Table 1.

Selected model performance validation based on holdout.

Disease	Years	Compared Models	Selected Model	Gamma Deviance	RMSE	R-Squared	MAE	MAPE
Ischemic heart disease	DW = 6 FD = 24	25	AVG Blender	0.0110	65.6272	0.9725	48.5973	8.3663
Stroke	DW = 6 FD = 24	29	eXtreme Gradient Boosting on ElasticNet Predictions	0.0140	82.1981	0.8000	57.5654	8.8361
Chronic obstructive pulmonary disease	DW = 12 FD = 24	35	AVG Blender	0.0117	79.8697	0.9126	61.3592	8.7560
Lower respiratory infections	DW = 8 FD = 24	25	AVG Blender	0.0108	192.8462	0.9045	127.5683	7.4239
Alzheimer’s disease	DW = 8 FD = 24	28	AVG Blender	0.0579	21.7291	0.8688	16.0741	19.9726
Lung cancer	DW = 12 FD = 24	35	eXtreme Gradient Boosted Trees Regressor with Early Stopping (Gamma Loss)	0.0115	32.7293	0.9372	24.7214	8.5870
Diabetes mellitus	DW = 6 FD = 24	26	AVG Blender	0.0053	50.9413	0.9499	37.2955	5.5817
Road injuries	DW = 6 FD = 24	25	Elastic-Net Regressor (L2/Gamma Deviance) with Forecast Distance Modeling	0.0338	25.7580	0.8410	19.6105	15.1879
Diarrheal Disease	DW = 10 FD = 24	25	AVG Blender	0.0175	108.7063	0.8274	74.4832	10.7970
Tuberculosis	DW = 12 FD = 24	41	eXtreme Gradient Boosted Trees Regressor with Early Stopping	0.0674	54.8689	0.7771	36.5015	21.7094

Note. After choosing the length of training data for the backtests, Derivation Window (DW), and the length of forecasted data (FD), models were compared and validated for each disease by the AML (automated machine learning) platform. The year 2018 was chosen as holdout, and the predicted values were compared to the actual values. Model selection was based on the Gamma Deviance or root mean square error (RMSE). Other calculated estimators were R-squared, the mean absolute error (MAE) and the mean absolute percentage error (MAPE). The total number of compared models, as well as the final selected model, are listed. These final selected models were either the AVG (average) Blender, the eXtreme Gradient Boosting model or the Elastic-Net Regressor.