Utilizing machine learning for predicting mortality in patients with heat-related illness who visited the emergency department.

BACKGROUND

In the context of climate change and global warming, heat-related illness (HRI) is anticipated to escalate and become a major concern. Patients with severe HRI primarily present to the emergency department (ED), but there are no prediction tools for mortality in HRI patients who visit ED. The objective of this study was to use machine learning approaches to establish prediction models for mortality in patients with HRI who visit ED.

METHODS

We included all patients aged 20 and above with a final diagnosis of HRI who visited the EDs of three hospitals (Chi Mei Medical Center, Chi Mei Hospital Liouying, and Chi Mei Hospital Chiali) between January 2010 and October 2021. Patients who had transferred to other hospitals or had insufficient data were excluded. A total of 11 predictive feature variables were used in the algorithms. The primary outcome was in-hospital mortality or impending death discharge. We used machine learning algorithms including Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), Light Gradient Boosting Machine (LightGBM), Multilayer Perceptron (MLP), and eXtreme Gradient Boosting (XGBoost) to establish prediction models for mortality in such patients. Accuracy, sensitivity, specificity, and area under curve (AUC) were used as indicators to evaluate the performance of prediction models.

RESULTS

Out of the 820 HRI patients included in the analysis, 1.5% had mortality. All six prediction models had a high AUC, ranging from 0.825 to 0.991, and LightGBM which included peripheral oxygen saturation (SpO and Glasgow Coma Scale (GCS) score on arrival as the two main features had the highest AUC. The accuracy, sensitivity, and specificity of LightGBM were 0.976, 1.000 and 0.975, respectively.

CONCLUSION

Machine learning-based prediction models are promising tools in accurately predicting mortality in HRI patients who present to the ED.