Sparsifying machine learning models identify stable subsets of predictive features for behavioral detection of autism

Table 5 Summary of accuracies for modules 2 and 3 with best classifier, best parameters, and different feature sets

Module	3	3	3	2	2	2
Number of features	10	10	5	5	5	5
Best classifier	L2 LR	L1 Lin SVM	L2 LR	LDA	L1 Lin SVM	L2 LR
Optimal parameters	C = 1	C = 0.5	C = 10	S = 0.8	C = 0.5	C = 0.05
Area under ROC	0.95	0.95	0.93	0.93	0.93	0.92
Precision	0.99	0.99	0.99	0.98	0.98	0.98
Recall/sensitivity	0.90	0.95	0.88	0.97	0.98	0.93
Specificity	0.89	0.87	0.89	0.50	0.58	0.67
Balanced accuracy	0.90	0.90	0.88	0.74	0.78	0.80
F1 score	0.94	0.97	0.93	0.97	0.98	0.95

LR denotes logistic regression, L1 Lin SVM denotes L ₁ penalized linear SVM, and S denotes the LDA shrinkage parameter

ISSN: 2040-2392