Table A.5.

Search space for hyperparameters by technique and the best configuration obtained for predicting AvTT per hour for each hospital.

Technique	Search space	Best configuration
		CHRU	CHIH	HNFC
LGBM	max_depth: [1–12]	6	12	5
	n_estimators: [50–1000]	143	808	768
	num_leaves: [31–100]	40	61	45
	learning_rate: [0.001–1]	0.8498	0.0103	0.0094
	subsample: [0.5–1]	0.99	0.92	0.5
	colsample_bytree: [0.5–1]	0.94	0.6	0.63
MLP	Dense layers: 2	2	2	2
	nb_neurons: [100–500]	(100, 302)	(125, 325)	(362, 562)
	alpha: [0.00001–0.01]	0.000029	0.003674	0.000034
	learning_rate_init: [0.0001–0.1]	0.0355	0.003611	0.000543
	max_iter: [50–200]	154	86	177
	tol: [0.00001–0.01]	0.003731	0.002888	0.000661
	momentum: [0.00001–0.01]	0.0073	0.006057	0.00479
	Early stopping: 20	20	20	20
LSTM	LSTM layers and neurons: 1, (110)	1, (110)	1, (110)	1, (110)
	Dense layers and neurons: 2, (128, 1)	2, (128, 1)	2, (128, 1)	2, (128, 1)
	Activation function: ReLU	ReLU	ReLU	ReLU
	Dropout: 0.5	0.5	0.5	0.5
	Loss function: ‘mse’	‘mse’	‘mse’	‘mse’
	Optimizer: Adam	Adam	Adam	Adam
	Early stopping: 15	15	15	15
	Max. epochs: 100	100	100	100
	Batch size: [40–250]	93	142	112
	Learning rate: [0.005–0.01]	0.00841	0.00595	0.00894
Prophet	n_changepoints: [20–100]	45	75	35
	seasonality_prior_scale: [0–50]	36.82	18.17	12.13
	holidays_prior_scale: [0–50]	23.47	22.42	34.77