Application of a Temporal Fusion Transformer and Long-Term Climate and Disease Data to Assess the Predictive Power and Understand the Drivers for Malaria and Dengue

TL;DR

This study uses advanced AI models and climate data to predict and understand malaria and dengue outbreaks, helping public health teams prepare in advance.

Contribution

The study introduces a deep-learning framework using Temporal Fusion Transformers to predict malaria and dengue with high accuracy and interpretability.

Findings

The best malaria model achieved an R2 of 0.95 and an MAE of 4.98.

Extreme temperature and rainfall metrics were the strongest predictors of disease outbreaks.

ENSO and IOD improved longer-range malaria forecasts and revealed non-stationary climate–disease relationships.

Abstract

Public health relevance—How does this work relate to a public health issue?
Malaria and dengue remain major vector-borne disease burdens whose transmission is strongly shaped by climate variations and forcing.These diseases are difficult to detect in real time because clinical case data often lag behind environmental changes.

Malaria and dengue remain major vector-borne disease burdens whose transmission is strongly shaped by climate variations and forcing.

These diseases are difficult to detect in real time because clinical case data often lag behind environmental changes.

Public health significance—Why is this work of significance to public health?
The model identifies which climate features become most informative at different prediction horizons, giving public health programs clearer signals about when environmental changes become actionable.By extracting climate “risk-profiles” linked to elevated disease incidence, the study helps translate complex environmental information into practical indicators for malaria and dengue surveillance teams.

The model identifies which climate features become most informative at different prediction horizons, giving public health programs clearer signals about when environmental changes become actionable.

By extracting climate “risk-profiles” linked to elevated disease incidence, the study helps translate complex environmental information into practical indicators for malaria and dengue surveillance teams.

Public health implications—What are the key implications or messages for practitioners, policy makers and/or researchers in public health?
The study provides concrete temperature and rainfall ranges that typically precede higher malaria and dengue activity, offering clear environmental cues to monitor before cases rise.The deep-learning framework shows that routinely available climate variables can support reliable early prediction, improving situational awareness and guiding when to intensify vector-control or preparedness measures.

The study provides concrete temperature and rainfall ranges that typically precede higher malaria and dengue activity, offering clear environmental cues to monitor before cases rise.

The deep-learning framework shows that routinely available climate variables can support reliable early prediction, improving situational awareness and guiding when to intensify vector-control or preparedness measures.

Vector-borne diseases are strongly influenced by climate, yet the magnitude and temporal variability of climate–disease relationships remain poorly quantified. Outbreaks occur abruptly, and responses are typically delayed, underscoring the need for predictive tools that can support proactive interventions. This study applies Temporal Fusion Transformers (TFTs) to long-term, high-resolution climate datasets and to weekly malaria and dengue case records from South Africa and Vietnam to assess predictive performance and identify key environmental drivers. The models incorporated diverse climatic predictors and large-scale climate indices and were trained using multi-horizon forecasting with novel loss functions and physics-based constraints. The best malaria model achieved an R2 of 0.95 and an MAE of 4.98, while leading dengue models reached R2 values up to 0.90. Variable-importance analyses derived from model-learned weights showed that extreme temperature and rainfall metrics were consistently the strongest predictors, with ENSO (El Niño Southern Oscillation) and IOD (Indian Ocean Dipole) improving longer-range malaria forecasts. Furthermore, climate–disease risk dynamics were explored, revealing specific temperature and rainfall thresholds associated with elevated transmission and highlighting non-stationary relationships across decades. These findings demonstrate accurate, interpretable forecasting offered by TFTs and represent a valuable tool for early warning and understanding of complex climate–disease interactions.

Figures