Figure 1.

Classification performance. Subject mean validation accuracy for the six approaches using only 10 features and 100 patterns. Subject range is indicated by the error bars. The performance appears to increase with increased classifier complexity and tailoring, and the non-linear methods perform better than the linear (p < 0.05). The mean difference between wrapper non-linear and wrapper linear is small, suggesting that a high degree of classifier and subset tailoring is more critical than non-linearity. The random feature selection performance is significantly lower than the high-performing wrapper classifiers (p < 0.01).