元宇宙數位雙胞胎模型評估免疫糞便潛血精準大腸癌篩檢

Abstract

背景： 糞便潛血免疫法 (Fecal immunochemical test, FIT)篩檢效益評估已經從族群層級的實證醫學證據轉向個人層級為中心的精準醫學。除了過去對影響評估效度的自我選擇偏差與干擾因子的校正之外，評估N對1試驗及考慮個人因子的大腸腺腫-大腸癌自然病史予以不同篩檢策略建議包括個人化篩檢間隔與個人風險分層篩檢建議評估及長程追蹤效益等挑戰亦興起新穎方法學考量。傳統上使用電腦模擬的評估模型無法適用此類複雜情境，其原因包括需要多重虛擬假設，且未能考慮真實世界數據 (Real World Data, RWD)的屬性和特徵。 本論文旨在發展元宇宙數位雙胞胎模型，以評估以族群基礎之FIT篩檢效益，評估不同篩檢間隔等均一性篩檢與個人化風險為基礎篩檢建議之沉浸式篩檢政策。 方法： 元宇宙基礎架構乃透過數位雙胞胎設計創建RWD虛擬分身，經過RWD和虛擬世界數據 (Virtual World Data, VWD)橋接合成，進行隨機化沉浸式介入評估。VWD透過生成式人工智慧引擎利用兩種機器學習演算法生成。首先，利用馬可夫過程從2004年至2022年臺灣全國FIT篩檢數據中學習主宰不同部位大腸直腸癌 (近端和遠側部位)自然病史參數，並評估對不同部位大腸直腸癌其不同篩檢間隔之篩檢效益之差別。其次，我們利用隨機森林機器學習模型辨別社區民眾個人特徵於大腸直腸腺瘤和癌症之個人化風險分位數，並嵌合於第一階段馬可夫過程模型，以評估個人化介入與篩檢政策之效益。 結果： 三階段馬可夫過程結果顯示近端大腸直腸癌從臨床前可偵測期 (Preclinical Detectable Phase, PCDP)到臨床期(Clinical Phase, CP)進展速度較遠側大腸直腸癌快。若將癌症分為早期及晚期之五階段馬可夫過程結果亦也顯示相同結果，早期大腸癌在近端部位的FIT敏感度低於遠側部位，從腺腫到癌症近端部位轉移較快且敏感度低於遠側部份。若欲達到降低大腸直腸癌晚期發生率達25%的效益，則對有較高近端大腸直腸癌風險個案需要每年篩檢一次，而對遠側大腸直腸癌潛在風險較高的個案而言，篩檢間隔可設定於每兩年篩檢一次。 利用隨機森林機器學習演算法找到可鑑別近端和遠側大腸直腸癌之個人特徵，並利用貝氏網路演算產生多階段個人化十分位風險評估模式，予以建立個人化篩檢策略，如高風險者(風險分數大於75%)：每年篩檢一次，中高風險者(風險分數50-75%)：每兩年篩檢一次，低風險者(風險分數25%-50%)：每三年篩檢一次，最低風險者(風險分數0%-25%)：每六年篩檢一次，此個人化篩檢策略於降低晚期大腸癌發生率可達32%。 結論： 本論文顯示應用元宇宙數位雙胞胎模型可做為以病人為中心精準大腸直腸癌FIT篩檢策略設計，並用於評估篩檢效益。

Background: Evaluation of population-based Fecal immunochemical test (FIT) service screening has turned from universal evidence-based medicine to patient-centered personalized medicine. Therefore, a series of concerns and recommendations are raised and suggested including those threats to validity (self-selection and confounding factors) associated with the effectiveness of screening, N to 1 trial impasse, the individually-tailored disease natural course of colorectal neoplasia, various screening policies such inter-screening interval and risk-based screening strategy, and the logistics of long-term follow-up. The conventional simulation models cannot be accommodated to satisfy these complex situations because the use of this traditional approach requires numerous virtual assumptions and fails to take into account the real world data (RWD) property and inherent characteristics. This thesis is therefore to develop a metaverse-based digital twin model for evaluating the effectiveness of population-based FIT service screening with a series of immersive screening policies with emphasis on inter-screening interval and risk-based screening strategies. Methods: The metaverse-based infrastructure was framed by adopting the digital twin design to spin out the avatar of the RWD, after synthetic bridging between the RWD and the virtual world data (VWD), which was further randomized to immersive interventions for evaluation. The VWD was generated by generative artificial intelligence engine by leveraging two machine learning algorithms. The first was to learn the parameters governing the site-(proximal and distal location)-based disease natural course of colorectal cancer with Markov process from the RWD of Taiwan nationwide FIT screening data between 2004 and 2022. The effectiveness of inter-screening interval by site was then evaluated. To develop the personalized disease natural course of colorectal adenoma and carcinoma, we leveraged the random forest of machine learning model to identify important variables based on community-based integrated screening data. The decile of personalized risk assessment model of colorectal adenoma and carcinoma was therefore constructed. The effectiveness of individually-tailored intervention and risk-based screening policy was evaluated. Result: The results of transition rates and sensitivity of FIT test using three-state Markov process show the proximal CRCs progressed faster from pre-clinical detectable phase (PCDP) to clinical phase than the distal CRCs. The findings of five-state Markov process also show the proximal CRCs had faster transition from early to late AJCC stage during PCDP and further progressed to the late stage of clinical phase. The test sensitivity of early AJCC stage was lower in the proximal site than the distal site. The similar findings on faster transition rates from adenoma to carcinoma and lower sensitivity were also noted. If effectiveness of CRC mortality reduction is set at least 25% annual screening is required for those who are potential of developing proximal CRCs and biennial screening is sufficient for those who are potential of developing distal CRCs. Individual features were extracted to distinguish the proximal CRCs form the distinguished CRCs by leveraging the random forest of machine learning algorithm. Risk-based-screening strategy (annual for 75% or higher risk, biennial for 50-75% moderate high risk four-yearly regime for 25%-50% lower risk and six-yearly regime for the lowest risk of 0-25%) in the light of the incorporation of these distinguished features into decile-based personalized risk model was demonstrated to reduce of incidence of advanced-stage CRC 32%. Conclusion: This thesis demonstrates the clinical usefulness of leveraging a metaverse digital twin model for evaluating and designing precision and patient-centered FIT screening strategies for CRC.