利用盛行率調查資料建立並驗證結核病預測模型—針對愛滋病陰性或未知的族群

Abstract

背景 目前常見的系統性篩檢結核病多使用長期咳嗽(逾兩週)或是任一結核病相關症狀做為指標。但此兩種篩檢指標分別有高敏感度低特異度和高特異度低敏感度，造成政策無法依照有限資源做最佳取捨。本研究目的為發展一結核病預測模型，並加入結核病臨床特徵和危險因子來進行預測。 方法 本研究使用南非尚比亞減少結核病與愛滋病的試驗(ZAMSTAR)中的橫斷性盛行率調查資料。其中僅篩選出愛滋病陰性或未知的參與者資料進行分析。臨床症狀和生活習慣之危險因子皆納入模型選擇考量中。我們使用南非模型建立資料庫建立多變項邏輯斯迴歸模型，計算預測因子預測結核病的能力，使用向後刪去法選擇AIC最小的模型當成預測最佳而不過度擬合之模型。模型簡化為分數系統後於南非模型驗證資料與尚比亞資料中驗證分數系統的預測能力，同時和目前常用的篩檢指標進行比較。最後，我們提出不同的篩檢模式，結合初步篩檢與確診工具，進行成本效益分析。 結果 最終模型所選出的預測因子與其相對應之分數為:長期咳嗽(2),短期咳嗽(2),夜間盜汗(2),酒精使用(2),個人結核病史(2),家族中結核病史(1),抽菸(1)和體重下降(1)。分數系統的整體預測能力(AUC:0.68,95%CI:0.64-0.72)高於目前常使用的篩檢指標(常期咳嗽AUC:0.58,任一症狀 AUC:0.60)。本研究所建立的預測系統在成本效益分析中也優於長期咳嗽或任一症狀做為初步篩檢工具的篩檢流程。 結論 本研究建立的預測模型整體預測與效用都優於目前的症狀指標，能提供彈性的篩檢工具選項，針對系統性篩檢計畫選擇最適於其計畫預算之篩檢工具。

Background Current strategies of systematic screening for active tuberculosis (TB) in resource-limited settings use prolonged cough or any TB symptoms as the initial screening criteria. The suboptimal sensitivity or specificity of these criteria lead to the difficulty of selecting the best strategy for TB screening. A prediction model of prevalent TB was developed to improve the use of relevant clinical information and assist decision-making of systematic screening. Methods Cross-sectional data from the prevalence survey of Zambia South Africa Tuberculosis and AIDS Reduction trial were used. Participants with HIV negative/ unknown were included in the analysis. Potential predictors included TB symptoms and risk factors, i.e., gender, life style, history of TB, and body mass index. Multivariable logistic regression model was constructed in the South Africa training set (No. of TB cases: 355) using the backward elimination method based on AIC, and the result was converted into risk scores. Performance of the scoring system was evaluated in the South Africa validation set (No. of TB cases: 176) and Zambia validation set (No. of TB cases: 107). Cost-effectiveness analyses were conducted comparing the scoring system against the conventional strategies (prolonged cough and any TB symptoms). Results The predictors and corresponding scores in the final model were cough >=2wks (5), cough <2wks (2), night sweats (2), personal TB history (2), ever drink (2), household TB history (1), ever smoke (1), and weight loss (1). AUC of the scoring system in South Africa validation set was 0.68 (95%CI: 0.64-0.72), compared against those for prolonged cough (0.58, 0.54-0.62) and any symptoms (0.6, 0.56-0.64). In Zambia dataset the AUC of the scoring system was 0.66 (0.60-0.72), which was also higher than those of current tools. Systematic screening algorithms with the scoring system as an initial screening tool dominated the algorithms using the two conventional tools in cost-effectiveness analyses. The ACER of the algorithm using the new scoring system ranged from 428 to 1848$USD per TB case detected under various cutoff score values in the SA validation set, and it ranged from 171 to 8,461$USD per TB case detected in the Zambia dataset. Conclusions This new prediction model had better discrimination performance and superior cost effectiveness than the conventional tools. It also provides flexible screening options for TB systematic screening programs to target those with high TB risk under budget constraints.