Correlations Between COVID-19 and Dengue

This paper presents a neural network-based correlation model demonstrating similar trends between COVID-19 and Dengue cases, which is extended into an LSTM framework to predict Dengue infections in regions with insufficient data by leveraging COVID-19 statistics and external factors.

The Gist

Imagine the world as a giant, chaotic dance floor. On this floor, two very different viruses are trying to get people to dance: Dengue (spread by mosquitoes) and COVID-19 (spread by people breathing on each other).

Usually, scientists study these two dancers separately. But this paper asks a fascinating question: "Are they dancing to the same beat?"

The authors, a team of researchers, decided to use a Digital Crystal Ball (called a Neural Network) to see if they could predict when Dengue would spike by looking at how COVID-19 was behaving. Here is the story of their discovery, broken down into simple parts.

1. The Big Idea: Two Diseases, One Rhythm?

For a long time, people thought Dengue and COVID-19 were totally unrelated. One is a tropical mosquito problem; the other is a respiratory pandemic.

However, the researchers noticed something strange in countries like Brazil, Peru, and Colombia. When COVID-19 cases went up, Dengue cases often went up too. When COVID went down, Dengue seemed to follow. It was like two dancers who, despite wearing different shoes, were accidentally stepping on each other's toes at the exact same time.

2. The Tool: The "Digital Crystal Ball" (Neural Networks)

To figure out if this was a coincidence or a real pattern, the team built a Neural Network.

• Think of it like a super-smart student. You give this student a stack of data (numbers of sick people, weather reports, holiday calendars) and say, "Figure out the pattern."
• The student looks at the data, makes a guess, sees if it's wrong, and tries again. Over and over, it learns the "rules" of the dance floor.

3. The Ingredients: What Makes the Dance Happen?

The researchers didn't just look at the virus counts. They fed the "student" three main ingredients to see what influenced the dance:

• The Weather (Temperature & Humidity): Mosquitoes love heat and rain. If it's hot and humid, Dengue usually spikes. The student learned that while weather matters, it wasn't the only thing driving the pattern.
• The Holidays: This was a big surprise. The student noticed that during big holidays (like Christmas or Carnival), the number of cases for both diseases changed.
  • Analogy: Imagine a party. When the music stops for a holiday, people might stay home (lowering COVID spread), but they might also let their guard down and let mosquitoes bite them (changing Dengue spread). The model found that holidays act like a "volume knob" for both diseases.
• The Population: They weighed the data by how many people lived in a city, so a big city like São Paulo counted more than a small town.

4. The Experiment: Teaching the Model

The team first taught their "Digital Crystal Ball" using data from Brazil, the country with the most data.

• The Result: The model got really good at predicting COVID-19 cases if it knew the Dengue numbers, the weather, and the holidays.
• The "Contraction" Effect: The model was so good at finding the trend (the ups and downs) that it sometimes smoothed out the jagged edges. It was like a weather forecast that correctly predicted "it will rain," but didn't quite predict the exact intensity of the storm.

5. The Magic Trick: Predicting the Unknown

This is where the paper gets really cool. The researchers took their trained "Digital Crystal Ball" and tried to use it on countries where Dengue data was missing or messy, like Cambodia (in Asia) and Kenya (in Africa).

• The Problem: These countries had good COVID-19 data, but their Dengue records were incomplete or hard to find.
• The Solution: They told the model, "We know the COVID-19 numbers for Kenya. We know the weather. We know the holidays. Now, use the pattern you learned in Brazil to guess what the Dengue numbers should be."
• The Outcome: The model successfully predicted the peaks and valleys of Dengue in Kenya and Cambodia, even without having the actual Dengue numbers to start with! It was like guessing the plot of a movie sequel just by watching the first movie and knowing the director's style.

6. The "Time Machine" (LSTM)

The researchers also tried a more advanced version of their model called LSTM (Long Short-Term Memory).

• Analogy: A standard model is like looking at a photo and guessing what happens next. An LSTM is like watching a video. It remembers what happened yesterday, last week, and last month to understand the flow of time.
• This helped them predict future trends even better, accounting for the fact that diseases don't just happen in a vacuum; they have a history.

7. The Takeaway: Why Does This Matter?

The paper concludes that these two diseases are surprisingly linked, likely because they are both affected by the same human behaviors (holidays, lockdowns) and environmental factors (heat, rain).

Why should you care?

• For Health Officials: If you see a huge spike in COVID-19, you might want to prepare for a Dengue spike too, even if you don't have perfect mosquito data yet.
• For the Future: This "Digital Crystal Ball" can help countries that don't have enough money or staff to track every mosquito bite. They can use the data they do have (like COVID stats) to guess where the next Dengue outbreak will hit, allowing them to send medicine and mosquito nets to the right place before the disaster happens.

In short: The researchers built a smart computer that learned that when the world dances to the rhythm of a pandemic, the mosquitoes often dance right along with it. By understanding that rhythm, we can be better prepared for the next step.

Technical Summary

1. Problem Statement

The paper addresses the lack of research regarding the correlation between Dengue fever and COVID-19, two significant viral diseases that often co-circulate, particularly in tropical and sub-tropical regions (e.g., Latin America, Southeast Asia, Africa).

• Context: While both diseases have been studied individually, their simultaneous transmission (syndemic) poses a challenge to public health. There is conflicting evidence regarding cross-immunity; some studies suggest it, while others indicate prior Dengue infection may increase COVID-19 risk.
• Data Scarcity: Many countries, particularly in Africa and Southeast Asia, lack sufficient or reliable Dengue surveillance data, making it difficult to predict outbreaks or allocate resources effectively.
• Objective: The authors aim to develop machine learning models to:
  1. Quantify the correlation between COVID-19 and Dengue case trends.
  2. Incorporate external variables (climate, social behavior) to improve predictive accuracy.
  3. Utilize robust COVID-19 data to predict Dengue outbreaks in regions with poor Dengue reporting.

2. Methodology

The authors employ a multi-stage approach using Neural Networks (NN) and Long Short-Term Memory (LSTM) models.

A. Data Sources & Preprocessing

• Sources: Pan-American Health Organization (PAHO) for Dengue; World Health Organization (WHO) for COVID-19; Visual Crossing for climate data (temperature, humidity, rainfall).
• Scope: Epidemiological weeks 30–100 (Jan 2020 – early 2022).
• Normalization: Due to the magnitude difference between COVID-19 and Dengue cases, the authors use base-10 logarithms of case counts. For cross-country predictions, they normalize by population (cases per million).
• External Variables:
  • Holidays: Binary indicators (1/0) for weeks containing national holidays.
  • Climate: Weighted averages of temperature and humidity based on population density (e.g., using Salvador and São Paulo for Brazil; Lima for Peru; Bogotá for Colombia).

B. Model Architecture

1. Correlation Model (Standard Neural Network):

   • Structure: 1 Input Layer, 4 Hidden Layers, 1 Output Layer.
   • Activation: ReLU for hidden layers; Linear for output.
   • Loss Function: Mean Absolute Error (MAE).
   • Training: 150 epochs with cross-validation.
   • Inputs: Dengue cases, COVID-19 cases, Holiday status, Temperature, Humidity.

2. LSTM Model (Time Series Analysis):

   • Structure: Recurrent Neural Network with Input, Hidden, and Output layers, utilizing Input, Output, and Forget gates to handle temporal dependencies.
   • Purpose: To capture time-lags and sequential patterns that standard NNs miss.
   • Scaling: MinMaxScaler (0–1) applied before training, then inverse-scaled for interpretation.

3. Cross-Regional Prediction Strategy:

   • Models trained on South American data (Brazil, Peru, Colombia) are applied to test datasets from Cambodia and Kenya to predict Dengue trends using only available COVID-19 and climate data.

3. Key Contributions

• Novel Correlation Model: The paper defines a specific "Correlation Model" that successfully integrates biological data (infection counts) with socio-environmental factors (holidays, climate) to model the relationship between two distinct viral epidemics.
• Data Imputation via Transfer Learning: The authors demonstrate a method to estimate Dengue infections in data-scarce regions (Cambodia, Kenya) by leveraging the strong temporal correlation with COVID-19 data from South America.
• Variable Importance Analysis: The study systematically isolates the impact of holidays versus climate on disease transmission, providing insights into which factors drive variance in the models.
• Peak Index Definition: The authors propose a theoretical framework for "Peak Indices" (Relative and Absolute) to standardize the identification of epidemic peaks across different diseases and populations.

4. Results

• Correlation Trends:
  • Brazil: A strong positive correlation was observed between COVID-19 and Dengue peaks (specifically between epidemiological weeks 60–80 in 2021).
  • Peru & Colombia: Trends were present but less uniform; Peru showed random correlation, while Colombia showed a slight negative correlation, suggesting country-specific dynamics.
• Model Performance (Brazil):
  • Holidays Only: The model captured the general trend but underestimated the amplitude (variance) of the actual data.
  • Climate Only: The loss curve decreased more gradually, indicating climate factors alone are less predictive than holidays for COVID-19 trends in this context.
  • Combined (Holidays + Climate): The combined model produced the most accurate trend alignment. However, a systematic bias was noted: predictions were shifted down by approximately 0.2 on the logarithmic scale compared to actual data.
• Cross-Regional Predictions:
  • The model successfully predicted Dengue peaks in Cambodia and Kenya using South American training data.
  • Predicted peaks in Kenya (weeks 60 and 110) and valleys (weeks 50 and 100) aligned with expected seasonal patterns, validating the utility of using COVID-19 as a proxy for Dengue in data-poor regions.
• LSTM Performance:
  • The LSTM model provided a more comprehensive view of future values, accurately fitting the general trends of both diseases. It successfully predicted a rise and subsequent fall in Dengue cases post-2022.

5. Significance and Future Directions

• Public Health Policy: The models offer a tool for health policymakers to anticipate Dengue outbreaks in regions with weak surveillance systems by monitoring COVID-19 trends and climate variables.
• Syndemic Management: Understanding the simultaneous transmission of these diseases helps in resource allocation, especially given that lockdowns may have inadvertently increased Dengue transmission by disrupting vector control.
• Future Work:
  • Latent Heat Flux (LHF): The authors suggest LHF (energy transfer via evaporation/condensation) might be a superior climate predictor compared to simple temperature/humidity, though data availability is currently limited.
  • Peak Indices: Formalizing peak detection algorithms could improve correlation studies for other viral epidemics.
  • Lockdown Impact: Further investigation is needed on how public health interventions (lockdowns) specifically alter mosquito breeding cycles and Dengue transmission rates.

In conclusion, the paper establishes that while COVID-19 and Dengue are biologically distinct, their epidemiological curves exhibit significant temporal correlations driven by shared external factors. Machine learning models can effectively leverage these correlations to fill data gaps and improve global disease surveillance.