Evans Michelle V, Ihantamalala Felana A, Randriamihaja Mauricianot, Herbreteau Vincent, Révillion Christophe, Catry Thibault, Delaitre Eric, Bonds Matthew H, Roche Benjamin, Mitsinjoniala Ezra, Ralaivavikoa Fiainamirindra A, Razafinjato Bénédicte, Raobela Oméga, Garchitorena Andres
MIVEGEC, Univ. Montpellier, CNRS, IRD, Montpellier, France.
NGO Pivot, Ranomafana, Ifanadiana, Madagascar.
Malar J. 2025 Jan 30;24(1):30. doi: 10.1186/s12936-025-05266-0.
The increasing availability of electronic health system data and remotely-sensed environmental variables has led to the emergence of statistical models capable of producing malaria forecasts. Many of these models have been operationalized into malaria early warning systems (MEWSs), which provide predictions of malaria dynamics several months in advance at national and regional levels. However, MEWSs rarely produce predictions at the village-level, the operational scale of community health systems and the first point of contact for the majority of rural populations in malaria-endemic countries.
This study developed a hyper-local MEWS for use within a health-system strengthening intervention in rural Madagascar. It combined bias-corrected, village-level case notification data with remotely sensed environmental variables at spatial scales as fine as a 10 m resolution. A spatio-temporal hierarchical generalized linear regression model was trained on monthly malaria case data from 195 communities from 2017 to 2020 and evaluated via cross-validation. The model was then integrated into an automated workflow with environmental data updated monthly to create a continuously updating MEWS capable of predicting malaria cases up to three months in advance at the village-level. Predictions were transformed into indicators relevant to health system actors by estimating the quantities of medical supplies required at each health clinic and the number of cases remaining untreated at the community level.
The statistical model was able to accurately reproduce village-level case data, performing nearly five times as well as a null model during cross-validation. The dynamic environmental variables, particularly those associated with standing water and rice field dynamics, were strongly associated with malaria incidence, allowing the model to accurately predict future incidence rates. The MEWS represented an improvement of over 50% compared to existing stock order quantification methods when applied retrospectively.
This study demonstrates the feasibility of developing an automatic, hyper-local MEWS leveraging remotely-sensed environmental data at fine spatial scales. As health system data become increasingly digitized, this method can be easily applied to other regions and be updated with near real-time health data to further increase performance.
电子健康系统数据和遥感环境变量日益容易获取，催生了能够进行疟疾预测的统计模型。其中许多模型已被应用于疟疾早期预警系统（MEWS），这些系统可在国家和地区层面提前数月预测疟疾动态。然而，MEWS很少在村庄层面进行预测，而村庄层面是社区卫生系统的运作规模，也是疟疾流行国家大多数农村人口的首个接触点。
本研究开发了一种超本地化的MEWS，用于马达加斯加农村的卫生系统强化干预。它将经过偏差校正的村庄层面病例通报数据与空间分辨率低至10米的遥感环境变量相结合。利用2017年至2020年195个社区的月度疟疾病例数据训练了时空分层广义线性回归模型，并通过交叉验证进行评估。然后将该模型整合到一个每月更新环境数据的自动化工作流程中，以创建一个能够在村庄层面提前三个月预测疟疾病例的持续更新的MEWS。通过估计每个卫生诊所所需医疗用品的数量以及社区层面未治疗病例的数量，将预测转化为与卫生系统行为者相关的指标。
该统计模型能够准确再现村庄层面的病例数据，在交叉验证期间的表现几乎是空模型的五倍。动态环境变量，特别是那些与积水和稻田动态相关的变量，与疟疾发病率密切相关，使该模型能够准确预测未来发病率。回顾性应用时，与现有的库存订单量化方法相比，MEWS的改进超过50%。
本研究证明了利用精细空间尺度的遥感环境数据开发自动、超本地化MEWS的可行性。随着卫生系统数据日益数字化，该方法可轻松应用于其他地区，并可通过近实时卫生数据进行更新，以进一步提高性能。
