Feature Signature Discovery for Autism Detection: An Automated Machine Learning Based Feature Ranking Framework

Table 3

Comparative training estimates of AutoML models on ASD datasets.

Dataset	Data analysis technique	Feature selection method	Classifier	Training metrics (%)		#Features
				Toddler	Built-in		None	SVM, cost = 1; γ = 1; polynomial kernel	97.5	98.4	All
Feature selection (FS)	SES; α = 0.05, Maxk = 2	SVM, cost = 1; γ = 1; polynomial kernel	99.2		99.5	9
Aggressive FS	SES; α = 0.05, Maxk = 2	SVM, linear kernel; cost = 1	90.8		97	9
Child	Built-in	None	Ridge logistic regression; λ = 1	94.6	97.6	All
	Feature selection (FS)	SES; α = 0.05, Maxk = 2	SVM, linear kernel; cost = 1	100	100	10
	Aggressive FS	Both feature selection and aggressive feature selection yielded the same signature
Adolescent	Built-in	None	Ridge logistic regression; λ = 1	81.0	94.6	All
	Feature selection (FS)	LASSO, penalty = 1	Ridge logistic regression; λ = 1	75.5	91.3	12
	Aggressive FS	SES; α = 0.05, Maxk = 2	SVM, radial basis function kernel; cost = 1; γ = 1	69.6	86.4	6
Adult	Built-in	None	SVM, radial basis function kernel; cost = 1; γ = 1	96.0	99.2	All
	Feature selection (FS)	SES; α = 0.05, Maxk = 2	SVM, linear kernel; cost = 1	91.7	96.5	11
	Aggressive FS	Both feature selection and aggressive feature selection yielded the same signature