大腸直腸癌風險預測及監測間隔規劃：篩檢方法及監測策略

根據國際癌症研究機構指出，大腸直腸癌是全球常見癌症及癌症死因，且近年發病率及死亡率皆呈現增長的趨勢，使得大腸直腸癌於未來醫學研究領域上極具重要性。本研究旨在解決兩項大腸直腸癌的分析情境：篩檢方法與監測策略。篩檢方法為利用重要風險因子來建構大腸直腸癌風險分類模型以預測受測者的進階性大腸腫瘤風險分數。監測策略為考量時間因素來建構回診時間之監測間隔推薦模型以替受測者的個人化需求提供合適的回診時程規劃。本研究使用國立台灣大學醫學院附設醫院的健康管理中心所提供的長期追蹤之健康檢查資料為研究數據，並以已完成大腸鏡檢查且有確定記錄篩檢結果的資料為最終樣本進行分析。篩檢方法採用廣義估計方程式作為預測模型，並使用減少多數抽樣法來減輕不平衡資料影響，同時進一步嘗試原始模型參數、參數平均法、參數投票法及參數奇異值分解法等四種建模方法來強化模型的準確性。監測策略採用階層式模型作為預測模型，並納入重要風險因子來增進模型擬合之配適度。篩檢方法採用ROC曲線下面積、敏感度及特異度進行模型評估，而監測策略採用最小化資訊量準則為之。篩檢方法表明結合年齡、性別、家族大腸直腸癌病史及現在平均每日抽煙量等重要風險因子與過去大腸鏡檢查資訊之大腸直腸癌風險分類模型能有較高的模型準確性，並使用參數平均法能強化模型準確性。監測策略以追蹤時間長度作為時間尺度，來建構可動態改變且相對簡單穩定的一般線性監測間隔推薦模型，並依照欲越過限制的機率幅度來計劃適當的監測間隔，且納入重要風險因子後能進一步優化模型擬合程度。

關鍵字

大腸直腸癌；進階性大腸腫瘤；篩檢方法；監測策略；風險預測；監測間隔；廣義估計方程式；階層式模型

並列摘要

According to the International Agency for Research on Cancer, colorectal cancer (CRC) is a common cancer and cancer cause of death globally, and the morbidity and mortality rates have shown an increasing trend in recent years, making CRC extremely important in the field of medical research in the future. This study aims to solve two analysis scenarios of CRC: screening and surveillance. The screening is to develop a CRC risk classification model using the important risk factors to predict the subject's risk score for advanced colorectal neoplasia. The surveillance is to take the time factor into consideration to develop a monitoring interval recommendation model of the follow-up visit to provide an appropriate revisit schedule for the subject's personalized needs. The data used for this study was the long-term follow-up health examination data provided by the Health Management Center of National Taiwan University of Medicine Hospital, and the cases that have completed the colonoscopy and had a definitive outcome recorded as the final sample for analysis. The screening adopted the generalized estimating equation as the prediction model and utilized the undersampling method to alleviate the impact of imbalanced data. The prediction model was developed further using four modelling methods including original, average, voting and singular value decomposition of model parameters to enhance the discriminatory power of the model. The surveillance adopted the hierarchical model as the prediction model and included important risk factors to improve the model fitting. Model evaluation was assessed by receiver operating characteristic curve, sensitivity and specificity in the screening and Akaike information criterion in the surveillance. The screening showed that the CRC risk classification model combining age, gender, family history of CRC, current daily average amount of smoking and other important risk factors with past colonoscopy information had higher predictive power. The use of modelling method with average of model parameters could strengthen the discriminatory power. The surveillance took the time since screening as the timescale to construct a not only dynamically changeable but also relatively simple and stable monitoring interval recommendation model using linear growth model for planning the suitable monitoring intervals in terms of limiting the probability of exceeding the specified threshold. The model fitting of the recommendation model containing important risk factors was further improved.

並列關鍵字

Colorectal cancer ； Advanced colorectal neoplasia ； Screening ； Surveillance ； Risk prediction ； Monitoring intervals ； Generalized estimation equation ； Hierarchical model