Forecast-to-warning for Injury Surveillance in Megacity based on Seasonal and Holiday Effects: A District-level Study

Imagine you are trying to predict the weather for your neighborhood. You know that rain often comes in seasons, and you know that storms might hit harder after a big holiday when everyone is out and about. In the world of public health, scientists do something similar but with injuries instead of rain. They keep a daily log of how many people get hurt and rush to the hospital. This is called surveillance. Usually, these logs are just used to look back at the past and say, "Oh, more people fell off bikes in July." But what if we could look forward? What if we could say, "Hey, next Tuesday might be a dangerous day, so let's get extra doctors ready"? That is the goal of forecasting. To do this, scientists use math models that act like crystal balls, trying to find patterns in the chaos of daily life, such as the rhythm of the week or the special chaos of public holidays. The big question is: Can we turn these math predictions into a simple, colorful warning system—like a traffic light—that tells hospitals when to be on high alert?

This paper is a story about trying to build that traffic light system for a massive, super-fast city called Shenzhen in China. The researchers looked at four different hospitals in different neighborhoods and asked: "Can we predict injury spikes before they happen, especially around big holidays like the Spring Festival?" They tested several different "crystal balls" (math models) to see which one worked best. They found that while no single model was perfect for every hospital, a tool called Prophet was particularly good at guessing the direction of change—telling them if injuries were about to go up or down. They also discovered that holidays are a huge deal; for example, during the Spring Festival, injury numbers actually dropped at some hospitals because people went home to their families, but during Labor Day, they went up in other areas because workers were out having fun.

However, the most interesting part of the story is the "traffic light" itself. The researchers tried to turn their math predictions into three colors: Green (normal), Yellow (caution), and Red (danger). Here is the twist: just because the math was good at guessing the exact number of injuries didn't mean it was good at guessing the right color. The system was pretty good at saying "Green" when things were calm, but it struggled to catch the "Yellow" and "Red" days. It often missed the high-risk days or got confused about when the danger started. The paper suggests that while we can predict injuries, turning those numbers into a reliable warning system is tricky. It's like having a weather app that knows it's going to rain but can't tell you exactly when the umbrella is needed. The authors conclude that to make this work in real life, every hospital needs its own custom settings, because what works for one neighborhood doesn't necessarily work for the next.

The Story of the Injury Crystal Ball

The Setup: A City of Millions and a Mountain of Data
Shenzhen is a megacity, a place where millions of people move around, work, and play every day. Because it's so big and busy, injuries happen all the time. The city has a network of "sentinel hospitals" (think of them as the city's main injury report cards) that count every injury that comes through the emergency room. The researchers grabbed the daily injury counts from four of these hospitals: Bao'an, Guangming, Nanshan, and Longgang. They wanted to see if they could predict the future injury count for the next week.

The Tools: Five Different Crystal Balls
To make their predictions, the team didn't just guess; they used five different mathematical "crystal balls" (models) and compared them to see which one was the sharpest:

1. ARIMA & SARIMA: These are the old-school, classic models. They are great at finding patterns that repeat every week or every year, like a clock.
2. SARIMAX: This is the classic model but with an extra feature: it can look at outside factors, like holidays.
3. LSTM: This is a fancy, modern "neural network" (a type of artificial intelligence) that tries to learn complex patterns by reading through tons of past data, kind of like a student who memorizes every textbook.
4. Prophet: This is a newer tool designed to be flexible. It breaks the data down into a trend (is it going up or down?), seasonality (does it happen every week?), and holidays. It's like a chef who knows exactly how much salt to add based on the day of the week and the season.

The Experiment: Rolling the Dice
The researchers didn't just test these models once. They used a "rolling window" method. Imagine you are predicting the weather for next week. You use data from the past to make a guess, then you wait a week, see what actually happened, add that new data to your history, and make a new guess for the following week. They did this over and over again for each hospital.

The Findings: One Size Does Not Fit All
Here is where the story gets interesting. The researchers found that no single model was the winner for everyone.

• For Bao'an and Guangming hospitals, Prophet was the champion. It made the most accurate guesses about the numbers and was the best at telling if the injury count was going up or down.
• For Nanshan hospital, the LSTM (the AI model) was the best at guessing the exact numbers.
• For Longgang hospital, the old-school ARIMA model actually did the best job with the numbers.

This taught them a big lesson: You can't just pick one "best" model for the whole city. Each hospital is like a different neighborhood with its own personality, and the model needs to be tuned to fit that specific place.

The Holiday Effect: When the Calendar Changes the Game
The researchers also wanted to know: Do holidays mess up the predictions? In China, holidays like the Spring Festival (Chinese New Year) and National Day are huge. People travel, parties happen, and routines change.

• They found that Spring Festival was a big deal. At almost all the hospitals, injury numbers dropped during this time. It seems like when everyone goes home to their families, the city quiets down, and fewer people get hurt in the city.
• Labor Day was different. In Longgang, injury numbers went up. The researchers guessed this might be because workers in that area were out having fun, traveling, or doing outdoor activities instead of being at their factories.
• They tried different "windows" to see how long the holiday effect lasted. They tested if the effect lasted 1 day, 3 days, 5 days, or 7 days before and after the holiday. They found that a 5-day window (2 days before, the holiday, and 2 days after) was usually the sweet spot, but it varied by hospital.

The Traffic Light: Green, Yellow, Red
The ultimate goal wasn't just to guess a number like "42 injuries tomorrow." They wanted to turn that number into a simple warning:

• Green: Everything is normal.
• Yellow: Be careful, things might get busy.
• Red: Danger! Get ready for a surge.

They set up rules to turn the predicted numbers into these colors. For example, if the predicted number was higher than the average plus a little bit, it turned Yellow. If it was way higher, it turned Red.

The Problem: The Light Was Flickering
Here is the tricky part. Even though the models were okay at guessing the numbers, they were not great at guessing the colors.

• Green worked well. The system was good at saying "It's a normal day."
• Yellow was a mess. The system struggled to tell the difference between a normal day and a slightly busy day. It missed a lot of Yellow days.
• Red was also hit-or-miss. In some hospitals, the system caught the Red days, but in others (like Nanshan), it missed them almost all the time.

The researchers realized that being "close" to the right number isn't the same as getting the right warning. A model can be slightly off and still get the color wrong, which is a big problem if you are trying to save lives.

The Conclusion: Customize or Fail
The paper ends with a clear message: We can predict injuries, and we can use holidays to help us, but we can't use a "one-size-fits-all" approach.

• Don't use the same model for every hospital. Some need AI, some need old-school math, and some need the flexible Prophet tool.
• Don't assume holidays affect everyone the same way. The Spring Festival might calm things down in one area but spike them in another.
• The warning system needs work. Turning numbers into a simple Red/Yellow/Green light is harder than it looks. The system needs to be tested and tuned specifically for each hospital before it can be trusted to tell them when to call in extra staff.

In short, the researchers built a cool new tool for the city, but they learned that to make it work, you have to treat every hospital like a unique individual, not just a number in a spreadsheet.