睡眠呼吸中止症候群的預測統計模型

Abstract

背景： 以前阻塞性睡眠呼吸中止症嚴重度預測模型的研究均有使用流 行病學及臨床的參數來估計的睡眠呼吸障礙指數（AHI），然而睡 眠呼吸障礙指數（AHI）相對於阻塞性睡眠呼吸中止嚴重度的統 計特性一直沒被深入探討，而目前似乎也沒有人應用預測模型來 估計如何以介入性計畫來改善某些危險因子而達到降低風險的目 的。 目標： 我們嘗試探討AHI的統計特性，然後利用不同多重廻歸模型來預測每個危險因子對AHI變化的相對貢獻，最後藉由改善顯著危險因子而降低阻塞性睡眠呼吸中止的風險性。 方法: 全部500個個案隨機由原本樣本中2005年至2007間在一家北台灣的醫學中心接受夜間睡眠檢查的2110名病人（1599名男性及511名女性）中隨機選取，其平均AHI為每小時28.32次(標準差26.38)，4個危險因子的資料包括身體質量指標（BMI）、頸圍、抽菸及病歷回溯自述性高血壓，多重線性廻歸模型先被用來區分明顯危險因子，利用AHI的連續變項的特性以對比危險模型及多重分佈羅吉斯廻歸模型進一步用來估計每一個危險因子的臨床廻歸係數，每一個羅吉斯廻歸模型的ROC曲線均被畫出來評估每個預測模型的預測準度，基於預測模型所計算出來的危機值，貝氏定理演算出在改善高血壓控制及BMI後睡眠呼吸中止風險性下降幅度。 結果： AHI分數呈現正偏斜的分佈，傳統的阻塞性睡眠呼吸中止的切點為5、15、30，定義為輕度、中度、重度，相當與AHI分佈的20%、40%及60%，4個危險因子包含年齡、性別、BMI及頸圍等在多重線性模型下被歸類為統計上顯著，在多重分佈羅吉斯廻歸模型下，高血壓是一個針對中度阻塞性睡眠呼吸中止顯著的預測因子，. 藉由細分比較AHI＜5，5-14，15-29及＞30，每一個廻歸係數比重均差不多，而這個結果可以被對比危險模型所支持，利用貝氏定理去推導出來的事後機率（Posterior odds）及概似比（likelihood ratio）由預測模型來導出risk score公式，如將BMI由30改善到標準的23，則中度睡眠呼吸中止的事後機率由21.7下降到6.9。 結論： 阻塞性睡眠呼吸中止的危險預測模型有助於設計介入計畫來改善睡眠呼吸中止的危險因子

Background： The predictive model for obstructive sleep apnea (OSA) measured by apnoea - hypopnoea index using epidemiological and clinical parameters has been proposed in previous studies. However, how the statistical property of AHI corresponds to the severity of OSA has been hardly addressed. Nor the application of predictive model to estimating risk reduction after the administration of intervention program that modulates the modifiable predictors. Objectives： We are tempted to investigate the statistical property of AHI, then applied different multiple regression model to estimating the relative contributions of each predictors to the variation of AHI, and finally to estimate the risk reduction of OSA by modifying the significant predictors. Methods： A total of 500 samples were randomly selected form 2110 patients (1599 males and 511 females) who underwent nocturnal sleeping examination between 2005 and 2007 in one medical centre, northern Taiwan, with the average AHI of 28.32 (SD=26.38). Data on four predictors were collected, including body mass index (BMI), neck circumference, smoking and self-reported hypertension by reviewing medical chart. The multiple linear regression model was first applied to identifying significant predictors. Proportional odds model and multinominal logistic regression model were further adopted to estimating clinical weights (regression coefficients) of each predictor by taking AHI as categorical and ordinal property. Each receiver operating characteristics (ROC) curve corresponding to each logistic regression model was plotted to assess predictive validity of each predictive model. Based on risk score computed from the predictive model, Bayesian approach given different prevalence rate of OSA was applied to calculating the posterior risk in order to project risk reduction due to the modification of hypertension and BMI. Results ： AHI score is a positively skewed distribution. The conventional cutoffs on 5, 15 and 30, defined as mild, moderate, and severe state, were commensurate with 20%, 40%, and 60% of the distribution of AHI. Four predictors were identified statistically significant, age, gender, BMI, and neck circumference in multiple linear regression model. In the multinominal logistic regression model, hypertension was a significant predictor, particularly affecting the moderate OSA. By looking at piecewise comparison of AHI < 5, 5-14, 15-29, and 30+, the contribution of clinical weight to each outcome are similar. These findings are supported by using proportional odds model. Posterior odds using Bayesian approach by prior odds (prevalence odds of OSA) and likelihood ratio based on the density of composite risk score calculated by multiplying the relevant covariates with each clinical weight obtained from the predictive model. Reduction of BMI from 30 to 23 yields the posterior risk of OSA defined by AHI greater than 15 from 21.7 to 6.9. Conclusion： Predictive model for risk prediction for the OSA is helpful for the design of program intervention in modifying the risk factor.