運用搜尋引擎紀錄檔與機器學習模型於傳染病監測之研究

Abstract

傳染病無可避免的會造成大量的死亡以及重大的社會及經濟上的損失。傳染病監測因此早已成為重要的保健研究議題。在2009年，Ginsberg等人觀察到搜尋引擎紀錄檔可以被用來即時的評估當前傳染病的嚴重性狀態。在本論文中，我們將傳染病監測當成一個分類的問題並且應用Google的查詢統計資料來對登革熱傳染病的嚴重性狀態做分類。23個與登革熱有關的關鍵字查詢紀錄檔資料用作機器學習訓練與測試的觀察值，研究中也評估不同的機器學習模型在傳染病監測上的成果。根據在五年真實世界資料集上的實驗，證明了搜尋引擎紀錄檔可以被用來建立準確的傳染病狀態分類器。此外，經過學習的分類器也會表現的比傳統的回歸模型好。我們也應用了各種不同的機器學習模型如生成模型(generative model),判別模型(discriminative model)，序列化模型(sequential model)，以及非序列化模型(non-sequential model)來證明他們在傳染病監測上的適用性。

Epidemics inevitably result in a large number of deaths and always cause considerable social and economic damage. Epidemic surveillance has thus become an important healthcare research issue. In 2009, Ginsberg et al. observed that the query logs of search engines can be used to estimate the status of epidemics in a timely manner. In this paper, we model epidemic surveillance as a classification problem and employ query statistics from Google to classify the status of a dengue fever epidemic. The query logs of twenty-three dengue-related keywords serve as observations for machine learning and testing, and a number of machine learning models are investigated to evaluate their surveillance performance. Evaluations based on a 5-year real world dataset demonstrate that search engine query logs can be used to construct accurate epidemic status classifiers. Moreover, the learned classifiers generally outperform conventional regression approaches. We also apply various machine learning models, including generative, discriminative, sequential, and non-sequential classification models, to demonstrate their applicability to epidemic surveillance.