Table 4.

Confusion matrixes for the cubic kernel SVM algorithm with a 600 ms window length and 10% overlap.

Stage 1: 5-fold Cross-Validation – Training Dataset
		Predicted class
		C&J	AS	BJ	BP	TRANS
True class	C&J	$\mathrm{228\; (5.9\%)}$				$\mathrm{22\; (0.6\%)}$
	AS		$\mathrm{107\; (2.8\%)}$			$\mathrm{25\; (0.6\%)}$
	BJ			$\mathrm{92\; (2.4\%)}$		$\mathrm{13\; (0.3\%)}$
	BP		$\mathrm{1\; (0.0\%)}$		$\mathrm{138\; (3.6\%)}$	$\mathrm{4\; (0.1\%)}$
	TRANS	$\mathrm{2\; (0.1\%)}$	$\mathrm{7\; (0.2\%)}$	$\mathrm{5\; (0.1\%)}$	$\mathrm{5\; (0.1\%)}$	$\mathrm{3215\; (83.2\%)}$
Stage 2: LOSO – Test Dataset
		Predicted class
		C&J	AS	BJ	BP	TRANS
True class	C&J	$\mathrm{434\; (8.1\%)}$	$\mathrm{4\; (0.1\%)}$			$\mathrm{90\; (1.7\%)}$
	AS	$\mathrm{3\; (0.1\%)}$	$\mathrm{214\; (4.0\%)}$		$\mathrm{2\; (0.0\%)}$	$\mathrm{51\; (0.9\%)}$
	BJ			$\mathrm{174\; (3.2\%)}$		$\mathrm{29\; (0.5\%)}$
	BP	$\mathrm{2\; (0.0\%)}$			$\mathrm{280\; (5.2\%)}$	$\mathrm{43\; (0.8\%)}$
	TRANS	$\mathrm{47\; (0.9\%)}$	$\mathrm{12\; (0.2\%)}$	$\mathrm{12\; (0.2\%)}$	$\mathrm{20\; (0.4\%)}$	$\mathrm{3961\; (73.3\%)}$

Classification performance is reported from the two stages of validation as total counts and % of total. Blank cells correspond to a count of zero. LOSO = “leave one subject out.” Training = sensor data from movements of Set 1 and Set 3 of the experimental protocol, used to train the classifier. Test = sensor data from movements of Set 2 of the experimental protocol, not used to train the algorithm. Results are the average across multiple iterations. Green = correct prediction, red = misclassification.