A Machine-Learning Framework for Forecasting Climate-Induced Shifts in Mosquito Vector Ecology and Arboviral Disease Risk Across the Middle East

Technical Summary: A Machine-Learning Framework for Forecasting Climate-Induced Shifts in Mosquito Vector Ecology and Arboviral Disease Risk Across the Middle East

Problem Statement
Vector-borne diseases (VBDs), particularly those transmitted by Aedes mosquitoes (dengue, chikungunya, Zika), are expanding their geographic and seasonal footprints due to climate change and urbanization. While global modeling efforts exist, the Middle East remains under-represented in high-resolution, predictive frameworks. Existing approaches in the region face critical limitations: sparse and fragmented entomological surveillance data, a reliance on single-dimension models (climate or land cover only), a lack of integration between mechanistic transmission dynamics and machine learning, and insufficient quantification of uncertainty across climate scenarios. Consequently, public health authorities lack actionable, spatially explicit tools to anticipate where and when vector habitats and transmission risks will intensify.

Methodology
The study develops a comprehensive, ensemble machine-learning (ML) framework designed to forecast Aedes vector suitability, abundance, and population-weighted arboviral transmission risk across the Middle East (18°–41°N, 34°–63°E) at a 5-km resolution.

1. Data Assembly and Synthetic Generation:

   • Entomological Data: Due to the lack of proprietary surveillance access, the authors generated a synthetic but ecologically realistic dataset for Aedes aegypti and Aedes albopictus. This was created using a hybrid statistical-stochastic framework calibrated to known ecological constraints, incorporating spatial logistic models for presence/absence and zero-inflated negative binomial distributions for abundance. The dataset includes 18,420 presence records, 22,775 absence records, and 9,110 abundance observations across 312 administrative districts, with simulated seasonal trajectories.
   • Climate Data: Projections were derived from six General Circulation Models (GCMs) from the Coupled Model Intercomparison Project Phase 6 (CMIP6) under three Shared Socioeconomic Pathways (SSPs): SSP1-2.6 (low emissions), SSP2-4.5 (intermediate), and SSP5-8.5 (high emissions). Data were downscaled to 5-km resolution using the Bias-Corrected Spatial Disaggregation (BCSD) method and bias-corrected against ERA5 reanalysis.
   • Covariates: The model integrated socio-environmental variables including land-use/land-cover, NDVI, night-time lights (urbanization proxy), population density, and a synthetic human mobility index.

2. Machine-Learning Architecture:

   • An ensemble of five supervised algorithms was constructed: Random Forest (RF), Extreme Gradient Boosting (XGBoost), LightGBM, Support Vector Machines (SVM), and a Deep Neural Network (DNN).
   • Models were trained separately for presence/absence classification and abundance prediction.
   • Hyperparameters were optimized via Bayesian optimization. The final ensemble prediction utilized weighted averaging based on out-of-sample AUC and inverse prediction variance.
   • Validation employed spatially stratified 80/20 splits and k-fold spatial cross-validation to mitigate autocorrelation bias.

3. Risk Modelling and Interpretability:

   • Transmission Risk: Vector suitability outputs were coupled with temperature-dependent extrinsic incubation period (EIP) functions and human exposure metrics (population density × mobility) to generate probabilistic transmission risk maps.
   • Interpretability: SHapley Additive exPlanations (SHAP) and partial dependence plots were used to elucidate non-linear drivers and identify critical climatic thresholds.

Key Results

• Model Performance: The ensemble framework demonstrated high predictive accuracy, achieving a mean out-of-sample AUC of 0.92 ± 0.01. The ensemble outperformed individual models, reducing variance in abundance predictions by 11%.
• Dominant Drivers: SHAP analysis identified mean monthly temperature, diurnal temperature range, and vapor-pressure deficit (VPD) as the primary climatic drivers (52% of variance), while urbanization and vegetation metrics contributed an additional 28%. Non-linear thresholds were identified: suitability sharply increases above 20°C, peaks at 28–30°C, and declines above 35°C.
• Spatial and Temporal Shifts:
  • Expansion: All SSP scenarios project a pronounced northward and altitudinal expansion of suitable Aedes habitats. Under the high-emission SSP5-8.5 scenario, high-suitability zones are projected to cover over 53% of the Middle East by the mid-21st century, including previously unsuitable arid interiors and high-altitude regions (e.g., Zagros Mountains).
  • Seasonality: Higher emission scenarios lead to earlier seasonal peaks (advancing by 2–3 months) and extended transmission windows.
• Risk Projections:
  • Under SSP2-4.5, approximately 45 million additional people are projected to face high-risk exposure by the 2050s.
  • Under SSP5-8.5, this figure rises to ~63 million by the 2080s, with risk extending into interior semi-arid plains and high-density urban centers.
  • Temperature thresholds and urbanization intensity were found to account for ~68% of the projected risk variance.

Significance and Claims
The paper claims to provide a novel, region-specific framework that bridges gaps in global vector-climate modeling for the Middle East. Its significance is threefold:

1. Methodological Rigor: It integrates multi-model CMIP6 projections with an ensemble ML approach and mechanistic transmission proxies (EIP), producing uncertainty-quantified forecasts that are both scientifically rigorous and operationally actionable.
2. Regional Relevance: By focusing on the under-represented Middle East, the study delivers intelligence tailored to the region's heterogeneous climates and socio-political landscape, addressing the specific needs of national and transboundary health authorities.
3. Operational Utility: By embedding interpretability (SHAP) and uncertainty quantification throughout the pipeline, the framework lowers barriers for policymakers, facilitating targeted surveillance, resource prioritization, and climate-adaptive mitigation strategies.

The authors conclude that climate change, urbanization, and ecological factors synergistically drive the expansion of Aedes habitats, necessitating adaptive public health strategies that account for earlier onset of vector activity, longer transmission seasons, and the emergence of risk in previously marginal zones. The study emphasizes that these findings provide a durable, updateable platform for guiding surveillance and control interventions in a region of increasing epidemiological vulnerability.