肝細胞癌個人化動態預測模型

Personalized Dynamic Prediction Model for Hepatocellular Carcinoma

王威淳(Wei-Chun Wang)

指導教授 ： 陳祈玲

共同指導教授 ： 陳秀熙

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(2017年)

https://doi.org/10.6342/NTU201701025

摘要

背景
在過往的研究中已發展了許多對於肝癌罹病風險的預測模型，但這些模型通常僅適合用於預測目標族群之平均風險。隨著醫療科技的進步，如何將這些預測模型轉而運用於個人化肝癌風險評估與預測已經成為重要的議題。

研究目的
本論文的目的在於
（1）藉由貝氏臨床推理方法建構個人化肝癌罹病風險預測模型，藉以對平
均風險族群與高風險B型肝炎帶原者來進行風險分層;
（2）以時間相依寇斯模型，將HBV DNA指數和肝功能指數(ALT)對於B型
肝炎帶原者以及其他相關指數(AST，ALT，AFP，血糖和血小板)對
於高危險群病患之肝癌罹患風險在納入其動態變化之考慮下建立動
態預測模型;
（3）運用(2)之訊息結合時間相依寇斯廻歸模型所建立之肝癌罹病風險分
數架構以風險分數為基礎之B型肝炎帶原動態肝癌罹病預測模型，並
據以釐清肝癌罹病機轉之動態與中介變化過程以及預測其罹患肝癌
之可能。

材料與方法
本研究使用包含1999至2007年之社區族群肝癌篩檢世代追蹤以及肝癌高風險(B型肝炎帶原者)監測兩個分析資料。社區肝癌篩檢世代追蹤資料採用兩階段方式提供社區民眾肝癌篩檢服務，研究期間共納入98552位民眾資料進行分析，所收集資料包含B型肝炎與C型肝炎感染狀態、肝功能指數、甲型胎兒蛋白指數、肝癌家族病史、人口學變項資料、生活型態資料以及相關之生化檢驗數值。第一階段篩檢為肝癌高風險之民眾則依其風險高低安排接受每三個月或每六個月之腹部超音波肝癌篩檢。參與篩檢民眾後續之罹患肝癌狀態則藉由比對追蹤至2007年之全國癌症登錄資料庫獲得。肝癌高風險(B型肝炎帶原者)為REVEAL-HBV研究監測計劃之世代追蹤資料，共計納入3584位B型肝炎帶原且收案時無罹患肝癌之個案，除提供常規腹部超音波進行肝癌監測以及收集前述之相關資訊外，並檢測其B型肝炎病毒基因型資訊。此監測計劃之肝癌罹病狀態則藉由對疑似個案進行之確診取得。
研究首先運用肝癌風險概似對比值以逐步更新演進之方法在貝氏架構下建構肝癌風險評估系統。為進行個人化肝癌風險分層與風險預測之目的，本研究以一系列之傳統與多階段模式統計方法結合實證資料評估個人化實證肝癌風險。所使用之傳統風險評估模型包含以是否發生肝癌之二元變項為結果進行分析之邏輯斯迴歸分析，並運用廣義線性混和模型將常期追蹤資料中之相關性納入考量。本研究並進一步以發生肝癌之時間為結果運用寇斯等比風險迴歸分析進行肝癌發生風險評估。由於肝癌世代長期追蹤資料包含對於肝功能指數、甲型胎兒蛋白以及血小板記數之重複量測資料，本研究進一步以時間相依之寇斯迴歸模型將預測因子之動態變化納入模型考量，以持續更新之預測因子訊息建構包含時序變化動態之各人化肝癌風險預測，並據以進行肝癌風險分層與動態肝癌防治評估。
為了解肝癌罹病進程之動態結構變化，本研究並建構四階段馬可夫模型，分別以B型肝炎DNA指數與肝癌風險分數定義不同程度之肝癌疾病進程狀態，以評估其病程進展機轉。基於上述四階段模型，本研究以貝氏馬可夫鏈蒙地卡羅方法進行各肝癌罹病階段之轉移以及發生肝癌速率估計，達到建構不同程度之肝癌罹病風險動態變化之目的並據以評估肝癌罹病進程在B型肝炎DNA指數與風險層級間變化之危險層級。

結果
本研究在羅吉斯迴歸分析、寇斯迴歸模型及時間相依寇斯迴歸模型均為一致。利用持續更新之預測因子訊息建構包含時序變化動態之各人化肝癌風險預測，於問診及診驗後，區分具不同個人風險的族群，本論文的案例中，其罹患肝癌風險可能被更新至0.10%至80%之間。
利用高危險的B型肝癌帶原世代，REVEAL-HBV世代，可知收案時的HBV-DNA的增加與肝癌罹癌風險增加有關，其調整後的風險比介於1.12至5.63之間，ALT>=45 IU/L相較於<15 IU/L的調整後的風險比為1.84 (95% CI: 1.21-2.78)，若考慮隨時間變化的HBV DNA及ALT，其調整後風險比則會增加至HBV DNA的1.79-5.99及ALT的2.46。本論文以時間相依寇斯迴歸模型估計結果建構肝癌風險分數。
利用四階段馬可夫模型，我們建構了包含3個以風險分數定義的轉移狀態及罹患肝癌的吸收狀態，結果發現由低至中危險群及中至高危險群的前進轉移速率分別為4% (95% CI: 4.1-4.8%)及3% (95% CI: 2.6-3.3%)，而由中至低危險群及高至中危險群的回復速率則分別為8% (95% CI: 7.8-9.1%)及13% (95% CI: 11-14%)，高危險群的罹癌風險速率為3.2%，相對於中危險群約為6倍，後者約為低危險群的5倍。在考量轉移狀態間的動態變化之後，低、中及高危險群之12年肝癌累積危險性分別為每千人25、69及205例。

結論
本研究利用臺灣地區世代發展一肝癌個人化動態預測模式，此模式不僅可用於族群危險分層，並可做為B型肝炎患者的個人化預防和治療策略之用。

關鍵字

貝氏 ； 個人化預測模型 ； 動態預測模型 ； 肝細胞癌 ； 馬可夫模型 ； 寇斯迴歸模型 ； 社區研究

並列摘要

Background
The predictive model for the risk of hepatocellular carcinoma (HCC) has been developed in previous studies but such kinds of models are often presented for predicting the mean risk of the underlying population. More importantly, the translation of these predictive models into a personalized prediction model for individual risk of hepatocellular carcinoma (HCC) has increasingly gained attention.
The objectives of this thesis are to (1) to build up a predictive model for individual risk prediction for HCC by using a Bayesian clinical reasoning algorithm in order to stratify risk groups for average-risk subjects and high-risk hepatitis B carrier; (2) to build up a dynamic prediction model, considering the dynamics of HBV DNA level and ALT level for hepatitis B carrier and time-varying covariates (including AST, ALT, AFP, AC sugar, and platelet) for average-risk subjects, for the risk of HCC with time-dependent Cox regression model; (3) to build up a dynamic risk-score-based prediction model with the formulation of risk score based on the same information on time-varying covariates for hepatitis B carrier as seen in (2) in order to elucidate the dynamics of intermediate events defined by risk-score-based categories and also to predict the final outcome of HCC.

Materials and Methods
Two study cohorts were enrolled including a community-based screening cohort for general population between 1999 and 2007 and a hospital-based high-risk (i.e. hepatitis B carrier) cohort. For the community cohort, a two-stage design for liver cancer screening were provided for 98552 subjects. Information on HBV and HCV infection status, liver function test, AFT, family history of liver cancer, demographic characteristics, life style variables and relevant biomarkers were collected. Subjects detected as high risk received abdominal ultrasonography for detecting HCC at the intervals of three and six months, depending on the level of risk. The occurrence of HCC were ascertain by the linkage of the nationwide cancer registry till the end of 2007. Considering the REVEAL-HBV hospital-based high-risk population, 3584 subjects who were HBV carriers and free of HCC were enrolled and received regular surveillance of HCC. In addition to the information mentioned above, HBV genotype were also measured. Confirmatory diagnosis of HCC were provided to subject with clinical suspicion.
For the derivation of individual-tailored risk stratification and prediction, a series of statistical approaches including the conventional models and multistate models were applied. Due to the updated information derived from repeated evaluation of biomarkers such as AST, ALT, and platelet count, a time-varying Cox regression model was applied to derive the risk scores used for the following Markov model analysis.
For the derivation of dynamic process along the evolution of HCC, four-state Markov models using HBV DNA vial loading and the risk scores derived from the results of time-dependent Cox model were regarded as the definition of state space.

Results
The findings on the identification of risk factors were consistently noted in logistic regression, Cox proportional hazards regression, and time-dependent Cox regression models. Using Bayesian clinical reasoning algorithm, the posterior individual risk of HCC could be updated to a range between 0.10% and 80%.
In the high-risk population of HBV carrier, REVEL cohort, the adjusted hazard ratio (aHR) of baseline HBV DNA in the levels of 300-9999, 10^4-99999, 10^5-999999, and >= 10^6 increased from 1.12 (95% CI: 0.62-2.03) to 5.63 (95% CI: 3.13-10.13) compared to those <300 copies/mL. The aHR for ALT >= 45 IU/L was 1.84 (95% CI: 1.21-2.78) than ALT<45 IU/L. These figure were inflated when applying the dynamic value in the repeated examination [1.79 (95% CI: 1.06-3.03) to 5.99 (95% CI: 3.58-10.01) for HBV DNA, and 2.46 (95% CI: 1. 59-3.82) for ALT >= 45 IU/L]. A risk score based on the multivariable time-dependent Cox model was derived.
In the four-state Markov model, the progression rates from low- to intermediate- and from intermediate- to high-risk group were 4.4% (95% CI: 4-4.8%) and 3% (95% CI: 2.6-3.3%), respectively. The regression rates from intermediate- to low- and from high- to intermediate-risk group were 8% (95% CI: 7.8-9.1%) and 13% (95% CI: 11-14%). The hazards rate of HCC from the high-risk group was 3.2%, which was 6-fold than the intermediate risk group. The hazard rate of HCC for the intermediate risk group was about 5-fold than the low-risk group. The 12-year cumulative risk of HCC for risk score <=10, 11-14, and >=15 was 25, 69, and 205 per 1000, respectively, making allowance for three transients states pertaining to dynamics of risk-score group (including time-invariant and time-varying covariates).

Conclusions
We developed a novel personalized dynamic predictive model for the risk for HCC among Taiwanese subjects. The proposed dynamic prediction models are not only useful for the risk classification of HCC and also useful for the surveillance of personalized treatment to HBV.

並列關鍵字

Bayesian ； Personalized Prediction Model ； Dynamic Prediction Model ； Hepatocellular Carcinoma ； Markov model ； Cox regression model ； Community-based study

參考文獻

Angulo P. (2002)

Nonalcoholic fatty liver disease. N Engl J Med; 346:1221-31.

Universal hepatitis B vaccination in Taiwan and the incidence of hepatocellular carcinoma in children. Taiwan Childhood Hepatoma Study Group. N Engl J Med. ;336:1855–9.

Ultrasound screening and risk factors for death from hepatocellular carcinoma in a high risk group in Taiwan. Int J Cancer.;98:257–61.

Risk of hepatocellular carcinoma across a biological gradient of serum hepatitis B virus DNA level. JAMA; 295: 65–73