🧬 biology

TREA-Net: A Transferable Residual Epidemiological Adaptation Network for Dengue Incidence Forecasting

The paper proposes TREA-Net, a lightweight and transferable neural network that combines environmental epidemiological projections with a gated residual correction mechanism to significantly improve multi-week dengue forecasting accuracy in data-scarce regions by leveraging knowledge from data-rich areas.

Original authors: Inesh Shukla, Madhurima Panja, Tanujit Chakraborty, Chittaranjan Hens

Published 2026-07-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Inesh Shukla, Madhurima Panja, Tanujit Chakraborty, Chittaranjan Hens

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine the world of science as a giant, bustling library where researchers are constantly trying to predict the future. One very important corner of this library is dedicated to epidemiology, the study of how diseases spread. Think of it like trying to guess the path of a runaway ball in a crowded room. To make good guesses, scientists usually need two things: a deep understanding of the rules of physics (how the ball bounces) and a massive history of how the ball has rolled in the past. But what happens when a new room opens up, and you only have a few seconds of video to figure out how the ball moves there? This is the challenge facing health officials in many parts of the world. They need to predict outbreaks of diseases like dengue fever to send help before it's too late, but they often don't have enough historical data to train their computer models. This paper tackles that exact problem: how to use what we know from places with lots of data to help places with very little data, without needing to start from scratch.

The researchers behind this study, Inesh Shukla and his team, have built a clever new tool called TREA-Net. You can think of TREA-Net as a "smart translator" for disease forecasts. Usually, if you want to predict a dengue outbreak in a new city, you'd need years of local data to teach a computer how to do it. But TREA-Net is designed for cities that are just starting to keep records, perhaps only having 78 or 104 weeks of data (about 1.5 to 2 years). The system works by taking a "big brain" computer model that has already learned a lot from data-rich countries like Colombia and Nicaragua, and then giving it a tiny, lightweight "adjustment kit" to fit the new, data-scarce location.

Here is how the magic happens. The team uses two main ingredients. First, they have a "mechanistic" model called ETSIR. Imagine this as a weatherman who knows the rules of nature: he knows that if it rains a lot and gets warm, mosquitoes will breed, and dengue cases will likely go up. This model is based on the laws of biology and physics. Second, they have a "neural" model, which is a powerful AI that has learned to spot patterns in mountains of past data. The problem is that the AI might get confused by local quirks, and the weatherman might be too simple to catch complex trends. TREA-Net combines them by asking the AI to make a prediction, then asking the weatherman to make a prediction, and finally using a tiny, flexible "residual" module to figure out the difference between the two. It learns a correction rule that says, "When the AI and the weatherman disagree in this specific way, adjust the forecast by this amount."

What makes TREA-Net special is its ability to "transfer" this correction rule. The team trained this correction module on data from Colombia and Nicaragua, which have long, detailed histories of dengue. Then, they tried to use it in Mexico and Malaysia, where the data was much shorter. They found that they didn't need to retrain the whole system or change the number of cities being monitored. Instead, they only needed to tweak two tiny numbers (like adjusting the volume and the bass on a stereo) to make the system work perfectly for the new location.

The results were quite promising. When they tested this setup, TREA-Net improved the accuracy of the predictions in 9 out of 10 different scenarios they tried. In fact, when they paired TREA-Net with a very advanced AI called TiRex, it produced the most accurate forecasts of all, reducing the error rate significantly. For example, in Mexico, the system was able to predict dengue cases 8 weeks in advance with much greater precision than the AI could do alone. The researchers also showed that they could provide a "confidence range" for these predictions, telling health officials not just how many cases to expect, but how sure they could be about that number.

The paper suggests that this approach is a game-changer for places with limited resources. It proves that you don't need a massive supercomputer or decades of local data to get a good early warning system. By borrowing knowledge from places that have already seen a lot of outbreaks and applying a tiny, smart adjustment, health agencies in new or under-resourced regions can get the lead time they need to prepare. The authors emphasize that this isn't a magic bullet that solves everything, but it is a lightweight, portable framework that suggests we can do much better at protecting people from dengue, even when the data is scarce. They have even made their code available so other scientists can try it out and see if it works for their own local challenges.

Technical Summary: TREA-Net for Dengue Incidence Forecasting

Problem Statement
Accurate multi-week dengue forecasting is critical for vector-control interventions and healthcare resource allocation. However, newly established surveillance systems often lack the historical data required to train reliable neural forecasting models. While pre-trained time-series foundation models offer promising zero-shot capabilities, they often fail to capture local epidemiological dynamics due to their reliance on statistical patterns learned during pretraining rather than mechanistic knowledge. Conversely, purely mechanistic models (e.g., SIR variants) struggle with nonlinear interactions and behavioral responses. The central challenge addressed is how to transfer mechanistic epidemiological knowledge from data-rich source regions to structurally different, data-scarce target systems while preserving a pre-trained forecasting backbone and minimizing target-specific adaptation.

Methodology
The authors propose TREA-Net (Transferable Residual Epidemiological Adaptation Network), a modular transfer-learning framework designed for dengue forecasting under limited data. The methodology consists of three core components:

ETSIR Mechanistic Prior: The framework utilizes an Environmental Time-Series Susceptible–Infected–Recovered (ETSIR) model to encode climate-sensitive dengue dynamics. This model incorporates nonlinear effects of precipitation and temperature alongside lagged incidence. Unlike traditional approaches that embed mechanistic equations into the loss function, TREA-Net uses the ETSIR projection as a compact mechanistic prior input feature. Future weather inputs are replaced by climatological averages from the training period to prevent information leakage.
N-Invariant Gated Residual Correction: TREA-Net employs a lightweight, shared-weight neural module that learns a residual correction. This module takes as input the normalized forecasts from a pre-trained backbone (e.g., LSTM, TiRex) and the ETSIR projection, along with their absolute discrepancy. It outputs a gate value ( $g$ ) and a residual adjustment ( $\Delta$ ) to correct the backbone forecast. Crucially, this module is node-invariant (N-invariant), meaning it contains no node-specific parameters or fixed spatial topology. It operates independently on each horizon-region pair, allowing it to be transferred across surveillance systems with different numbers of administrative units.
Two-Scalar Target Adaptation: To adapt the model to a new, data-scarce target region without retraining the backbone or the correction module, TREA-Net introduces only two global, target-specific parameters: a scale factor ( $\gamma$ ) and an offset ( $\beta$ ). These parameters adjust the magnitude and offset of the transferred residual correction. This design ensures that adaptation requires learning only two parameters, regardless of the number of regions in the target system, thereby preventing overfitting in low-data regimes.

Key Contributions

Transferable Mechanistic Adaptation: The paper formulates epidemiological guidance as a latent residual correction using ETSIR projections and pre-trained forecasts. The resulting adapter is point-wise and N-invariant, requiring only two global parameters for target adaptation, which facilitates transfer across systems with varying numbers of monitored regions.
Multi-Country Low-Data Evaluation: The authors demonstrate the framework's efficacy by transferring knowledge from data-rich systems (Colombia and Nicaragua) to data-scarce targets (Mexico and Malaysia) using only 78 or 104 weeks of target observations. The evaluation spans five neural backbones and ten transfer settings.
Open Framework: The authors commit to releasing the TREA-Net implementation and evaluation pipeline to support reproducible assessment of user-defined forecasting models under realistic low-data transfer settings.

Results
Experiments were conducted on weekly subnational dengue data from Colombia, Nicaragua, Malaysia, and Mexico. The framework was tested against five forecasting backbones: LSTM, N-HiTS, TCN, PatchTST, and the pre-trained foundation model TiRex.

Performance Improvement: Across five backbones and ten transfer settings, TREA-Net improved the corresponding backbone in 9 out of 10 settings with statistically significant gains (Wilcoxon signed-rank test, $\alpha=0.05$ ).
Comparison with Baselines: The standalone ETSIR model produced substantially higher errors, particularly in data-scarce scenarios. Hybrid models like ETSIR-PINN showed improvements but did not outperform the strongest sequence-model baselines.
Optimal Configuration: Integrating TREA-Net with the zero-shot TiRex foundation model achieved the lowest Mean Absolute Error (MAE) across all target datasets, reducing error by an average of 7.1% relative to the zero-shot TiRex baseline.
Ablation Studies: Removing the ETSIR feature increased MAE by 7.5% (Malaysia) and 11.5% (Mexico), confirming the value of the mechanistic prior. Training the correction solely on target data (without multi-source transfer) performed worse, highlighting the benefit of multi-source learning. Replacing the two global adaptation parameters with node-specific parameters provided no consistent improvement and increased the risk of overfitting.
Uncertainty Quantification: When combined with Conformal Prediction (EnbPI), TREA-Net maintained empirical coverage while reducing the 8-week prediction-interval width by 29.6% in Mexico compared to the zero-shot baseline.

Significance
The paper positions TREA-Net as a lightweight and portable early-warning framework specifically designed for health agencies operating with limited surveillance data and computational resources. By decoupling the epidemiological correction from the forecasting backbone and requiring only minimal target-specific adaptation, the framework addresses the critical gap in deploying reliable epidemic forecasts in newly established or data-scarce surveillance systems. The authors claim that this approach supports earlier vector-control planning, healthcare resource allocation, and risk-informed decision-making in resource-constrained regions without the need for extensive retraining or large historical datasets.

Technical Summary: TREA-Net for Dengue Incidence Forecasting

More like this