APBN New Site

APBN Developing Site

Beating the Heat With Predictive Machine Learning

A hybrid machine learning method can now accurately predict the incidence of heatstroke, hospitalisations, and deaths in large cities, paving the way for better allocation of healthcare resources.

Hot days are getting hotter, and cool days growing scarce. We are no stranger to the consequences of rising global temperatures. However, in recent years, the world has seen a steady increase in heatwaves, deadly ones that have severely impacted both the livelihood and health of many populations. Despite being a natural occurrence, heatwaves are amongst the most dangerous natural hazards as they can cause a range of severe symptoms, from heat exhaustion to heat strokes, which can damage the central nervous system and even lead to death.

To prepare for, or better yet, prevent cases of heatstroke, scientists have tried to develop predictive models for heat-related illnesses. Leveraging past weather information like temperature and humidity, these systems are trained to predict the likelihood of heatstroke and stratify the risk. However, most models constructed to date have limited accuracy and have failed to account for the severity of heatstroke in their forecasts, which is an important factor needed to predict the days where cases of heatstroke are likely to spike.

Taking this gap as an opportunity for innovation, a team of scientists led by the National Cerebral and Cardiovascular Center Research Institute, Kansai University, and the National Institute for Environmental Studies have developed a new machine learning model that can accurately predict the incidence of heatstroke using only publicly available information like demographics, weather forecasts, and time of year. The team tested their model in Japanese cities and have reported successful results.

“Our AI prediction model, which can also predict spikes in the incidence of heatstroke, is based on easily obtained weather and demographic data, so it can be widely applied,” explained the first author of the study, Soshiro Ogata.

The team’s novel approach employs a hybrid model that combines the curve-fitting feature of the best performing generalised additive model (GAM) and the decision tree methods of extreme gradient boosting decision tree (XGBoost), which was trained with historical data. The GAM method makes use of smooth spline functions to fit non-linear data, like the occurrence of heatstroke over the year, while the XGBoost method can help scientists combine the output of many weak decision trees, or predictions, to create a stronger, more accurate forecast.

To validate their model, they tried to predict the total number of heatstroke cases, hospital admissions, and deaths over a 12-hour period in 16 Japanese cities with around 10 million people in total. After a series of trials and errors, the researchers were finally able to achieve highly accurate predictive values, which can help to guide public health officials to optimise medical resources in emergency medicine and public health settings. Moreover, with these accurate forecasts, scientists can also build alert systems to inform citizens of the daily risks of heatstroke and support people at high risk of heat-related illnesses to self-manage and take necessary precautions.

Besides predicting the chances of heatstroke, the tool’s GAM method also revealed several interesting predictors of high hospital admissions and deaths related to heatstroke, or in other words, features that can increase the risks of heatstroke. These factors include high temperature, a small change in maximum temperature between consecutive hot days, high solar radiation, and having a large population of people aged 65 years and older.

The project’s findings have shed light on the valuable benefits of applying artificial intelligence strategies to solve health care challenges. In future, experts hope that city-specific hybrid machine learning models, like the one introduced in the project, can play a larger role in public health administration and other fields. [APBN]

Source: Ogata et al. (2021). Heatstroke predictions by machine learning, weather information, and an all-population registry for 12-hour heatstroke alerts. Nature Communications, 12, 4575.