代謝症候群與低血紅素雙向時序及全死因關係研究

Abstract

背景: 過去已有研究報告陳述血紅素濃度與代謝症候群之間的關聯，但兩者間的雙向關係尚未被釐清。此外，低血紅素合併代謝症候群與全死因死亡間的相關性也甚少研究探討。本研究欲運用社區長期追蹤世代研究資料對血紅素濃度與代謝症候群雙向關係及兩因子與全死因死亡相關性進行探究。 方法: 本研究將三個台灣社區共97460名在2000年至2009年參與社區整合式篩檢者（年齡大於20歲以上）納入這個世代研究，我們設計一項前瞻性追蹤世代研究，包括兩個世代追蹤，其一由初始時無貧血的參與者組成，另一個則是由初始時無代謝症候群的參與者組成。在追蹤期間確定了代謝症候群和貧血的事件病例，並採用寇克斯比例風險回歸模型與調整相關干擾因子以評估代謝症候群對貧血的影響，另一方面，利用寇克斯比例風險回歸模型與調整了相關干擾因子來評估貧血對代謝症候群的影響。使用Kaplan-Meier 方法來估計四組的存活曲線。寇克斯比例風險回歸模型推估共變數(年齡、性別、吸菸、飲酒、嚼檳榔及基礎線貧血或代謝症候群狀態) 對存活時間的影響，以及採用監督式機器學習方法中的支援向量機、邏輯式迴歸分析及貝氏網絡三種方法，以產生預測貧血及代謝症候群的發病與否的模型。最後針對貧血或代謝症候群有無追蹤死亡至2012年，利用寇克斯比例風險回歸模型以評估貧血或代謝症候群對全死因死亡影響，亦以支援向量機及邏輯式迴歸機器學習方法建構貧血及代謝症候群對全死因死亡的預測模型。 結果: 與沒有代謝症候群的個案相比，患有代謝症候群的受試者貧血發生的風險降低了 8%。[調整後風險比=0.92 (95%信賴區間0.87-0.97)]。與沒有貧血的參與者相比，患有貧血的參與者發生代謝症候群的風險降低了18%。 [調整後的風險比=0.82 (95%信賴區間0.78-0.86)] 執行統合分析結果發現，貧血與代謝症候群之間存在負相關，於固定效應模型的勝算比為0.79 (95%信賴區間: 0.76-0.83) ，而在隨機效應模型中的勝算比為0.77 (95%信賴區間: 0.34- 1.24)，雖然後者沒有統計顯著。以監督式機器學習方法(支援向量機、邏輯式迴歸分析及貝氏網絡)預測代謝症候群發病及貧血發生的風險，發現貝氏網路模型作預測貧血風險其預測能力C統計值為0.617，於三者中表現最佳，而支援向量機預測代謝症候群風險預測能力達0.639，於三者中表現最佳。追蹤期間無貧血或代謝症候群者的全死因死亡率最低，有貧血或代謝症候群者死亡率次之，死亡率最高的族群是有貧血合併代謝症候群的參與者，勝算比為2.18 (95%信賴區間:1.94- 2.45)。考量年齡、性別、吸菸、飲酒、嚼檳、貧血及代謝症候群因子，利用支援向量機預測能力達0.841。 結論: 這是第一個以族群為基礎的世代研究來評估貧血與代謝症候群的雙向關係及其對全死因死亡影響的研究。此研究結果對這兩種疾病在個人化醫療上具有重要意義。

Background Previous studies have reported the association between hemoglobin (Hb) levels and metabolic syndrome (MetS), but its temporal sequence based on a population-based study has been not elucidated yet. Also, the association of all-cause mortality with Hb and MetS has rarely been addressed. The study was aimed to assess the bidirectional relationship between Hb and MetS and reveal the association between Hb/MetS and all-cause mortality based on a longitudinal community-based study cohort. Methods After enrolling the 97460 screening participants (aged greater than 20 years old) in a cohort of over eight years, a prospective follow-up cohort study was designed by following the two normal cohorts over nine years (during 2000 to 2009), including two normal cohorts, anemia-free and free of metabolic syndrome (MetS) at baseline. The incident cases of metabolic syndrome and anemia was ascertained. Cox proportional hazards regression model was adopted to assess the effect of metabolic syndrome on anemia and vice versa with adjustment for other relevant confounding factors. Based on the baseline, participants were divided into four groups---- no anemia or metabolic syndrome, anemia, metabolic syndrome, and anemia accompanying metabolic syndrome. Kaplan-Meier method was used to estimate the survival curves of these four groups. Cox proportional hazards regression model was applied to estimate the effect of covariates (age, sex, smoking, alcohol consumption, betel quid chewing, and baseline status of anemia or metabolic syndrome) on survival. Three supervised machine learning methods including support vector machine (SVM), logistic regression analysis and Bayesian network were used to generate models for predicting the development of anemia or metabolic syndrome. Deaths were ascertained by follow-up until the end of 2012. Cox proportional hazards regression model was used to assess the effect of Hb and MetS on all-cause mortality. SVM and logistic regression method were applied to construct the all-cause death predictive models. Results Subjects with MetS as opposed free of Mets yielded an 8% decreased risk for incident anemia. [adjusted hazard ratio =0.92 (95% CI 0.87-0.97)] making allowance for other confounding factors. Participants with anemia versus free of anemia yielded an 18% decreased risk for incident metabolic syndrome. [adjusted hazard ratio =0.82 (95% CI 0.78-0.86)] after considering other confounding factors. The negative association between anemia and MetS based on Meta-analysis has been corroborated by the fixed-effect model (OR=0.79, 95% CI: 0.76-0.83) and the random effect model (OR=0.77, 95% CI: 0.34-1.24) although the latter was not statistically significant. Using supervised machine learning methods (SVM, logistic regression analysis and Bayesian network) to predict the risk of metabolic syndrome and anemia, revealed that C statistic (Area Under Curve) prediction of risk of anemia was 0.617 using Bayesian network, which was the best among them. The highest C statistic (0.639) for prediction the risk of anemia using SVM was found. During the follow-up period, those without anemia or MetS at baseline had the lowest all-cause mortality rate followed by those with baseline anemia or metabolic syndrome, and the group with the highest mortality was participants with baseline anemia and metabolic syndrome. Compared with those without anemia or MetS, the adjusted hazard ratio was 2.18 (95%CI: 1.95-2.45) for subjects with anemia and MetS. Conclusion This is the first study to report both causal temporal sequences between anemia and Mets and association of all-cause mortality with anemia and MetS based on a large population-based cohort data. Our findings have a significant implication in personalized medical care for both diseases.