Climate Change Poses a Significant Challenge to the Current Dynamics of Disease Outbreaks. This Study

Improves Outbreak Prediction Using an Attention-Based LSTM Model Optimized by the Honey Badger

Algorithm (HBA) for Hyperparameter Tuning.

Using Disease, Climate, and Geographic Data From 2015–2024 in Different Barangays in Koronadal, South

Cotabato, the Model Predicts Incidence Over 1-, 3-, 6-, and 12-Month Horizons. Attention Mechanisms

Enhanced Long-Term Pattern Detection, While HBA Reduces Overfitting and Boosts Accuracy. Results Show

the HBA-LSTM Reduces Mean Squared Error by 43.7% Over Standard LSTM and 22.2% Over Attention

Models. Similar Reductions are Seen in RMSE, MAE, and MAPE. Though Effective, Further Tuning and

Alternative Architectures are Suggested for Improved Generalization.

Developing predictive models that effectively utilize climate data is essential, as they significantly enhance

public health interventions. Deep learning is particularly well-suited for processing complex, high-dimensional

data, making it ideal for modeling interactions between climate and disease. Auto-Regressive Integrated Moving

Average (ARIMA) and Gated Recurrent Units (GRU) have been applied to predictive disease outbreaks using

climate data ( [23]; [33]).

LSTM models are subsets of Recurrent Neural Networks that overcome vanishing gradients through their

memory cells. However, they face challenges in overfitting and long-term prediction accuracy. In stock market

return forecasting, LSTM models demonstrate potential but still struggle with generalization to unseen data,

indicating significant overfitting([32]).

Overfitting can be addressed through various techniques, including grid search and L2 regularization ([7]) and

optimization technique fine-tunes model parameters to search for relevant values within the search space. When

faced with local optima, it can escape from these traps, enhancing the overall optimization process ([13]:[14]).

Since outbreak data is irregular and non-stationary, attention mechanisms may not effectively capture these

trends. The researchers propose an Attention-based LSTM optimized for hyperparameters using the Honey

Badger Algorithm. This approach aims to improve forecasting with lags of 1, 3, 6, and 12 months, comparing it

to stacked LSTM and Attention stacked-LSTM models.

Climatic and geographic factors significantly influence the incidence of vector-borne diseases, such as dengue

fever, malaria,

and

chikungunya.

Temperature,

humidity,

and

noise. Integrating LSTM with time-series data significantly enhances prediction accuracy. Metrics such as

RMSE, MAE, and MSE demonstrate that these hybrid models reduce prediction errors and generate smoother

Recent research on spatiotemporal attention mechanisms has enhanced prediction accuracy in multivariate time

series forecasting by capturing complex dependencies among variables. Models like LSTNet utilize

spatiotemporal self-attention for improved results.[29] Advances such as STAM ([11]) and symmetric kernel

attention mechanisms ([25]) further refine long— and short-term dependencies. Integrating attention modules

into LSTM models improves memory retention and predictive accuracy ([22]), while multi-head attention

enhances spatiotemporal relationship analysis ([36]).

The Honey Badger Algorithm (HBA) is a bio-inspired metaheuristic optimization technique that mimics honey

badgers’ foraging behavior to balance exploration and exploitation, making it effective for solving complex

optimization problems ([15]; [14]). In deep learning, HBA fine-tunes weights, minimizes loss, and enhances

generalization. The Global Optimization Honey Badger Algorithm (GOHBA) boosts convergence and search

diversity by optimizing hyperparameters like learning and dropout rates to prevent overfitting [16].

fundamental values [28].

MSE in a regression task squares the errors before averaging, giving more weight to larger deviations. This is

particularly useful in outbreak prediction, as it penalizes extreme errors, ensuring the model is sensitive to sudden

spikes in disease cases. [31].

transformation makes the error more interpretable because it expresses the error in the same units as the original

data.

Data Preprocessing

Irrelevant columns were removed from the dataset to focus solely on the relevant features. Since there were no

missing values, imputation was not required. Categorical variables were converted to numerical values using

label encoding with the sci-kit-learn library. These preprocessing steps effectively structured the dataset for

training purposes.

Feature Engineering

Key climatic factors such as temperature, precipitation, and humidity, along with geographic features like Land

Surface Temperature and NDVI, were preserved. Lagged outbreak cases were created at 1, 3, 6, and 12-month

Feature Scaling

LSTM models are sensitive to input values, so feature scaling was used to ensure equal contribution from all

features include feature-engineered variables such as cases from previous months, and climatic and geographic

the encoder are passed to the decoder. The encoder outputs store feature-rich representations for attention

processing.

Exploitation Phase: In the exploitation phase, the focus is on fine-tuning hyperparameters through localized

3.72

2.23

6.60

3.18

1.99

13.68

RMSE, MAE, and MAPE. Specifically, in the 1-month lookback, HBO reduced MSE by 43.7% compared to S-

LSTM and by 22.2% compared to the Attention model. Similar trends are observed in all error metrics, with the

3-month lookback showing a 43.8% reduction in MSE versus LSTM. This pattern of superior performance

continues through the 6-month and 12-month windows.

Fig. 3. Time Series Visualization of S-LSTM, S-Att-LSTM-HBO LSTM

lowest RMSE values across all time windows (4.04, 3.34, 4.30, and 3.15), this improvement is not enough to

accurately capture the extreme variability in dengue transmission.

population_si

ze: 20

(3)

Dropout: [0.2,0.4],

c : 1.2

ze: 20

Att Dropout: [0.2,0.4],

Embedding Dim: [64,256]

Learning Rate: [0.0001,0.005]

beta : 6

max_iteratio

ns: 5

LSTM Units: [128,512]

ze: 10

(12)

Dropout: [0.2,0.4],

Att Dropout: [0.2,0.4],

Embedding Dim: [64,256]

[128,512]

c : 1.2

beta : 6

max_iteratio

ns: 5

population_si

Table 2 shows the optimal Honey Badger Optimization configurations across various lookback windows. The

analysis demonstrates a strategic tradeoff between window size and optimization parameters, with smaller

windows (1, 3) requiring more extensive exploration (population size: 20, iterations: 15), while larger windows

(6, 12) perform better with focused search strategies (population size: 10, iterations: 5). Performance metrics

vary significantly across configurations, with window size 6 achieving superior prediction accuracy (RMSE:

21.19), window size 1 minimizing absolute error (MAE: 17.38), and window size 3 offering the most

proportionally accurate forecasts (MAPE: 52.99%). However, computational demands escalate with window

provides valuable insights into the potential and limitations of predictive epidemiological modeling, especially

using hyperparameter optimizations. It highlights fundamental constraints within the current modeling methods,

likely stemming from a lack of sufficient data on non-linear interactions among environmental factors, vector

dynamics, and human movement patterns.

These results underscore the need for better modeling techniques that consider additional factors, such as climate

variability, population immunity, and vector abundance. Using multi-modal ensemble frameworks could also

improve understanding of the complex relationships driving dengue transmission.

The researchers recommend conducting a more detailed analysis of outbreak dynamics, particularly focusing on

non-linear relationships such as mobility data, internet trends, and changes in land use.

When modeling, especially with deep learning models, it is crucial to choose the appropriate hyperparameters

A. G., Sampaio,