# LSTM‐Based Recurrent Neural Network Predicts Influenza‐Like‐Illness in Variable Climate Zones

## TL;DR

This study uses an LSTM neural network to predict flu trends in different climate zones, finding that seasonal patterns, not specific climate variables, mainly drive flu activity.

## Contribution

The study introduces a cross-regional LSTM model to predict influenza-like illness trends, revealing the role of seasonal patterns over absolute climate variables.

## Key findings

- All three regions showed strong seasonality in flu trends, with Hawaii having the highest ILI values.
- Climate variables showed weak to moderate correlations with ILI, with temperature showing the strongest negative correlation.
- Cross-regional LSTM models performed comparably, suggesting seasonal patterns drive ILI trends more than specific climate variables.

## Abstract

Influenza virus is responsible for a recurrent, yearly epidemic in most temperate regions of the world. Flu has been responsible for a high disease burden in recent years, despite the confounding presence of SARS‐CoV‐2. However, the mechanisms behind seasonal variance in flu burden are not well understood. This study seeks to expand understanding of the impact of variable climate regions on seasonal flu trends. To that end, three climate regions have been selected. Each region represents a different ecological zone and provides different weather patterns.

A long short‐term memory (LSTM)‐based recurrent neural network was used to predict influenza‐like‐illness trends for three separate locations: Hawaii, Vermont, and Nevada. Flu data were gathered from the Center for Disease Control as weekly influenza‐like‐illness (ILI) percentages. Weather data were collected from Visual Crossing and included temperature, wind speed, UV index, solar radiation, precipitation, and humidity. Data were prepared and the model was trained as described previously.

All three regions showed strong seasonality of flu trends with Hawaii having the largest absolute ILI values. Temperature showed a moderate negative correlation with ILI in all three regions (Vermont = −0.54, Nevada = −0.56, Hawaii = −0.44). Humidity was moderately correlated in Nevada (0.47) and weakly correlated with ILI in Hawaii (0.22). Vermont ILI did not correlate with humidity. Precipitation and wind speed were weakly correlated in all three regions. Solar radiation and UV index showed moderate correlation in Vermont (−0.33, −0.36) and Nevada (−0.53, −0.55), but only a weak correlation in Hawaii (−0.15, −0.18). When trained on the complete data sets, baseline model performances for all three datasets at +1 week were equivalent. Models trained on one region and used to predict cross‐regional data performed uniformly and equivalent to baseline.

Results indicate that climate variables were weak to moderate predictors in all regions. Initial modeling attempts revealed uniform performance in all regions. Despite strong climate differences, cross‐regional LSTM models performed comparably, suggesting that seasonal patterns, rather than absolute climate variables, drive ILI trends. Additionally, this data suggests that absolute climate changes may not be influential as relative seasonal changes.