- 學位論文
推理模式應用於心血管疾病精準預防
Reasoning-based Models for Precision Prevention of Cardiovascular Disease
摘要
背景 為達到世界衛生組織 ”2025年前降低非預期死亡達25%” 之目標，目前全球對於心血管疾病防治有三大主要防治策略：以平均風險為主之族群介入(Population-Wide Approach，PWA)、高風險族群介入(High-Risk Approach，HRA)，以及新興發展之精準預防介入(Precision Preventive Approach，PPA)。如何在考慮疾病前期階段之動態心血管疾病進展中納入個人層次多樣之風險因子以達到精確預測個人心血管疾病風險之目標，改善過往預測工具多僅於族群平均風險層面之缺點是當前達到精準心血管疾病預防目標之關鍵。運用多階段統計模型結合大數據分析發展包含疾病前期進展之動態個人化心血管疾病風險評估模式，有助於精準心血管防治之發展。 材料與方法 本論文使用以族群為基礎之世代追蹤研究設計，運用各項癌症及慢性病重複測量及長期追蹤之整合式社區篩檢資料進行分析。在大數據分析與機器學習方面，則運用重新抽樣為基礎之方法及研究設計，訓練資料集進行估計，並利用測試集資料進行交叉驗證。運用圖形因子設計結合動態貝氏網絡分析以利於機器學習演算法之發展。 本研究以族群為基礎的世代追蹤研究進行流行病學分析。以迴歸為基礎的隨機過程分析、動態貝氏網絡分析運用於整合式篩檢世代追蹤資料。使用資料包含自1999年參與基隆社區的整合式篩檢計畫(Keelung Community-based Integrated Screening，KCIS)以及2005至2018年間參與彰化整合式篩檢計畫(Changhua Community-based Integrated Screening，CHCIS)之民眾。此社區整合式篩檢計畫包含多維度的測量生物醫學與人體相關指標測量，以及含蓋生活習慣之資料之收集。此社區長期追蹤資料亦收集包含高血壓、代謝症候群、糖尿病以及大腸直腸癌之主要結果。 本論文應用五階段馬可夫模型評估代謝症候群分類(Refined MetS-Related Classification，RMRC)以及兩個主要結果(心血管疾病與心血管疾病相關死亡)之疾病自然病程，並同時考慮其他競爭死因之影響。繼而以四階段馬可夫模評估高血壓前期及高血壓的疾病自然進程。 在機器學習運用方面，本研究以監督及非監督式演算法發展以人工智慧為基礎之心血管疾病以及疾病前期(Intermediate)狀態之新分類。在此基礎上，本研究運用隱馬可夫模型(Hidden Markov Model，HMM)之動態貝氏網絡分析發展結合生物路徑與新人工智慧分類之前期疾病(Intermediate state)進展，釐清此新分類之中間狀態的生物機轉。 結果 (1)對於心血管新興風險因子之探討 隨著f-Hb濃度漸升，調整其他相關因素之心血疾病疾病風險亦呈劑量濃度效應之方式漸增，而對於心血管疾病死亡也顯現出相同的趨勢。 在f-Hb與CRP之相關分析結果方面，調整年齡、性別、代謝症候群、運動、喝酒及抽菸後，f-Hb 濃度在50-99 ng/ml，100-499 ng/ml，以及500 ng/ml以上之CPR陽性危險勝算比則介於1.8-2.5倍之間。 (2)以迴歸為基礎的隨機過程模型(Regression-based Stochastic，RBS) 在考慮不同嚴重程度之代謝疾病隨機過程模型中，約30%之輕度代謝疾病狀態個案會改善成為無疾病狀態。將此一疾病自然改善納入考量後，輕度、中度，以嚴重代謝疾病狀態進展至心血管疾病之年風險分別為1.6%、4.7%，以及20.2%。男性進展成為代謝症候群之風險較高，而女性發生心血管疾病之風險則較高。女性以及年輕族群由輕度代謝疾病改善成為無病狀態之可能性較高。 運用四階段馬可夫模型藉由高血壓前期與高血壓間之淨返回速率(由四階段模型中以返回速率扣除進展速率得到)評估發展成為第2期高血壓之10年風險結果顯示，社區觀察資料相較於無介入之血壓自然進展在風險分層10%、50%，以及100%之族群分別下降達9%、42%，與77%。社區觀察資料與無介入之血壓自然進展相較，在高血前期呈現較高之比例；而對於高血壓則呈現較低之比例。此一結果顯示在社區中經過多年逐漸推行之生活型態改變下高血壓風險逐漸下降，而由於健康意識之提升，高血壓前期個案之偵測則漸增。 (3)利用有向無環圖(DAG)模型構建代謝症候群因果圖 動態貝式分析之建構主要應用於對於代謝症候群及心血管疾病之個人化風險預測。以條件為有抽菸、嚼檳習慣、教育程度為小學，以及尿酸、天門冬胺酸轉胺酶(GOT)、丙胺酸轉胺酶(GPT)、血液尿素氮（BUN）、肌酸酐異常之80歲男性，若目前狀態無罹患任一代謝症候群(FMD)，而10年發展為代謝症候群、發生心血管疾病及死於心血管疾病之機率分別為2.3%、76%及21%。而有抽菸、飲酒習慣、教育程度為高中、且有尿酸異常之68歲男性屬中度風險，未來10年發展至代謝症候群、心血管疾病及死於心血管疾病之風險為7.7%、48%及4%。而有規律運動、學歷為大學，且各項生化指標為正常之55歲女性屬低風險族群，相對應的風險值分別為6.7%、39% 及0 .23%. (4)隱馬可夫模型(Hidden Markov Model，HMM)之動態貝氏網絡分析結果 利用高斯隱馬可夫模式以代謝分數(標準化代謝因子總和)服從常態分佈下，比較兩個至七個不同隱狀態個數之模型估計結果，顯示五個隱狀態模型表現較其他隱狀態個數好。初始狀態對應五類狀態的比例分別為30%、15%、31%、19%及5%，而五個狀態對應的代謝分數則介於-4.37至4.60之間。Viterbi演算法對於可對於每個人的隱狀態路徑進行解碼。並由發生心血管疾病之存活曲線看出這五個族群的分類是具有辨別性的。 結論 心血管疾病預防以由傳統一般風險族群介入策略與高風險族群介入進展至目前的個人化精準預防策略。結合複雜之統計模型與大數據分析方法建構涵蓋生物機轉之動態風險預測模式並且奠基於此涵蓋疾病前期與心血管疾病以達到包含初段、次段以及末段之個人化精準預防策略至關重要。
並列摘要
Background Prevention of cardiovascular disease (CVD) to reach the goal of the WHO 25 by 25 has oriented population-wide approach (PWA) and high-risk approach (HRA) into precision preventive approach (PPA). To this end, a precise and individualized risk prediction model for CVD is required to accommodate a large number of features accounting for occurrence of CVD and model the detailed intermediate states of CVD on individual level in order to improve the weakness of the previously developed prediction models the basis of population-average rather than individual-specific risk. So doing may require a complex multistate statistical model and also big data analytics for developing such a kind of the reasoning-based model that is translated into the precision prevention of CVD. Materials and Methods A population-based cohort design with the integration of the community-based screening data from various cancers and chronic diseases with repeated measurement and longitudinal follow-up was used for analysis. The population-based cohort was used for epidemiological analyses, regression-based stochastic analysis, dynamic Bayesian DAG model, and Bayesian network model. A five-state Markov process was used to describe the dynamics of natural course on the refined MetS-related classification (RMRC) associated with the risk for two prognostic outcomes, CVD and its death. The similar four-state Markov model was also applied to modelling the natural course of prehypertension and hypertension. Individually-tailored risk prediction model for CVD with intermediate events of MetS was developed with the application of Bayesian causal DAG model. The ML algorithms of supervise and unsupervised learning are applied to derived the new AI-based classification for the outcomes of intermediate state and CVD. The dynamic Bayesian network analysis with the hidden Markov model was used to annotate the new AI-based classification with the biological pathways of intermediate state of CVD. Results (1)Identification of Novel Risk Factors Associated with CVD When the f-Hb concentrations increased, the adjusted hazard ratios for CVD increased in a dose-response manner. The similar pattern was observed for deaths from cardiovascular disease. Regarding the association between f-Hb and CRP, after adjusting for sex, age, MetS, exercise habit, drinking and smoking status, we found that the risks of having elevated CRP for f-Hb were 1.8 to 2.5-fold for the concentration of 50-99, 100-499 and 500+ ng/ml. (2)Results of Regression-based Stochastic (RBS) Models Approximately 30% mild state subjects were potential of being reversible to healthy state and the risk for CVD increased with the severity of MetS being from less than 1.6%, 4.7%, and 20.2% annually. Regression from mild to the state free of disease was more likely to observe in female and young age group. Applying the estimated results of the four-state Markov model for pre-hypertension and hypertension to the net score for calculating the probability of progressing to stage 2 hypertension after 10 years following the disease natural history for those with net score at first decile was as low as 9%, compared with the value of 42% and 77% for those at fifth and tenth decile. Compared the observed with the predicted percentage derived from the natural history model, the percentage of prehypertension was higher in the expected compared to the observed percentage whereas hypertension was lower in the expected percentage compared to the observed. These findings suggest that hypertension has been reduced due to the gradual introduction of life style modification and pre-hypertension has been increased possibly due to the enhanced awareness. (3) Bayesian DAG model for building up the causal diagram of MetS Based on the developed dynamic Bayesian Analysis, the individual risk of MetS and CVD can be predicted. An 80 years old male with habits of smoking and betel quids chewing, elementary school educated, and elevated UA, GOT, GPT, BUN, and Cr would develop MetS and CVD and CVD death in 10-years with probability of 2.3%, 76% and 21% given FMD at baseline. An intermediate risk subject (68 years old male with habits of alcohol drinking and smoking, and senior high school graduated with elevated uric acid) had the corresponding figures of 7.7%, 48%, and 4%. A low risk female (55 years old with habit of regular exercise, and college graduated, with biochemical measures of interests in this study in the normal range) had the corresponding risk of 6.7%, 39%, and 0.23%. (4) Results of Dynamic Bayesian Network with Hidden Markov Model The estimated results of the Gaussian hidden Markov model with 2 to 7 states following the normal distribution for the metabolic score show that 5-state HMM performed better than others. The initial distribution of the 5 state partitioned the study population into five categories with proportions of 30%, 15%, 31%, 19%, and 5%. The mean value of the metabolic score in the HMM ranged from -4.37 to 4.60. The Viterbi algorithm for the pathway decoding deciphered individual disease course. Conclusion The evolution of CVD prevention has been initiated from population-based approach and high-risk approach toward a personalized prevention approach. Building up reasoning-based models for such a purpose requires a complex multi-state statistical model and also big data analytics are important for developing a precision prevention map of CVD on both population and individual level.