建立共識自組織對映於預測登革熱擴散的時空範圍

Abstract

發展登革熱每日滾動預測擴散模式目的，掌握登革熱疫情傳染的動向，採用時空預測的框架，達到精準匡列登革熱高風險區，縮小打擊範圍的進程，以2010年至2015年臺灣臺南市和高雄市的登革熱疫情作為研究案例。 我們將疫情分成兩個階段，疫情前期發病的登革熱病例資料作為訓練資料集，疫情中後期發病的登革熱病例資料作為測試資料集，為了發展預測模式，而建構時空擴散區的概念，並提出四種時空擴散每日滾動預測模式，模式一預設固定時空邊界、模式二歸納時空邊界、模式三利用空間異質性與模式四利用氣候、環境與空間異質性，模式三與模式四時空擴散區滾動預測模式都採用MST-DBSCAN (Modified space time DBSCAN algorithm)與自組織對映類神經網路採用共識決演算法，我們定義擴散預測的綜合指標，挑選最佳的登革熱滾動擴散預測模式。 根據研究模式預測成果，在登革熱疫情非爆發年份之2010年至2012年，模式四利用氣候、環境與空間異質性的綜合指標表現最佳，該模式在疫情中後期，每日預測高達80%的登革熱病例，在登革熱疫情爆發年份之2014年與2015年，模式三利用空間異質性的綜合指標表現最佳，該模式在疫情中後期，每日預測高達90%的登革熱病例，彙整上述模式預測細節，發現登革熱的傳染半徑約800公尺，在疫情非爆發年份發病間距是從6至11日，在疫情爆發年份發病間距是從6至9日。 最後我們開發動態地圖，視覺化呈現登革熱每日滾動預測的擴散特徵，根據最佳模式生成的時空擴散區細節，協助我們歸納登革熱的傳染半徑與發病間距。

The objectives of the study are to develop the daily rolling prediction models for the diffusion of dengue fever and to track the spread of the epidemic. The spatiotemporal prediction framework would accurately estimate the high-risk areas of dengue fever in near future for effective spatially targeting of epidemic control. Dengue fever epidemics in Tainan and Kaohsiung Cities of Taiwan from 2010 to 2015 were used for the case study. Our approach divided the epidemic into two stages: an early stage for data training and a middle-late stage for data testing. To develop our predictions, we incorporated the concept of spatial-time diffusion zone. Four models for daily rolling predictions of dengue fever diffusion are proposed. Model I is based on a fixed spatial-time boundary, while Model II summarizes the spatial-time boundary. Model III utilizes the concept of spatial heterogeneity, and Model IV integrates the concepts of climate, environment, and spatial heterogeneity. Both Models III and IV utilized MST-DBSCAN (Modified Space Time Density-Based Spatial Clustering of Applications with Noise) and a self-organizing map modified by consensus decision-making algorithm. A comprehensive indicator also defined for determining the best daily rolling prediction model for dengue fever diffusion. Our results show that, for the non-outbreak years of dengue fever (2010-2012), Mode IV, incorporating climate, environment, and spatial heterogeneity, performed the best with the comprehensive indicators. This model accurately predicted up to 80% of daily dengue cases during the middle-late stage of the epidemic. Conversely, for the outbreak years of dengue fever (2014-2015), Model III, considering spatial heterogeneity, performed the best with the comprehensive indicators. This model achieved daily prediction accuracy of up to 90% of dengue cases during the middle-late stage of the epidemic. Our model analysis also found that the infection radius of dengue fever was approximately 800 meters. In the non-outbreak years, the serial intervals ranged from 6 to 11 days, while in the outbreak years, the serial intervals ranged from 6 to 9 days. Finally, we created dynamic mapping animations for visualizing the diffusion patterns generated by the daily rolling prediction models of dengue fever. It provided detailed spatial-time diffusion zones, enabling us to summarize the infection radius and serial intervals of dengue fever.