Table 4.

Hyperparameter ranges for each machine learning model. Abbreviation definitions (in order of appearance from left to right): ML machine learning, learn rate learning rate, subsamp. subsampling per tree, cols/tree columns per tree, max iter. maximum iterations, SVM support vector machine, SVM-Gaussian SVM with Gaussian radial basis function kernel, SVM-Linear SVM with linear kernel, Log. Regression logistic regression, lasso lasso regression with L1 penalization, ridge ridge regression with L2 regularization, DFNN deep feedforward neural network, LSTM bidirectional long short term memory neural network, BrainNet CNN BrainNet convolutional network⁵², ReLU leaky slope rectified linear activation unit slope for x < 0.

	Naïve Bayes	Random forest	Extremely random trees	Adaptive boosting	Gradient boosting	SVM-Gaussian
Nonlinear classical ML	NA	estimators [50,5e3] max nodes [5,50]	estimators [50,5e3] max nodes [5,50]	estimators [50,5e3] learn rate [0.1, 0.9]	estimators [50,5e3] learning rate [5,50] max depth [1,10] subsamp. [0.2, 0.8] cols/tree [0.2, 1]	C [1e−4, 1e5] max iter. [1e4, 1e5] gamma [1e−2, 1e2]

	SVM-linear	Log. regression (Lasso)	Log. regression (Ridge)
Linear classical ML	C [1e−4, 1e5] max iter. [1e4, 1e5]	C [1e−4, 1e4] max iter. [1e4, 1e5]	C [1e−4, 1e4] max iter. [1e4, 1e5]

	DFNN	LSTM	BrainNet CNN
Deep learning	hidden layers [1,3] initial width [16, 256] dropout fraction [0.1, 0.6] L2 penalty [1e−4, 2e−2]	hidden layers [1,3] initial width [16, 256] dropout fraction [0.1, 0.6] L2 penalty [1e−4, 2e−2]	hidden layers [0, 2] initial width [16,64] dropout fraction [0.1, 0.6] ReLU leaky slope [0.1, 0.5]