Table 2.

Classification performance of machine learning algorithms for functional primitives.

Algorithm	PPVs for functional primitives				Overall PPV
	Reach	Transport	Reposition	Idle
LDA	93 ± 1.47%	91 ± 1.65%	93 ± 1.47%	92 ± 1.56%	92.5 ± 1.52%
NBC	77 ± 2.42%	71 ± 2.61%	83 ± 2.16%	85 ± 2.06%	80.2 ± 2.30%
SVM	92 ± 1.56%	90 ± 1.73%	92 ± 1.56%	93 ± 1.47%	92 ± 1.56%
KNN	86 ± 2.00%	87 ± 1.94%	85 ± 2.06%	89 ± 1.80%	87.5 ± 1.90%

Positive predictive values (PPV) with associated 95% confidence intervals are shown. PPV reflects how often a primitive was actually made when the algorithm identified it as such, was calculated for the primitives of reach, transport, reposition, and idle. Primitive-level PPVs were computed in one-vs.-all analysis (e.g., reach vs. transport + reposition + idle combined). The overall PPV was assessed by combining data for all primitives and tallying all true and false positives. Overall classification performance was highest for linear discriminant analysis (LDA) and support vector machine (SVM), moderately high for k-nearest neighbors (KNN), and lowest for Naïve Bayes classifier (NBC).