How can we better predict heat-related deaths where extreme heat is unprecedented?
Researchers led by SoGE’s Dr Sarah Wilson Kemsley have demonstrated an improved approach to predicting heat-related mortality in Europe, using machine learning.
As climate change pushes Europe into unprecedented heat extremes, predicting heat-related mortality is crucial. But standard epidemiological models* were not originally designed for prediction and are largely calibrated on previous local temperature observations - so they can be weak in regions where extreme heat risk is new.
To meet this challenge the team trained a neural network on all regions of Europe. The neural network predicted out-of-sample mortality across Europe more accurately on average than standard epidemiological modelling, and learned the relationship between temperature and mortality directly from data (without requiring assumptions to be built in).
The neural network was also more confident in its predictions of heat-attributable deaths during the most extreme heat events. It most strongly outperformed traditional epidemiological models in historically cooler regions of Europe, and in areas that experienced temperatures beyond their local historical records.
“We found that a joint-learning framework performed well, particularly in regions breaking local temperature records. This makes a strong case for flexible, data-driven approaches in future heat-health warning systems”
The joint-learning approach was able to borrow information from hotter regions to improve predictions where extreme heat is rare, and can more readily include stressors like air pollution and humidity which can worsen heat-related mortality. These capabilities are becoming increasingly important as climate change pushes temperatures into locally unprecedented conditions.
The full open-access paper can be read here.
Authors: Dr Sarah Wilson Kemsley (Oxford, SoGE) Jowan Fromentin (Oxford); Bikem Pastine (SoGE, Oxford), Prof. Xiaowen Dong (Oxford), Dr. Tom Matthews (KCL), Dr Pierre Masselot (LSHTM), Prof. Katrin Meissner (UNSW Sydney), Prof. Sarah Perkins-Kirkpatrick (ANU) and Prof. Louise Slater (SoGE, Oxford).
*Distributed Lag Non-linear Models (DLNMs)
How can we better predict heat-related deaths where extreme heat is unprecedented?
Researchers led by SoGE’s Dr Sarah Wilson Kemsley have demonstrated an improved approach to predicting heat-related mortality in Europe, using machine learning.
