Machine learning approaches to identify thresholds in a heat-health warning system context

TL;DR

This paper introduces data-driven, model-based methods for setting heat-health warning thresholds, demonstrating improved prediction of adverse health days over traditional approaches through simulations and real-world case studies.

Contribution

It proposes novel threshold-estimation methods using regression trees, MARS, PRIM, and AIM, advancing beyond existing exposure-response based techniques.

Findings

Proposed methods outperform current thresholds in predicting adverse health days.

PRIM emerges as the most reliable method with consistent results.

Methods show promise in improving heat-health warning systems.

Abstract

During the last two decades, a number of countries or cities established heat-health warning systems in order to alert public health authorities when some heat indicator exceeds a predetermined threshold. Different methods were considered to establish thresholds all over the world, each with its own strengths and weaknesses. The common ground is that current methods are based on exposure-response function estimates that can fail in many situations. The present paper aims at proposing several data-driven methods to establish thresholds using historical data of health issues and environmental indicators. The proposed methods are model-based regression trees (MOB), multivariate adaptive regression splines (MARS), the patient rule-induction method (PRIM) and adaptive index models (AIM). These methods focus on finding relevant splits in the association between indicators and the health outcome but do it in different fashions. A simulation study and a real-world case study hereby compare the discussed methods. Results show that proposed methods are better at predicting adverse days than current thresholds and benchmark methods. The results nonetheless suggest that PRIM is overall the more reliable method with low variability of results according to the scenario or case.