Table 2.

Prediction performances of prediction models for the number of all heatstrokes among 6 models^a.

	GLM using WBGT only	GLM	GAM	RF	XGBoost	Consolidation of 16 GAMs specific to each cityb
	The number of all heatstrokes
Overall predictive accuracies per city per 12 h
RMSE in training	1.73	1.41	1.37	0.73	1.09	1.27
RMSE in testing	3.73	2.92	2.47	3.51	3.28	5.25
Predictive accuracies on days when the number of heatstrokes spikedc
MAPE per 1-day (%) in training	22.6	18.4	18.0	8.3	11.9	16.8
MAPE per 1-day (%) in testing	43.0	27.1	19.7	32.0	28.5	19.0
Total absolute percentage error (%) in training	18.8	7.3	8.3	5.5	5.9	7.5
Total absolute percentage error (%) in testing	48.8	30.5	21.9	37.2	31.9	19.7

GLM generalized linear model, GAM generalized additive model, RF random forest, XGBoost extreme gradient boosting decision tree, WBGT wet bulb globe temperature, RMSE root-mean-square error, MAPE mean absolute percentage error.

^aSmaller RMSE, MAPE, and total absolute percentage error show better predictabilities.

^bPrediction models specific to each of 16 cities were developed for city-specific prediction.

^cMAPE and total absolute percentage error were calculated after observed and predicted values were summed up per day (for MAPE) per the entire period (for total absolute percentage error) on days when the number of all heatstrokes was 80th percentile (corresponding to 53.6 in 2015, 57.8 in 2016, 60.6 in 2017, and 89.8 in 2018) and over in each year. MAPE is a mean value of absolute errors divided by observed values.