應用動態貝氏網絡結合馬可夫決策模型於COVID-19防疫造成台灣大腸癌篩檢延遲

Dynamic Bayesian Network Combined with Markov Decision Tree Model for Adapting Colorectal Cancer Screening Disrupted by COVID-19 Pandemic in Taiwan

范僑芯(Chiao-Hsin FAN)

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(2022年)

https://doi.org/10.6342/NTU202203202

摘要

前言
隨著2020年3月爆發的COVID-19全球大流行，如族群大腸癌篩檢一般的公共衛生疾病防治計畫在全球各國皆受到疫情影響而有所延滯。先前的研究顯示，由篩檢服務以及後續治療之延滯將造成大腸癌個案期別惡化以及死亡增加。然而目前仍鮮有研究著墨於此篩檢延滯造成衝擊的應對策略。本研究因此提出疫情對於篩檢及醫療照護衝擊的創新解決策略評估架構與方法。
研究目標
本研究首先發展架構於篩檢過程與大腸癌疾病進展之動態貝氏網絡(dynamic Bayesian Network，DBN)並運用於評估篩檢對於病灶之偵測與可達成之效益。本研究進一步運用所建構之貝氏動態網絡結合馬可夫決策樹提出對於COVID-19疫情期間所造成的篩檢與治療延滯的最佳優先化介入策略對於挽救此延滯之衝擊的效益。
研究方法
本研究之設計架構於貝氏動態網絡為基礎的馬可夫決策樹評估模型。運用貝氏有向圖形(Directed Acyclic Graphic，DAG)模型與D-鑑別(D-separation)節點指向原則以建構篩檢以及大腸鏡確診過程與包含無法直接觀察之正常無大腸腫瘤狀態、大腸腺腫、臨床症前病灶大腸癌之潛藏馬可夫疾病自然進展與可直接觀察之臨床期大腸癌疾病自然進程和與此自然病史相關聯之生物標記與期別分類之精簡連結，並且據以建立涵蓋篩檢過程與大腸癌疾病自然進展之動態貝氏網絡。在所建立之動態貝氏網絡基礎上，本研究首先運用馬可夫決策樹模型評估一年一次、兩年一次，以及三年一次之篩檢間隔延遲對於大腸腺腫、大腸癌篩檢間隔個案與大腸癌死亡之影響。本研究進一步在兩年一次糞便潛血檢查篩檢策略之情境下提出以改變免疫法糞便潛血濃度切點之優先介入挽救策略在面對COVID-19疫情造成篩檢延滯時得以挽救滅少的大腸腺腫、超額篩檢間隔個案與大腸癌死亡。
實證資料
本研究運用包含一百萬50-74歲台灣族群免疫法糞便潛血檢查大腸癌篩檢長期追蹤實證資料於訓練動態貝氏網絡架構之馬可夫決策樹模型參數。此台灣族群實證資料之訊息包含篩檢年齡、糞便免疫潛血檢測結果、是否接受大腸鏡轉介確診、是否罹患大腸癌以及其大腸癌偵測模式(包含盛行篩檢偵測個案、後續篩檢偵測個案、篩檢間隔個案、拒絕篩檢個案)、大腸癌期別，追蹤時間，以及是否死於大腸癌。
結果
基於動態貝氏網絡學習建立之大腸癌疾病自然進展參數結合台灣100萬篩檢族群篩檢情境之馬可夫決策樹模型於80%篩檢參與率之評估結果顯示，在延遲篩檢1年情境下，對於每年一次的篩檢策略將造成減少大腸腺腫之偵測1187例、增加篩檢間隔大腸癌488例、增加大腸癌死亡429例。此衝擊對於三年一次之篩檢策略分別為1579(減少大腸腺腫之偵測), 1292(增加篩檢間隔大腸癌), 以及531(增加大腸癌死亡)。對於現行台灣兩年一次的篩檢策略，延遲篩檢6個月即造成減少大腸腺腫之偵測717例、增加篩檢間隔大腸癌481例、增加大腸癌死亡243例，此衝擊在延遲篩檢1年分別為1435例(減少大腸腺腫偵測), 962例(增加篩檢間隔大腸癌),與485例(增加大腸癌死亡)。當篩檢延遲達1.5年，則此衝擊分別增加達1584例(減少大腸腺腫偵測), 978例(增加篩檢間隔大腸癌),以及609例(增加大腸癌死亡)。
運用切點改變挽救措施，在提高免疫法糞便潛血檢測數值為150ng/ml下，對於篩檢延遲造成無法偵側腺腫個案將可減少90%，若切點提高為450 ng/mL，則可減少49%。提高切點值對於篩檢延遲造成的額外篩檢間隔個案將可分別減少91%(150ng/ml)與57% (450ng/ml)。提高切點值對於篩檢延遲造成的大腸癌死亡亦可分別減少88% (150ng/ml)與43% (450ng/ml)。
結論
本研究發展動態貝氏網絡馬可夫決策樹模型並且運用於評估由COVID-19疫情造成常規篩檢延遲對長期效益影響以及是否可藉由改變免疫法糞便潛血檢測陽性切點之優先介入策略挽救篩檢延遲之衝擊。在此評估架構下本研究運用台灣大腸癌族群篩檢實證資料進行實證評估以提出篩檢延遲之介入策略。

關鍵字

動態貝式網絡分析 ； 馬可夫決策模型 ； 新冠肺炎 ； 大腸癌篩檢

並列摘要

Introduction
In parallel with the evolution of COVID-19 pandemic as of March 2020, public health prevention programs, population-based colorectal cancer screening for example, have been disrupted by a series of large-scale outbreaks around the world. The previous study has already shown the impacts of the delayed screening schedule and the delayed treatment on an increase in advanced colorectal cancer (CRC) and its death. Nonetheless, up to date the solution to salvage these impacts has been barely addressed. It is therefore imperative to envisage a new approach to ameliorating the percussion of the delayed schedules for screening and the subsequent medical regimes.
Objective
The aim of this thesis is first to develop a dynamic Bayesian network (DBN) that relates the screening process to the yields and the effectiveness of screening amenable to the underlying multistate disease natural history of colorectal neoplasm. The proposed DBN was further combined with Markov decision tree model to assess how the envisaged salvaged decision strategy can minimize the impacts of the procrastinated screening schedule affected by COVID-19 pandemic.
Methods
The study design was pursuant to the DBN-enshrined Markov decision tree model. The DBN was developed by using the directed acyclic graphic (DAG) approach with D-separation to parsimoniously link the screening process and the confirmatory diagnosis with the underlying disease natural history of colorectal neoplasm including the hidden states from free of neoplasm, adenoma, and pre-clinical detectable phase (PCDP) and the observed state of surfacing to clinical phase (CP) of CRC classified by relevant biomarkers and tumour staging. The proposed DBN was incorporated into the Markov decision tree model to assess the extent of the increased interval cancer and deaths from CRC because of the procrastinated screening schedule by annual, biennial, and triennial screening regimes. The salvaged decision strategy based on the change of the cutoff of fecal immunological test (FIT) was further proposed to reduce under-detected adenoma, interval cancers and deaths from CRC because of the disruptive effect of COVID-19 pandemic in comparison with no-salvage under the existing biennial FIT screening regime identical to Taiwan scenario.
Data
The empirical data used for training the parameters used in the DBN-enshrined Markov decision tree model were derived from a total of one million residents of Taiwan population eligible for biennial FIT screening aged between 50-74 years. The main context of data consisted of the age of screening, the positive or negative result, whether to referral, age and time to have various kinds of detection modes (including first screen, subsequent screen, interval cancer, and refuser), tumour staging, and deaths from colorectal cancer.
Results
Given the estimated transition parameters on the disease natural history and 80% attendance rate of one million residents, under-detected colorectal adenoma, excess interval cancers, and excess CRC deaths, respectively, for 1-year delay increased from 1187, 488, and 429 for annual regime to 1579, 1292, and 531 for triennial regime. Given the current biennial regime, under-detected adenoma, excess interval cancers, and excess CRC deaths were 717, 481, and 243 for 0.5-year delay, 1435, 962, and 485 for 1-year delay, and 1584, 978, and 609 for 1.5 year-delay.
The under-detected adenoma would be reduced by 90% to 49% when the cutoff was changed from 150ng/mL to 450ng/mL. The excess interval cancers would be reduced by 91% to 57% when the cutoff was changed from 150ng/mL to 450ng/mL. Likewise, the excess CRC deaths would be averted by 88% to 43% when the same cut-offs were changed.
Conclusions
The DBN-enshrined Markov decision tree model was proposed to assess the impacts of the procrastinated screening schedule on long-term effectiveness resulting from regular screening and also to evaluate the possible reduced harm with the adaptation of the cutoff of f-Hb concentration as a result of FIT. These assessments were demonstrated by using Taiwan nationwide colorectal cancer screening program.

並列關鍵字

Dynamic Bayesian Network ； Markov Decision Tree Model ； COVID-19 ； Colorectal Cancer Screening

參考文獻

1. ABURTO, José Manuel, et al. Quantifying impacts of the COVID-19 pandemic through life-expectancy losses: a population-level study of 29 countries. International journal of epidemiology, 2022, 51.1: 63-74.

2. BLAY, J. Y., et al. Delayed care for patients with newly diagnosed cancer due to COVID-19 and estimated impact on cancer mortality in France. ESMO open, 2021, 6.3: 100134.

3. Bowman MA, Katzoff JM, Garrison LP, Wills J. Estimates of physician requirements for 1990 for the specialties of neurology, anesthesiology, nuclear medicine, pathology, physical medicine and rehabilitation, and radiology: a further application of the GMENAC methodology. JAMA. 1983 Nov 18;250(19):2623-7.

4. Burke BT, Miller BF, Proser M, Petterson SM, Bazemore AW, Goplerud E, Phillips RL. A needs-based method for estimating the behavioral health staff needs of community health centers. BMC Health Services Research. 2013 Dec;13(1):1-2.

5. BUSKERMOLEN, Maaike, et al. Colorectal cancer screening with faecal immunochemical testing, sigmoidoscopy or colonoscopy: a microsimulation modelling study. Bmj, 2019, 367.