辨識族群中肝細胞癌高風險群體：從模式建構到決策分析

Abstract

對族群廣泛施行肝癌篩檢並不符合成本效益，藉由肝癌高風險群體的明確界定並對其進行追蹤檢查可提升肝癌病患之辨識效率。過去美國肝病研究學會主要根據肝癌病毒感染、性別與年齡定義肝癌高危人群，並建議該些特定群體定期進行肝癌監視篩檢。然目前的臨床指引未具有未來肝癌進展風險預測的功能，且肝細胞癌發展涉及許多的危險因子之獨立或共同作用，因此藉由有限數量之可預測性因素組建肝癌風險預測模型，進而確實定量多重肝癌病因之不等暴露程度而導致的肝癌結果差異，可強化肝細胞癌高危險群之選擇過程。然而，目前缺乏針對族群廣泛篩檢問題而研擬之肝癌高風險群體辨識模型，從模型建構、外推評估至決策應用等完整的研究與探討仍然有限。為了於族群層次中，充分準確地早期辨識無症狀之肝細胞癌高風險群體並促進篩檢決策，本論文共包含四部份子研究進行了一系列分析探討：(1) 肝細胞癌之危險因子分析；(2) 一般人群肝細胞癌風險預測模式之建構；(3) 一般人群肝細胞癌風險預測模式之評估；(4) 肝細胞癌篩檢優先選擇標準。 研究一、肝細胞癌之危險因子分析 目的 利用一個龐大且多中心的長期前瞻性追蹤世代資料庫執行肝細胞癌之危險因子分析。 材料方法 本研究匯集三個不同來源與特性組成之前瞻性追蹤世代資料，將12377位年齡介於20-80歲間未曾罹患肝細胞癌的成人納入分析。收案時即對研究個案以結構式問卷進行訪視，並收取血液檢體做病毒學及生化相關檢驗。研究世代之新發肝癌病例乃透過定期追蹤的例行健康檢查或連結國家癌症登記及死亡證明系統做確立。於191240.25之追蹤人年中，共計387名個案被確認為肝細胞癌新發生病例。利用Kaplan-Meier法與Cox迴歸模型評估各個危險因子和肝細胞癌後果之關聯性。使用模型選擇法辨別統計學顯著的危險因子，藉以產生最優預測模型。 結果 男性和女性之10年肝癌累積風險分別為2.3%與0.7%。隨年齡增長肝癌之10年風險隨之增加，年齡20-39, 40-49, 50-59與≧60歲分別為0.8%, 2.0, 3.7%與4.6%。除人口學因子，血清ALT濃度、慢性肝臟疾病史、一等親肝細胞癌家族病史、終生菸害累積暴露量、肝炎病毒感染、酒精飲用與糖尿病史皆為肝細胞癌之顯著危險因子。最佳模式組合的多變項Cox迴歸分析結果顯示，危險因子相互調整其作用後男性相對女性風險比值為3.02 (95%信賴區間: 1.95-4.68)；年齡每增加一歲，其風險比值增加1.08 (95%信賴區間: 1.07-1.09)；以血清ALT濃度<25 IU/L為參考值，≥25 IU/L之風險比值為3.54 (95%信賴區間: 2.84-4.41)；具有慢性肝臟疾病史與一等親肝細胞癌家族病史相對於不具有者之風險比值分別為2.67 (95%信賴區間: 2.02-3.53) 與2.05 (95%信賴區間: 1.60-2.63)；以抽菸包年數<18為參考值，≥18包年其風險比值為1.47 (95%信賴區間: 1.13-1.91)；B型肝炎病毒感染狀態陽性相對陰性之風險比值為12.89 (95%信賴區間: 7.86-21.15)。 結論 吾發現性別、年齡、血清ALT濃度、慢性肝臟疾病史、一等親肝細胞癌家族病史、終生菸害累積暴露量與肝炎病毒感染會影響罹患肝細胞癌之危險性大小。 研究二、一般人群肝細胞癌風險預測模式之建構 目的 從一般人群的角度為出發點建構族群之肝細胞癌風險預測模式，並且發展簡易使用的風險分數系統。 材料方法 本研究使用性別、年齡、肝炎病毒感染狀態、血清ALT濃度、慢性肝臟疾病史、一等親肝細胞癌家族病史和終生菸害累積暴露量進行模型組建。藉由隨機分割技術將全世代2/3樣本用於模型與風險分數建構 (訓練集)，其餘1/3樣本用以驗證預測效能 (驗證集)。將Cox迴歸模型得到的迴歸係數經由加權轉換為整數的風險計分，並計算各種風險計分下10 年內發生肝癌的預測機率，計分系統與對應的肝癌預測機率將被轉換成對照圖 (nomograms) 以利使用。模式的預測性效能以兩個面向來評估，包括：鑑別能力 (discrimination) 是以重複取樣之bootstrap Harrell’s c統計量進行估算，而校準度 (calibration) 是以校準曲線來量測。研究根據風險評分的三分位計算各風險層之肝細胞癌10年累積風險。 結果 模型1使用性別、年齡與ALT；模型2使用性別、年齡、ALT、慢性肝臟疾病史、一等親肝細胞癌家族病史和終生菸害累積暴露量；模型3使用性別、年齡、ALT、慢性肝臟疾病史、一等親肝細胞癌家族病史、終生菸害累積暴露量和HBsAg；模型4使用性別、年齡、ALT、慢性肝臟疾病史、一等親肝細胞癌家族病史、終生菸害累積暴露量與HBsAg或Anti-HCV建構模型。模型1-4其Harrell’s c統計量值分別為0.77，0.79，0.84與0.84。從訓練集產生的肝癌風險分數可以準確地將參與者分類至適當的風險分層，隨風險三分位層遞增肝細胞癌風險隨之遞增加 (p<0.0001)，於整體、訓練或驗證世代中可觀察到累積風險有一致的趨勢。 結論 利用不同常見危險因子組合建構之4個風險預測模型可鑑別長期肝細胞癌之發生。 研究三、一般人群肝細胞癌風險預測模式之評估 目的 評估不同風險模型產出之肝癌風險分數其性能與可外推性。 材料方法 研究透過赤池訊息準則與貝氏訊息準則評估模型適配。並將所有研究對象分層於獨立的子集進行次樣本分析，包含不同世代來源、性別、年齡與病毒感染狀態，以探討各風險分數對不同特性次族群之肝癌區分能力變化。為探究模式風險分數對於肝細胞癌的預測鑑別準確度受族群HBsAg陽性比例變化之影響吾進行敏感性分析，研究於不同的族群HBsAg陽性比例 (10%, 15%, 20%) 假設下隨機抽樣產生了100組資料樣本，並觀察c統計量之變異分佈。本研究也利用新穎的風險重分類指標進行模式評比，包括重分類改善指標淨值 (net reclassification improvement) 與相對整合性鑑別度改善指數 (relative integrated discrimination improvement index)。 結果 各模型分數對肝癌5年和10年的風險預測表現相似且良好，c統計量介於0.78-0.86，模型分數即便限制於三個不同的次子世代 (0.74-0.86) 或根據性別 (0.75-0.92)、年齡 (0.65-0.85) 與病毒感染狀態 (0.65-0.82) 分層仍保有良好的預測能力。敏感性分析結果顯示風險分數的預測準確度度受族群HBsAg陽性率變化之影響程度小，10年肝癌風險預測之c統計量值四分位距範圍為0.74-0.78 (模型1)，0.76-0.80 (模型2)，0.84-0.88 (模型3) 和0.87-0.90 (模型4)。模型2之赤池訊息準則與貝氏訊息準則最小。與模型1相比，增加慢性肝臟疾病史、一等親肝細胞癌家族病史和終生菸害累積暴露量等因子之模型2能更正確地進行肝癌風險分類，對5年與10年內肝細胞癌事件之NRI (30.8%與30.9%) 與rIDI (30.1%與17.0%) 指標達統計顯著改善。 結論 模型2為最佳配適模型。風險模型2衍伸之風險分數，於整體或次樣本群中對肝細胞癌的預測性能皆優於或相似於其他三個模型。 研究四、肝細胞癌篩檢優先選擇標準 目的 探討模式效益並根據肝癌風險分數制定肝細胞癌篩檢之優先選擇標準。 材料方法 利用肝癌風險分數三分位數將研究對象均分至三個風險層。篩檢負荷 (screening load) 與需要被篩檢的人數 (number needed to test, NNT) 被用於評估篩檢效率。臨床效用分析用來估計所有可能的10年肝細胞癌預測風險閾值下，對應之臨床效益與相對效用。為制定篩檢選擇標準，研究利用模型2風險分數三分位數與主要病毒因子感染狀態合組將研究對象進行分層。在不同的篩選標準下分別計算5年和10年內肝癌發生的預測敏感性和特異性。 結果 隨風險程度增加篩檢效率隨之漸增，模型2衍生之最高風險層其NNT最低 (16)，最低風險層其NNT最高 (257)。於10年之肝癌發生風險閾值為2%之下，模型2風險分數比起當前AASLD的篩檢準則可獲得更高的臨床效益與相對效用。以10年之肝癌發生風險2%為高危險群挑選標準，吾所建議之肝癌篩檢優先選擇標準相較於當前的AASLD準則，對肝癌罹病的預測診斷準確性在特異度約略微降低下 (5年：67.3%比71.4%，10年：67.8%比71.8%；對早發性肝癌之5年：66.9%比71.0%，10年：67.0%比71.0%)，敏感度可明顯增加 (5年：90.2%比76.2%；10年：89.4%比76.8%)，尤其對小於50歲發病之早發性肝癌更有效的大幅提升 (5年：76.9%比48.7%；10年：79.4%比40.6%)。 結論 模型2衍生風險分數結合肝炎病毒感染狀態所制定之肝癌篩檢優先選擇標準比AASLD標準可提高敏感度。

The feasibility of population-wide hepatocellular carcinoma (HCC) screening is in need of a cost-effective strategy. Persons at high risk of developing HCC should be offered entry into surveillance programs, which is generally used to facilitate a cost-effective screening program. According to current authoritative guidelines, routine screening for HCC is recommended for selected populations based on hepatitis virus infection status, gender and age. However, these variables can not lead to an accurate prediction of HCC risk. Building a predictive model with incorporation of multiple important risk factors to quantify an individual’s risk can strengthen the selection process for screening of HCC. To date, however, studies of HCC-risk model with relevance to population-wide screening is lacking. To develop an interactive risk assessment tool to promote risk communication and risk stratification among average-risk population for priority criteria of selecting individuals for HCC screening, a series of analyses have been performed based on a longitudinal cohort database on HCC etiology. This thesis consists of four parts: (I) Prospective analysis of risk factors for HCC; (II) Constructing HCC-risk model to develop risk scoring system in average-risk population; (III) Assessment of model performance and generalizability of HCC-risk scores; and (IV) Evaluation of priority criteria for selecting individuals for screening of HCC. Part (I): Prospective analysis of risk factors for HCC Specific aim: To conduct a prospective study on risk factors for HCC. Materials Methods: The database consists of 12377 adults without HCC (aged 20-80 years) at recruitment from three large prospective cohorts. Participants were personally interviewed using structured questionnaires and provided blood samples for serological tests. Newly-developed HCC was ascertained through follow-up examination and data linkage with national cancer and death registry profiles. After 191240.25 person-years of follow-up, 387 incident cases of HCC were identified. Kaplan-Meier method and Cox model were used to investigate the associations of risk factors and HCC. Stepwise regression was used to derive best-fitted model. Results: The 10-year cumulative risks were 2.3% for men and 0.7% for women. There was an increasing trend of cumulative risk with increasing age, with the 10-year cumulative risks were 0.8%, 2.0, 3.7% and 4.6% for age groups 20-39, 40-49, 50-59 and ≧60 years, respectively. Besides demographic factors, significant risk factors for HCC were serum alanine transaminase (ALT) levels, past history of chronic liver disease (CLD), first-degree family history of HCC, cumulative smoking, chronic hepatitis virus infection, alcohol intake, and prior diabetes. A multivariate stepwise regression revealed that adjusted hazard ratios (95% confidence interval) of HCC were 3.02 (1.95-4.68) for men vs. women, 1.08 (1.07-1.09) for increment per year in age, 3.54 ( 2.84-4.41) for ALT ≥25 vs. <25 IU/L, 2.67 (2.02-3.53) for with vs. without history of CLD, 2.05 (1.60-2.63) for with vs. without first-degree family history of HCC, 1.47 (1.13-1.91) for smoking ≥18 vs. <18 pack-years, and 12.89 (7.86-21.15) for HBsAg positive vs. negative. Conclusions: Male gender, age, elevated serum ALT levels, past history of CLD, first-degree family history of HCC, pack-years of smoking, and chronic hepatitis virus infection were significantly associated with increased risks for HCC risk after adjustment for each other. Part (II): Constructing HCC-risk model to develop risk scoring system in average-risk population Specific Aim: To develop risk scoring systems by using common risk factors widely obtainable in average-risk population. Materials Methods: Gender, age, hepatitis virus infection status, ALT, history of CLD, first-degree family history of HCC, pack-years of smoking were used to construct risk models. Two thirds of the pooled cohort was randomly allocated for model development (training set), and the remaining one third for model validation (validation set). Regression coefficients derived from each variable from Cox models were used as weight to derive risk score. Cox model-derived predicted 10-year HCC risks were converted into nomograms. Discrimination and calibration were evaluated with the use of Harrell’s c-statistics on the basis of bootstrap resampling and calibration curve. Cumulative 10-year risk was calculated according to tertiles of risk scores. Results: Four models were constructed: model 1 with gender, age and ALT; model 2 with gender, age, ALT, history of CLD, first-degree family history of HCC and pack-years of smoking; model 3 with gender, age, ALT, history of CLD, first-degree family history of HCC, pack-years of smoking, and HBV infection.; model 4 with gender, age, ALT, history of CLD, first-degree family history of HCC, pack-years of smoking, and hepatitis B or C virus infection. Overall, model 1-4 had Harrell’s c-statistics of 0.77, 0.79, 0.84 and 0.84, respectively. In addition, there was an increasing trend in the cumulative HCC risk with increasing tertiles of risk scores (p<0.0001), and we obtained consistent patterns of cumulative HCC risk in all four models whether using the entire, training or validation cohort. Conclusions: Four risk models built with different combinations of common risk factors could distinguish individuals who will or will not develop HCC after an average follow-up of 15.5 years. Part (III): Assessment of model performance and generalizability of HCC-risk scores Specific Aims: To assess the performance and transportability of HCC-risk scores derived from different risk models. Materials Methods: Akaike's information criterion (AIC) and Bayesian information criterion (BIC) were used to evaluate model-based global fit. Subgroup analysis was conducted to explore whether discriminative abilities of different models varied according to distinct characteristics including sources of cohort, gender, age, and hepatitis virus infection status. Sensitivity analysis was used to investigate the impact of HBsAg-positive rate from 10% to 20% on the discriminatory accuracy of the risk scores for predicting HCC. At each level of HBsAg-positive rate, 100 random samples were generated and the distribution of c-statistics was determined. Besides using c-statistics, model performance was also compared based on reclassification of risk categories, including net reclassification improvement (NRI) and relative integrated discrimination improvement (rIDI) index. Results: Using risk scores from model 1-4, c-statistics for 5- and 10-year risk prediction ranged between 0.78-0.86. Furthermore, these c-statistics remained high in each of three individual cohorts (0.74-0.86) and in diverse subgroups stratified by gender (0.75-0.92), age (0.65-0.85) and hepatitis virus infection status (0.65-0.82). Sensitivity analysis revealed change in HBsAg-positive rates had little impact on predictive accuracy, with interquartile range of c-statistics for predicting 10-year HCC risk were 0.74-0.78 for model 1, 0.76-0.80 for model 2, 0.84-0.88 for model 3, and 0.87-0.90 for model 4. Among the four models, model 2 had the lowest AIC and BIC. Compared with model 1, adding history of CLD, first-degree family history of HCC and pack-years of smoking in model 2 resulted in significant improvement in NRI (30.8% for 5 years and 30.9% for 10 years) and rIDI (30.1% for 5 years and 17.0% for 10 years) for predicting HCC events. Conclusions: Model 2 was the best-fit model. In addition, the performance of model 2-derived risk score measured in terms of c-statistics and reclassification indices is equally excellent or superior to other three models overall or in diverse subgroups. Part (IV): Evaluation of priority criteria for selecting individuals for screening of HCC Specific Aim: To evaluate the clinical utility of HCC risk prediction models and compare with current guideline. Materials Methods: Participants were stratified according to tertiles of risk scores. Screening load and number needed to test (NNT) were calculated to explore the screening efficiency. Clinical utility analysis was used to determine net benefit and relative utility across risk thresholds defined by predictive 10-year risk for HCC. To develop screening criteria, participants were stratified according to tertiles of model 2-derived risk scores in combination with hepatitis virus infection status. Sensitivity and specificity were calculated under different screening criteria for the occurrence of HCC within 5-year and 10-year follow-up. Results: Screening efficiency increased with increasing risk tertile, with the lowest NNT (16) for the highest tertile and highest NNT (257) for the lowest tertile of model 2-derived risk score. At a threshold of 2% risk, model 2-derived risk score had greater net benefit and relative utility in comparison with the AASLD (American Association on Study of Liver Disease) criteria for selecting screening candidates. According to 2% risk threshold, our selection criteria, determined on the basis of tertiles of age-specific risk score and status of hepatitis B and/or C virus infection, had improved sensitivity when compared with AASLD criteria (5-year follow-up: 90.2% vs. 76.2%; 10-year follow-up: 89.4% vs. 76.8%), particularly for young-onset HCC under age 50 (5-year follow-up: 76.9% vs. 48.7%; 10-year follow-up: 79.4% vs. 40.6%), with slight loss of specificity (for all HCC events within 5-year follow-up: 67.3% vs. 71.4%, 10-year follow-up: 67.8% vs. 71.8%; for young-onset HCC events within 5-year follow-up: 66.9% vs. 71.0%, 10-year follow-up: 67.0% vs. 71.0%). Conclusions: Model 2-derived risk score combined with hepatitis virus infection had improved sensitivity, as compared with AASLD criteria.