建立並驗證結核病預測模型與結核病社區篩檢策略之成本效果分析

Development, validation, and cost-effectiveness analysis of a prediction model for community-based systematic screening of active tuberculosis

楊芷其(Chih-Chi Yang)

指導教授 ： 林先和

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(2019年)

https://doi.org/10.6342/NTU201903404

摘要

背景:
尚未被發現的結核病個案對於結核病控制是很大的挑戰。經由主動個案發現的策略可以有效找到結核病患，但是成本高昂。資源有限的國家適用以症狀為主的篩檢模式，然而此類工具所存在的敏感度及特異度問題使得整套個案發現模式需耗費大量資源，也因此造成了相當沉重的負擔。本研究發展並驗證一套活動性結核病預測模型，並與目前世界衛生組織所建議的症狀篩檢模式比較，進行成本效果分析。
方法:
研究族群來自南非/尚比亞減少結核病與愛滋病試驗(ZAMSTAR)中收集於2010年之盛行率資料。預測模型依照研究對象的愛滋病感染有無，分開建模。模型之依變項為經痰培養之結核分枝桿菌，預測因子包含結核病相關症狀、結核病危險因子以及過去個人/家戶病史。模型以多變項羅吉斯迴歸模型建立於南非資料集，基於赤池資訊準則(AIC)經逐步向後消去法選出最適模型，並轉換為分數系統，於尚比亞資料集做外部驗證。內部驗證為重複模型建立過程於1000組經自助抽樣法抽出之樣本。此外，預測模型與現有其他以症狀為基礎的篩檢策略透過成本效果分析，計算平均成本效果比(ACER)及增加成本效果比(ICER)，以選出具成本效果的策略。
結果:
針對愛滋病族群的最適模型所包含之預測因子與其相對應之分數如後：性別(1)、體重減輕(1)、近期咳嗽(2)、飲酒(2)及胸痛(1)，模型之預測能力在南非資料集中優於任一結核病相關症狀，於尚比亞資料集中之預測能力也不亞於目前的症狀篩檢策略。而針對非愛滋病族群的最適模型結果如後：體重減輕(3)、夜間盜汗(2)、短期咳嗽(3)、長期咳嗽(7)、飲酒(2)、吸菸(1)、性別(1)及家戶結核病史(2)，模型預測能力在南非資料集及尚比亞資料集中皆優於任一結核病相關症狀及長期咳嗽。另外，經成本效果分析發現，於所有可能的篩檢模式之中，20組分數系統切點組合及16組分數系統切點組合分別在南非資料集與尚比亞資料集被選為具成本效果的篩檢策略。當具成本效果的分數系統切點組合實際應用在結核病個案發現時，平均找到一個活動性結核病患的成本範圍將從246美金至1670美金(於南非)，及從186美金至6796美金(於尚比亞)。
結論:
本研究提出一套活動性結核病預測模型，除根據研究對象之愛滋病感染狀況分別建模，也與目前世界衛生組織建議之以症狀為主的篩檢策略在成本效果的考量上做綜合性的比較。模型除在預測能力及成本效果分析中優於現行篩檢策略，也在願意投入之金錢與檢出結核病個案之成效的決策點上提供多樣且彈性的選擇。

關鍵字

活動性結核病 ； 個案發現 ； 系統性篩檢 ； 預測模型 ； 分數系統 ； 成本效果分析

並列摘要

Background:
Undetected tuberculosis (TB) cases were a challenge to global TB control. Active case finding could effectively detect TB patients but costly. For those resource-constrained countries, symptom-based screening algorithms were applicable, but at the same time resulted in a massive burden to the screening program due to the low sensitivity or low specificity of the screening tools. This study aimed to develop and validate a score-based screening algorithm for active TB case finding in the community. Furthermore, through cost-effectiveness analysis, we aimed to determine the cost-effective strategies for active TB case finding among the competing algorithms including the model developed in this study and the symptom-based tools proposed by the World Health Organization (WHO).
Methods:
The study population was based on a 2010 TB prevalence survey in the Zambia/South Africa Tuberculosis and AIDS Reduction (ZAMSTAR) trial. We developed separate prediction models for HIV-positive and HIV-negative/unknown populations. The dataset was divided into two parts according to participants' countries for model development (South Africa) and external validation (Zambia). The outcome was prevalent culture-confirmed TB. The potential predictors included TB symptoms, TB risk factors, and previous TB history. The models were built on multivariable logistic regression and selected through stepwise backward elimination based on Akaike Information Criterion. The final model was converted to a scoring system. Secondly, the scoring system was compared with any TB symptom and prolonged cough, through the cost0effectiveness analysis. The Average Cost Effectiveness Ratio (ACER) and Incremental Cost Effectiveness Ratio (ICER) were computed to evaluate the cost-effective strategies among all competing choices.
Results:
The predictors selected in the final model for HIV population and the corresponding scores are shown as follows: gender (1), weight loss (1), ever drink (2), current cough (2), and chest pain (1). The model was presented better AUC than any TB symptoms in both South African and Zambian dataset. On the other hand, the model for non-HIV population included the predictors: weight loss (3), night sweats (2), cough less than two weeks (3), prolonged (more than two weeks) cough (7), ever drink (2), ever smoke (1), gender (1), and household TB history (2). The model also presented the higher AUC than any TB symptoms and prolonged cough among the participants from two different countries. Moreover, through cost-effectiveness analysis, twenty score cutoff combinations were selected as the cost-effective strategies among the tools proposed by WHO and other score cutoff combinations in the South African dataset. The similar results were presented in Zambian dataset, in which sisteen score cutoff combinations were cost-effective in comparison to other screening strategies. When the cost-effective score cutoff combinations were applied for TB screening, the ACER ranged from 246 USD to 1670 USD in the South African dataset, and ranged from 186 USD to 6796 USD in the Zambian dataset.
Conclusion:
The scoring system for active TB case finding presented higher performance than any TB symptoms and prolonged cough among study population. The results of cost-effectiveness analysis showed that both the scoring systems for HIV and the non-HIV population were cost-effective at some score cutoff combinations in comparison to the tools proposed by WHO, and therefore provided new strategies for active TB case finding with multiple options under budget consideration.

並列關鍵字

active tuberculosis ； case finding ； systematic screening ； prediction model ； scoring system ； cost-effectiveness analysis

參考文獻

1. Global tuberculosis report 2018, W.H. Organization, Editor. 2018: Geneva.

2. Golub, J.E., et al., Active case finding of tuberculosis: historical perspective and future prospects. Int J Tuberc Lung Dis, 2005. 9(11): p. 1183-203.

3. Ho, J., G.J. Fox, and B.J. Marais, Passive case finding for tuberculosis is not enough. International Journal of Mycobacteriology, 2016. 5(4): p. 374-378.

4. Corbett, E.L., et al., Comparison of two active case-finding strategies for community-based diagnosis of symptomatic smear-positive tuberculosis and control of infectious tuberculosis in Harare, Zimbabwe (DETECTB): a cluster-randomised trial. Lancet (London, England), 2010. 376(9748): p. 1244-1253.

5. Organization, W.H., Systematic Screening for Active Tuberculosis: Principles and Recommendations. 2013: Geneva.