Cox比例風險模型，以馬可夫過程建立的動態羅吉斯回歸模型。大腸直腸癌發生率的預測準確度則是使用時變與非時變的ROC curve來評估。
結果:在羅吉斯回歸以及Cox比例風險回歸模型裡呈現了糞便潛血濃度的差異與大腸直腸癌風險的關係。合併年齡以及性別之後，在非時間相依的分析裡AUC為0.75 (95%CI 0.71-0.76)。當糞便潛血濃度從1-5 ug/g上升到大於450 ug/g時，在時間相依的Cox比例風險回歸模型裡危險比也從0.73 (95% CI: 0.69 to 0.77)上升至22.5 (95% CI: 20.3 to 24.9)。類似的結果也可以在時間相依的布瓦松回歸模型裡將重複篩檢的糞便潛血濃度當作時間相依的變數時發現，由年齡、性別與隨時間變動的糞便潛血濃度構成的AUC在兩年、四年以及六年分別為0.912 (95%CI: 0.91-0.913), 0.879 (95%CI: 0.823-0.934), 及0.838 (95%CI: 0.794-0.884)。而在馬可夫過程動態羅吉斯回歸模型裡，糞便潛血濃度與腫瘤的發生率不只可以在臨床前期發現，也可以在臨床偵測期到臨床後期發現，考量動態糞便潛血的變化的多重階段馬可夫模型的AUC為0.880 (95% CI: 0.873-0.886)。
結論:這一個大型族群前瞻性世代研究不只呈現了糞便潛血濃度與大腸直腸癌發生率相關，更在動態羅吉斯回歸模型裡看到糞便潛血濃度在大腸直腸癌的引發及啟動都有顯著的意義，這些發現指出糞便潛血濃度在個人化的大腸直腸癌篩檢策略上可以有更好的應用。

關鍵字

糞便潛血 ； 預測 ； 大腸直腸癌 ； 風險

並列摘要

Background:Fecal immunochemical test (FIT) have proven to be the currently best non‐invasive test for CRC screening in asymptomatic individuals. The quantitative FIT is able to assess fecal hemoglobin concentration (f‐Hb). Most of CRC prediction model based on f-Hb were developed in symptomatic patients rather than in asymptomatic patients in CRC screening. Moreover, the use of subsequent FIT in repeated screen has been often neglected in development of CRC prediction model. Statistical models for the prediction model might be more precise when making allowance for time-varying property of f-Hb.
Aims:By using community-based integrated screening data, we aimed (1) to exam a dose-response relationship between f-Hb and incident CRC using the data of f‐Hb collected from a large population-based nationwide FIT screening;(2) to estimate the accuracy of predicting incident CRC by time-invariant and time-invariant approach;(3) to develop a Markov process with dynamic logistic regression model in association with time-dependent repeated f-Hb for CRC prediction.
Material and Methods:9,731 individuals with advanced CRC found in a cohort of 920,946 apparently healthy individuals, aged between 50 and 69 who had been invited to participate in population-based FIT screening with complete f‐Hb information since 2004. This cohort was followed-up from entry of study to the end of 2014 to ascertain colorectal cancer. Three statistical models including logistic regression model, Cox proportional hazards model with time-invariant and time-variant approach, dynamic logistic regression model with Markov process were used to develop CRC prediction model. The accuracy of predicting incident CRC was assessed by using time-invariant and time-variant receiver operating characteristics (ROC) curve analysis.
Results:A dose-response relationship between baseline f-Hb and the risk of developing colorectal neoplasia was demonstrated in both of logistic regression and Cox proportional hazard regression analysis. By combing baseline f-Hb with age and sex, the AUC was 0.75 (95%CI 0.71-0.76) using time-independent model. Compared with the group of f-Hb 1-5 ug/g, the impact of baseline f-Hb on the risk of CRC development increased from 0.73 (95% CI: 0.69 to 0.77)) of HR for undetected f-Hb to 22.5 (95% CI: 20.3 to 24.9) of HR for the group of f-Hb ≥ 450 ug/g by time-independent Cox proportional hazard regression analysis. The similar findings were observed by time-dependent Poisson regression model if taking repeated f-Hb measurements as time-invariant variable into account. The time-dependent AUC by age, sex, and time-varying f-Hb with 2 years, 4 years, and 6 years were 0.912 (95%CI: 0.91-0.913), 0.879 (95%CI: 0.823-0.934), and 0.838 (95%CI: 0.794-0.884), respectively.Using the dynamic logistic regression model with Markov process, a dose-response relationships not only noted for the incidence of entering preclinical detected phase (PCDP) but also the transition from PCDP to clinical detected phase (CP), both of which, in turn, yielded the occurrence of CP associated with f-Hb in a dose-response manner. The AUC of the ROC curve with multi-state Markov approach considering dynamic f-Hb changes was 0.880 (95% CI: 0.873-0.886).
Conclusions:Our results based on this large population-based prospective cohort study not only demonstrates a dose-response relationship between time-varying f-Hb and incident CRC but also show a dose-response relationship played in both of initiator and promoter of CRC when using the novel dynamic logistic regression model. These findings have significant implications for the better use of f-Hb concentrations on the individual-tailor CRC screening strategy.

並列關鍵字

Fecal immunochemical test ； colorectal cancer ； prediction

參考文獻

Meguid, R. A., Slidell, M. B., Wolfgang, C. L., Chang, D. C., & Ahuja, N. (2008). Is there a difference in survival between right- versus left-sided colon cancers? Ann Surg Oncol, 15(9), 2388-2394. doi:10.1245/s10434-008-0015-y

Auge, J. M., Pellise, M., Escudero, J. M., Hernandez, C., Andreu, M., Grau, J., . . . Group, P. (2014). Risk stratification for advanced colorectal neoplasia according to fecal hemoglobin concentration in a colorectal cancer screening program. Gastroenterology, 147(3), 628-636 e621. doi:10.1053/j.gastro.2014.06.008

Brenner, H., Kloor, M., & Pox, C. P. (2014). Colorectal cancer. Lancet, 383(9927), 1490-1502. doi:10.1016/S0140-6736(13)61649-9

Chen, L.-S., Yen, A. M.-F., Chiu, S. Y.-H., Liao, C.-S., & Chen, H.-H. (2011). Baseline faecal occult blood concentration as a predictor of incident colorectal neoplasia: longitudinal follow-up of a Taiwanese population-based colorectal cancer screening cohort. The Lancet Oncology, 12(6), 551-558. doi:10.1016/s1470-2045(11)70101-2

Chen, L. S., Yen, M. F., Wu, H. M., Liao, C. S., Liou, D. M., Kuo, H. S., & Chen, T. H. (2005). Predictive survival model with time-dependent prognostic factors: development of computer-aided SAS Macro program. J Eval Clin Pract, 11(2), 181-193. doi:10.1111/j.1365-2753.2005.00519.x