Researchers in Japan are using artificial intelligence to track heat stroke risks in real-time, according to a study published in Nature. By leveraging transformer-based pre-trained language models, they analyzed Japanese tweets to distinguish true heat stroke incidents from false ones. Their findings suggest that social media could play a crucial role in early public health surveillance, helping authorities respond more quickly to extreme heat events.
Image Credit: Leestudio/Shutterstock.com
Background
As global temperatures rise, heat stroke is becoming a growing concern in Japan, particularly for vulnerable populations. Traditional surveillance systems rely on emergency data, often leading to delayed responses. In contrast, event-based surveillance, which leverages social media, has shown promise in detecting public health risks early.
Previous studies have used machine learning (ML) and deep learning (DL) to classify morbidity-related tweets in English, but research on Japanese tweets—especially those concerning heat stroke—remains limited due to language complexity.
Earlier ML models demonstrated moderate accuracy, but recent advancements in transformer-based models, such as BERT and RoBERTa, have significantly improved disease-related tweet classification. This study addresses a crucial gap by evaluating ML and transformer-based DL models for classifying Japanese tweets about heat stroke, underscoring their potential for event-based surveillance and timely public health interventions.
Heat Stroke Detection with Transformers
To explore the feasibility of AI-driven event-based surveillance, researchers evaluated transformer-based Japanese language models for detecting heat stroke-related tweets. The study focused on Nagoya City, Aichi Prefecture, a region where rapid urbanization has led to rising temperatures and increased heat stroke cases.
Using data from the Nagoya City Fire Department (2017–2022), the team collected 27,040 tweets containing the Japanese word for “hot” via Twitter API v2. A time-series analysis revealed a correlation between tweet frequency and heat stroke cases. Text preprocessing included removing noise, normalizing characters, and geocoding location data.
A native Japanese-speaking public health specialist manually annotated the dataset, which was then split into training (60 %), validation (20 %), and test (20 %) sets. Researchers fine-tuned transformer-based models—BERT-base, RoBERTa-base, and LUKE Japanese base lite—on the dataset, using a Support Vector Machine (SVM) as a baseline. Model performance was evaluated using accuracy, precision, recall, and F1-score, with confusion matrices and heatmaps providing additional insight.
The study further assessed the feasibility of an event-based heat stroke surveillance system by mapping and animating classified tweets alongside emergency cases. The findings indicate that transformer-based models can effectively identify heat stroke-related events, reinforcing the potential of real-time public health monitoring and crisis response.
Results and Discussion
The study compared transformer-based models—BERT-base, RoBERTa-base, and LUKE Japanese base lite—against SVM for classifying Japanese tweets about heat stroke. LUKE Japanese base lite outperformed all models, achieving the highest accuracy (85.52 %) and F1-score (85.23 %), while SVM had the lowest performance (accuracy: 72.73 %). The transformer models significantly outperformed SVM, confirming their effectiveness in text classification.
Confusion matrix heatmaps showed that LUKE correctly classified 88 % of true cases, while BERT and RoBERTa both achieved 86 %, compared to SVM’s 69 %. Spatiotemporal visualizations of emergency transport data and classified tweets revealed a correlation, though the datasets did not fully overlap.
While the study highlights the potential of transformer models for event-based public health surveillance, challenges remain, including dataset bias and limited geographical coverage. Since data collection relied on tweets containing the keyword “hot,” selection bias may have been introduced. Expanding keyword selection and including tweets from all Japanese prefectures could improve model accuracy and generalizability.
Future research will focus on enhancing spatiotemporal analysis and scaling the heat stroke detection system from Aichi Prefecture to a nationwide early warning system. Additionally, the methodology can be adapted to monitor other public health risks, including emerging infectious diseases.
Conclusion
This study demonstrates the effectiveness of transformer-based models in event-based public health surveillance, particularly for detecting heat stroke risks through Japanese tweets. LUKE Japanese base lite achieved the highest accuracy, outperforming other models. Despite challenges such as dataset bias and limited geographical coverage, findings highlight the potential of deep learning for real-time public health monitoring.
Future work will expand data collection, refine spatiotemporal analysis, and scale the system for nationwide implementation. The methodology can also be adapted to track other health risks, contributing to proactive crisis response and early intervention strategies.
Journal Reference
Anno, S., Kimura, Y., & Sugita, S. (2025). Using transformer-based models and social media posts for heat stroke detection. Scientific Reports, 15(1). DOI:10.1038/s41598-024-84992-y. https://www.nature.com/articles/s41598-024-84992-y
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.