二階段抽樣病例世代研究設計估計多階段大腸直腸癌疾病自然史

Abstract

背景 評估多階段疾病的自然病史對於影響疾病進展之因子探討以及介入策略評估具有重要的角色。雖然隨機模型已廣泛應用於大規模族群篩檢資料進行效益評估以及預測多階段疾病，但在運用此一方法時所需要的族群資料在收集上仍然相當地耗費資源以及成本。再者，由於適用性與檢查費用之考量，經常無法對於所有篩檢群眾都施行昂貴的生物標記檢查，例如對於全國民眾施行基因檢測以找出對於某些疾病進展的高風險族群。因此，若使用病例世代的研究設計將可更有效益的運用篩檢資料進行研究。 目的 本論文的主要研究目的為發展廣義非線性回歸模型，且用於三階段的疾病進展模式，並將所獲得的資料結果與傳統的多階段隨機模型相互比較，以及將測量誤差(如敏感度)納入模型中加以考量。本研究將上述方法應用於台灣以糞便免疫化學檢驗(FIT)為工具的大腸直腸癌（CRC）全國民眾篩檢資料，同時將疾病的自然史參數以及糞便免疫化學檢查敏感度納入模型考量進行評估。 方法 在台灣全國大腸直腸癌篩檢計畫，從2004年1月1日至2009年12月31日邀請年齡在50〜69歲的民眾接受兩年一次的糞便免疫化學測試。在此期間，共1160884人接受篩檢，且重複篩檢率為28.3％。共有2494和195的大腸直腸癌個案於盛行篩檢和於兩年內接受的後續篩檢中被發現。在2年的追蹤期間內，共有694間隔癌。我們建構了以連續時間為基礎之三階段的馬可夫疾病進展模型並用於估計大腸直腸癌的自然病史。我們也基於上述三階段模式發展了廣義非線性回歸模型來推導疾病自然史參數，並發展出整合了多階段自然病史進展以及病例世代研究設計抽樣方式的方法使資料之運用更有效率。本研究亦探討在不同範圍的抽樣比例下利用所發展的方法評估疾病進展與使用整體資料之間之估計結果之差異。 結果 運用廣義非線性回歸模型於整體數據資料，所估計的大腸直腸癌臨床前期(PCDP)年發生率為每10萬人口75人(95％CI：每10萬人口58-92人)及臨床前期(PCDP)年進展率為0.31(95％CI：0.23 -0.40)，得到的平均臨床前期滯留期(MST)為3.23年(95％CI：2.5-4.35年)。運用自然病史模型於篩檢資料可估計的到MST約為3.2-4.3年，敏感度則為78-82%。進一步拓展模型於評估測量誤差與解釋變項的影響(性別和年齡)，結果顯示在PCDP發病率男性和老年人(60歲以上)之風險對比值分別為1.65 (95％CI：1.31-2.08)和2.05(95％ CI：1.61-2.64)，敏感度則為80％(95％CI：74-84％)。 基於上述運用於抽樣資料之模型，本研究藉由利用不同的抽樣比例所得到的樣本資料評估樣本數大小對於模型參數估計的影響顯示，隨著樣本數得減少，估計結果將趨於不穩定，對於解釋變項之統計檢定力亦隨之降低。藉由上述所建構的模型所估計得到的疾病自然史參數以及解釋變項對於疾病進展之影響，本研究進一步以模擬方法推估不同篩檢間隔時間對於篩檢間隔個案發生之影響。 結論 本研究發展了創新的多階段評估模型運用於病例世代研究中的抽樣資料，此一方法可以有效地評估疾病自然史的多階段結果，並將可能影響疾病進展的因素納入模型考量中。本研究亦將所發展的方法運用於台灣大腸直腸癌檢資料進行估計以及評估模型表現。本研究所提出的方法將可拓展為探究某些新的檢查與風險評估工具在不同疾病階段所扮演的角色之基礎，運用二階段病例世代研究設計配合所提出的方法將可達到有效率的運用族群篩檢資料之目的。

Background Elucidating multi-state disease natural history is of paramount importance for the identification of subjects at greater risk for disease progression, the determination of appropriate inter-screening intervals, and the evaluation of efficacy of interventions such as population-based screening program. While the application of stochastic models to estimate the force of multi-state disease progression using data on population-based screening program is well developed, the collection of such big data is quite costly. Moreover, it is often not feasible to accrue costly biomarkers such as genetic determinants based on the whole target population to quantify their roles played in the identification of subject at increased risk for disease progression. The application of case-cohort design is an alternative solution to address this issue with efficiency. Objectives The thesis aimed to develop a generalized non-linear regression model for fitting the data obtained from the three-stage design in comparison with the conventional multi-state stochastic model taking the measurement error such as sensitivity into account using data on the population-based fecal immunochemical test (FIT) for colorectal cancer (CRC) screening in Taiwan for the disease natural history for CRC and the sensitivity of FIT. Methods In the Taiwanese Nationwide Colorectal Cancer Screening Program, residents aged 50 to 69 years w consisting of 1160884 subjects with repeat screen rate of 28.3% were invited to receive a biennial FIT, between January 1, 2004 and December 31, 2009. A total of 2494 and 195 CRCs were detected during the prevalent screen and the subsequent screen, and 694 interval cancers were ascertained. A continuous-time, progressive 3-state Markov model was constructed for estimating the natural history of CRC. We developed a generalized non-linear regression model based on the three-state progression model for the derivation of the force driving the initiation and the progression of disease. A method incorporating both the nature of multistate disease progression and a sampling scheme based on a case-cohort design to utilize the data with efficiency was developed. The performance of the proposed method compared that of full data using a range of sampling proportions for the states of disease progression was then explored. Results Applying the generalized non-linear regression model to the full data, the estimated annual rate of CRC preclinical detectable phase (PCDP) incidence was 75 per 105 (95% CI: 58-92 per 10¬5) and that for PCDP progress was 0.31 (95% CI: 0.23-0.40) yielding a mean sojourn time (MST) for 3.23 years (95% CI: 2.5-4.35 years). The MST was around 3.2-4.3 years and the test sensitivity was 78-82% after fitting the data. The model considering measurement error and the effect covariates (sex and age) on PCDP incidence rate give the estimated hazard ratio for male and the elders (older than 60 years) of 1.65 (95% CI: 1.31-2.08) and 2.05 (95% CI: 1.61-2.64) and the sensitivity was estimated at 80% (95% CI: 74-84%). Applying the proposed algorithm for two-stage sampling scheme to the data derived by a series of sampling ratio using the covariate of sex and age demonstrated the influence of the reduction in sample size in terms of screen detected cancers and interval cancers. Conclusion A novel algorithm with a two-stage sampling design was developed to efficiently estimate the multistate outcome of the disease natural history and assess the effect of covariates on stage-specific transition rates. This new algorithm has been well demonstrated by using Taiwanese nationwide colorectal cancer screening program and can be easily extended to assess the causal effect of certain costly biomarkers on stage-specific transition on the basis of such a large population-based screened cohort.