📊 statistics

Temporal Modeling of Weekly Dengue Cases in Marikina City: A Time-Respecting Comparison of Classical, Machine Learning, and Deep Temporal Models

This study evaluates classical, machine learning, and deep learning models for forecasting weekly dengue cases in Marikina City using a leakage-safe temporal design, finding that while simple baselines performed well during internal validation, the LSTM model achieved superior predictive accuracy on the 2025 out-of-time holdout despite a general tendency to underpredict outbreak-level weeks.

Original authors: Kal-el John C. Bautista, Luis Arnold N. Respecio, Cyril Troy C. Vinuya, John Paul Q. Tomas, Karl C. Ablola, Bonifacio T. Doma

Published 2026-08-31

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Kal-el John C. Bautista, Luis Arnold N. Respecio, Cyril Troy C. Vinuya, John Paul Q. Tomas, Karl C. Ablola, Bonifacio T. Doma

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the humid, densely packed neighborhoods of tropical cities, a persistent threat lurks in the standing water of everyday life: the dengue fever virus. Transmitted by mosquitoes, this illness can overwhelm local clinics and strain community resources when cases spike unexpectedly. For city health officials, the difference between a manageable situation and a crisis often comes down to timing. Knowing that cases are likely to rise two weeks from now allows teams to stockpile medical supplies, deploy vector-control workers to specific streets, and alert the public before the surge hits. However, predicting these spikes is notoriously difficult. The number of reported cases does not follow a simple, straight line; it waxes and wanes with the seasons, reacts to local weather, and is influenced by how quickly people seek care. To navigate this uncertainty, scientists turn to forecasting, a field that uses past patterns to guess future outcomes. The challenge lies in distinguishing between a genuine signal that predicts the future and a random pattern that merely looks like one. If a model is tested incorrectly—for instance, by letting it peek at future data during its training—it might appear brilliant but fail completely when faced with a real, unseen outbreak.

This study focused on Marikina City in the Philippines, a place where dengue is a recurring public health concern. Researchers set out to build a system that could predict the number of weekly dengue cases exactly two weeks in advance, using only the history of reported cases themselves. They gathered data spanning nearly a decade, carefully removing years where reporting was too sparse to be useful, and organized the timeline so that the model learned from the past without ever seeing the future. The team tested eight different types of prediction tools, ranging from classical statistical methods that have been used for decades to modern machine learning algorithms and deep neural networks designed to recognize complex patterns in sequences. To ensure the results were trustworthy, they split their data into distinct time periods: one for teaching the models, one for tuning them, and a final, completely unseen year to serve as a strict test of their real-world performance.

The results revealed that the best tool depends heavily on the specific year being predicted. When the models were tested on data from 2024, a traditional statistical method called ARIMA performed the strongest among the trained systems. This approach, which relies on the idea that recent trends are the best guide for the immediate future, was slightly outperformed, however, by a very simple baseline: a moving average of the last two weeks. This finding served as a crucial reminder that complex algorithms do not always beat simple logic, especially when the data is noisy. But when the researchers applied the same models to the completely unseen data from 2025, the ranking shifted. A deep learning model known as a Long Short-Term Memory network, or LSTM, emerged as the clear winner. This type of artificial intelligence is designed to remember long sequences of events and learn how they connect over time. In the 2025 test, the LSTM model predicted the weekly case counts with an average error of just under eight cases, outperforming the traditional ARIMA model and the simple baselines.

Despite the success of the advanced models, the study highlighted a significant limitation that affects all forecasting methods: they struggle to predict the most dangerous moments. When the city faced weeks with very high numbers of cases, indicating an active outbreak, every model tended to underestimate the severity. The predictions were consistently lower than the actual numbers observed during these peak weeks. This suggests that while the models are excellent at tracking the general flow of the disease and anticipating moderate increases, they are less effective at capturing the sudden, explosive growth that characterizes a true epidemic. The researchers found that the most influential factors for the machine learning models were simply the number of cases reported in the current week and the average number of cases over the previous two weeks. The models relied heavily on this recent history rather than on complex seasonal patterns, reinforcing the idea that dengue transmission in this city is driven strongly by immediate, short-term momentum.

The study concludes that a surveillance-only approach, which uses only case counts without needing external data like weather or population movement, is a viable and practical way to support local health planning. The LSTM model showed the most promise for the future, offering the most accurate forecasts for the unseen year, while the ARIMA model remains a valuable, transparent, and computationally efficient alternative. However, the researchers caution that no single model should be used in isolation. Because the models consistently underpredicted the worst outbreaks, any real-world system must include simple baselines and human oversight to ensure that health officials are not caught off guard by a sudden surge. The ultimate goal is not to replace human judgment with a computer prediction, but to provide a reliable early warning that, when combined with expert review, helps communities prepare for the next wave of dengue before it arrives.

Technical Summary: Temporal Modeling of Weekly Dengue Cases in Marikina City

Problem Statement

Dengue fever remains a persistent public health challenge in tropical urban settings, where short-term surges in incidence can rapidly overwhelm local clinical services and vector-control operations. While city-level forecasts are critical for operational decision-making (e.g., staffing, inventory, and community interventions), existing forecasting studies often suffer from methodological flaws. These include random train-test splits that violate temporal order, information leakage from future data, insufficient baseline comparisons, and a lack of rigorous out-of-time evaluation. Furthermore, many local health units lack access to synchronized meteorological or mobility data, necessitating a "surveillance-only" approach that relies strictly on historical case counts and reporting dates.

This study addresses these gaps by developing and evaluating a leakage-safe, two-week-ahead forecasting framework for weekly dengue incidence in Marikina City, Philippines. The primary objective is to determine whether trained models (classical, machine learning, and deep temporal) can genuinely outperform simple persistence baselines when evaluated under strict temporal constraints.

Methodology

Data Sources and Preparation

The study utilized a unified dataset combining Department of Health–Epidemiology Bureau records (2016–2020) and Marikina Epidemiology Surveillance Unit (MESU) data (2023–2025).

Exclusions: The years 2021 and 2022 were excluded due to sparse reporting (0 and 3 cases, respectively), which would have disrupted lag and sequence modeling.
Dataset Structure: The final cleaned dataset comprised 417 weekly observations with 8,263 reported cases.
Temporal Segmentation: The data was split strictly by target year to prevent leakage:
- Training/Development: 2016–2020 and 2023 (313 weeks).
- Internal Validation: 2024 (52 weeks) for model selection and hyperparameter tuning.
- External Holdout: 2025 (52 weeks) as a fully unseen, out-of-time evaluation set.

Feature Engineering and Models

The task was defined as univariate regression: predicting the case count at $t+2$ using information available at $t$ .

Baselines: Five reference models were calculated without training: a current-count forecast and moving averages over 2, 4, 8, and 12 weeks.
Classical Models: ARIMA, SARIMA, and Exponential Smoothing (ETS). ARIMA(3,0,0) was selected via rolling-origin refitting; SARIMA and ETS used static full-horizon forecasts.
Machine Learning (Tree-Based): Decision Tree, Random Forest, and Gradient Boosting. Features included current counts, lags (1, 2, 4, 8, 12 weeks), rolling statistics (means, sums, standard deviations), trend slopes, and seasonal encodings (sine/cosine).
Deep Temporal Models: Long Short-Term Memory (LSTM) and Temporal Convolutional Neural Network (CNN). Inputs were constructed as 12-week case sequences. The LSTM used a single recurrent layer (32 units), while the CNN used two 1D convolutional layers.

Validation Protocol

A strict "target-year" assignment was employed. Observations were assigned to splits based on the year of the target variable, not the forecast origin. This ensured that 2025 targets were never used for model selection, hyperparameter tuning, or early stopping. For the 2025 holdout, earlier observed 2025 counts were permitted as causal inputs for later forecasts, maintaining temporal causality while keeping the targets unseen.

Key Results

Performance in Internal Validation (2024)

During the 2024 internal validation, ARIMA was the strongest trained model (MAE: 11.789, RMSE: 15.028, $R^2$ : 0.680). However, it did not outperform the simple baselines; the two-week moving average (MAE: 11.480) and the current-count forecast (MAE: 11.510) achieved slightly lower errors. Static seasonal models (SARIMA, ETS) performed poorly, yielding negative $R^2$ values.

Performance in External Holdout (2025)

The ranking shifted significantly in the unseen 2025 holdout:

LSTM: Achieved the best performance among all trained models (MAE: 7.951, RMSE: 10.651, $R^2$ : 0.698).
ARIMA: Ranked second (MAE: 8.652, RMSE: 11.686, $R^2$ : 0.636).
Tree-Based Models: Decision Tree and Random Forest showed moderate performance, while Gradient Boosting and Temporal CNN generalized less effectively.
Baselines: Both LSTM and ARIMA improved upon the current-count forecast (MAE: 8.846) and the two-week moving average (MAE: 9.298).

Error Analysis by Incidence Level

Moderate/Low Incidence: LSTM performed best during moderate weeks (MAE: 5.483), while the current-count forecast was most accurate during low-incidence weeks.
Outbreak Levels: All methods systematically underpredicted outbreak-level weeks. For the seven outbreak weeks (mean actual: 60.143), LSTM predicted an average of 45.143, and ARIMA predicted 42.082. The four-week moving average had the lowest outbreak-class error among selected methods but remained a benchmark.

Feature Importance

Tree-based models relied heavily on recent history. The Decision Tree assigned 72.7% importance to the current case count. Random Forest and Gradient Boosting prioritized two-week rolling sums and means, with current cases playing a secondary role. Seasonal history provided context but was not the dominant signal.

Significance and Claims

The paper claims three principal contributions:

Validation of Surveillance-Only Forecasting: It demonstrates that meaningful two-week-ahead forecasts are feasible using only historical case counts and reporting dates, providing a reproducible baseline for data-constrained settings.
Importance of Temporal Rigor: The study highlights that model superiority is not absolute but dependent on the evaluation stage. ARIMA was superior in 2024, while LSTM excelled in 2025. This underscores the necessity of complete, unseen external holdouts to avoid overstating model value.
Operational Realism: The results reinforce that simple baselines (current-count forecasts and moving averages) must remain part of the evaluation framework. In 2024, complex models failed to beat these baselines, suggesting that "complexity" does not guarantee "accuracy" in short-term infectious disease forecasting.

The authors conclude that while LSTM produced the strongest metrics on the external holdout, ARIMA remains a viable, interpretable, and computationally efficient alternative. They explicitly caution against autonomous deployment, recommending that operational use combine trained forecasts with simple baselines, uncertainty communication, and human review, particularly given the systematic underprediction of outbreak peaks.