Table 2.

Performance of the LDA classifier with shrinkage regularization trained using only fixations on button as target fixations (Trainset 1).

Sbj.	Test on fixations with 500 ms threshold					Test on fixations with 1000 ms threshold
	Specificity	Sensitivity			AUC	Sensitivity
		Button	Ball	Free cell	Button	Button	Ball	Free cell
1	0.87 ± 0.13	0.23 ± 0.17	0.05	0.12	0.64 ± 0.06	0.16	0.07	0.09
2	0.85 ± 0.10	0.29 ± 0.22	0.15	0.09	0.67 ± 0.09	0.28	0.08	0.05
3	0.84 ± 0.09	0.51 ± 0.24	0.26	0.29	0.76 ± 0.07	0.41	0.20	0.30
4	0.92 ± 0.12	0.29 ± 0.21	0.08	0.08	0.70 ± 0.08	0.30	0.05	0.11
5	0.83 ± 0.14	0.38 ± 0.11	0.20	0.29	0.77 ± 0.03	0.45	0.30	0.28
6	0.87 ± 0.09	0.23 ± 0.06	0.09	0.11	0.64 ± 0.07	0.19	0.12	0.13
7	0.93 ± 0.05	0.15 ± 0.09	0.03	0.02	0.57 ± 0.03	0.07	0.04	0.04
8	0.86 ± 0.11	0.43 ± 0.18	0.05	0.15	0.76 ± 0.04	0.34	0.13	0.10
Mean	0.87	0.31	0.11	0.14	0.69	0.28	0.12	0.14
Std.	0.11	0.20	0.08	0.10	0.09	0.13	0.09	0.10

For training, recordings with 500 ms dwell threshold were used. On these recordings, 5-fold cross-validation was used to estimate, as M ± SD over folds, sensitivity and Area Under [ROC] Curve (AUC) for Button fixations, as well as specificity, in the 500 ms dwell threshold data. For each fold, the classifier was trained on 80% of the set, its threshold was adjusted on 8% aiming to obtain 90% specificity, and testing was done on 12%. In the case of testing on Ball and Free cell fixations, as well as for all 1000 ms dwell threshold data, the train and test sets were not overlapped and, therefore, the classifier was applied to all test data. Sensitivity is the same as True Positive Rate (TPR), here, the rate of correct classification of the intentional dwells. Specificity is equal to one minus the False Positive Rate (FPR); FPR is the same as the rate of false alarms, and in this work it corresponds to incorrectly classifying a spontaneous fixation as the controlling one (i.e., an intentional dwell).