Researchers Use Social Media for Early Detection of Heat Stroke

Japanese researchers demonstrate the potential of combining social media posts and deep learning models for early detection of heat stroke risks. This approach opens new possibilities for leveraging real-time data in event-based surveillance, enabling timely detection and response to heat stroke threats. 

While previous studies have highlighted the potential of social media posts, such as tweets, to offer real-time insights into various events, its application in detecting heat stroke risks had not been explored.

A team of researchers, led by Professor Sumiko Anno from the Graduate School of Global Environmental Studies, Sophia University, Japan, along with Dr. Yoshitsugu Kimura, Yanagi Pearls, Japan, and Dr. Satoru Sugita, Chubu University, Japan, combined social media posts and transformer-based learning models to detect heat stroke risks in Nagoya City, Japan. 

The researchers utilized transformer-based deep learning models, including BERT, RoBERTa, and LUKE Japanese base lite, along with a machine learning model (support vector machine or SVM) to identify tweets containing the word “hot” in Japanese. The team successfully collected about 27,040 tweets over five years using the Twitter API. 

By preprocessing the text data and applying advanced deep and machine learning techniques, the models were trained and fine-tuned to identify tweets related to heat stroke events. These models were evaluated using key performance metrics such as accuracy, precision, recall, and F1-score.      

The findings were published in Scientific Reports on January 4, 2025. 

Anno explained, “By leveraging social media posts, we can enhance public health surveillance systems and facilitate the early detection of heat stroke risks. Our findings emphasize the importance of real-time data monitoring to combat the health challenges posed by climate change.”

The research highlighted the potential of combining Japanese tweets and transformer-based pre-trained language models for public health surveillance. Among the models tested, LUKE Japanese base lite achieved the highest performance metrics with an accuracy of 85.52%, followed by BERT-base (84.04%) and RoBERTa-base (83.88%). Whereas the SVM baseline model showed the lowest performance, with an accuracy of 72.73%.

The use of time-space visualizations and animated video showcased the potential for real-time event-based surveillance. Through mapping the locations of heat stroke-related emergency medical evacuations and matching them with geo-tagged tweets, the study demonstrated how social media data could provide an early warning system for heat stroke risks in urban environments. Time-space visualizations demonstrated how social media could be integrated with emergency response data to serve as an effective early detection tool for extreme weather events.

This research opens the door to future applications of deep learning and social media posts for real-time health monitoring systems. Looking ahead, the team plans to establish an early warning system for heat stroke in Aichi Prefecture, to eventually expand this system to a nationwide alert system for Japan. Achieving this will involve collaborations with local authorities to collect heat stroke data and conduct spatiotemporal analyses across all prefectures. 

“Our methodology can be extended and adapted for monitoring emerging and reemerging infectious diseases, broadening its application in public health surveillance,” concluded Anno.