[论文解读] Data-Driven Forecast of Dengue Outbreaks in Brazil: A Critical Assessment of Climate Conditions for Different Capitals

Lucas M. Stolerman, Pedro D. Maia+1 more|arXiv (Cornell University)|Dec 31, 2016

Mosquito-borne diseases and control参考文献 33被引用 5

一句话总结

本研究基于七个巴西首都的气候时间序列数据，开发了一种数据驱动的登革热疫情预测模型，识别出冬季气温和降水频率为关键预测因子。研究结果表明，疫情爆发前六至九个月的特定气候特征可实现准确的早期预警，城市间差异化的气候模式凸显了制定本地化干预策略的必要性。

ABSTRACT

Local climate conditions play a major role in the development of the mosquito population responsible for transmitting Dengue Fever. Since the {\em Aedes Aegypti} mosquito is also a primary vector for the recent Zika and Chikungunya epidemics across the Americas, a detailed monitoring of periods with favorable climate conditions for mosquito profusion may improve the timing of vector-control efforts and other urgent public health strategies. We apply dimensionality reduction techniques and machine-learning algorithms to climate time series data and analyze their connection to the occurrence of Dengue outbreaks for seven major cities in Brazil. Specifically, we have identified two key variables and a period during the annual cycle that are highly predictive of epidemic outbreaks. The key variables are the frequency of precipitation and temperature during an approximately two month window of the winter season preceding the outbreak. Thus simple climate signatures may be influencing Dengue outbreaks even months before their occurrence. Some of the more challenging datasets required usage of compressive-sensing procedures to estimate missing entries for temperature and precipitation records. Our results indicate that each Brazilian capital considered has a unique frequency of precipitation and temperature signature in the winter preceding a Dengue outbreak. Such climate contributions on vector populations are key factors in dengue dynamics which could lead to more accurate prediction models and early warning systems. Finally, we show that critical temperature and precipitation signatures may vary significantly from city to city, suggesting that the interplay between climate variables and dengue outbreaks is more complex than generally appreciated.

研究动机与目标

• 利用历史数据识别预测巴西首都登革热疫情的关键气候条件。
• 开发基于机器学习的预测系统，结合降维技术和气候时间序列分析。
• 确定气温和降水频率的特定气候特征是否可提前数月预测疫情爆发。
• 评估基于气候变量（尤其是疫情前冬季）的早期预警系统的可行性。
• 突出城市间气候-登革热关系的差异，挑战普遍适用预测规则的假设。

提出的方法

• 对七个巴西首都的气候时间序列数据应用降维技术。
• 使用机器学习算法检测流行病与非流行病气候条件之间具有高可分性的时段。
• 识别出疫情爆发前两个月的冬季窗口，此时气温和降水频率最具预测性。
• 采用压缩感知技术重建记录不完整的气温和降水数据。
• 基于历史气候-疫情配对数据训练预测模型，以识别早期预警信号。
• 在各城市间验证结果，评估气候-登革热关系的一致性与变异性。

实验结果

研究问题

• RQ1 哪些气候变量和时间窗口对巴西首都的登革热疫情最具预测性？
• RQ2 冬季气温和降水频率能否作为未来登革热流行病的可靠早期指标？
• RQ3 气候-登革热关系在巴西不同城市环境中的差异如何？
• RQ4 数据驱动模型在多大程度上可提升公共卫生干预的提前时间？
• RQ5 是否存在普遍适用的气候特征，还是预测模式需要基于城市进行校准？

主要发现

• 在两个月的冬季窗口期内，降水频率和平均气温是预测登革热疫情最具预测力的气候变量。
• 在大多数首都，基于冬季气候条件，登革热疫情可提前6至9个月预测。
• 里约热内卢在冬季降水频率约为每周一次且平均气温约为22°C时，显示出强烈的预测信号。
• 每个首都均表现出独特的气候特征，表明通用的预测模型可能无法有效适用。
• 夏季月份的预测价值较低，因为流行病年份与非流行病年份的气候条件在这一时期在定性上相似。
• 压缩感知有效恢复了缺失的气候数据，使在记录不完整的城市中也能开展分析。