Compared to the invasive type electromyogram (EMG) for measuring each motor unit action potential (MUAP) on muscle fiber in a small area, surface EMG signals can be recorded from the human body non-invasively and easily over wide skin areas. To extract the motion information, surface EMG signal processing has been studied, especially focusing on controlling electronic devices such as a human supporting prosthetic arms and smart interface (Sornmo and Laguna, 2005). Since motor commands are transmitted through internal nerve tissues or muscles, inferring such commands only from the surface EMG signals is not an easy task.

Thus many attempts to extract motor information from surface EMG have been made. A study of Englehart et al. (2001) used wavelet analysis and PCA analysis to classify two arm motions and two wrist motions. Also, Englehart and Hudgins (2003) used the absolute mean value, the zero crossing rate, and the wavelength to classify four arm and wrist motions. Momen et al. (2007) used RMS and FCMs to classify arm and wrist motions. Hudgins et al. (1993) obtained an ensemble average of EMG signals, and used a neural networks (NN) to classify four arm motions.

While the surface EMG-based motion inference has been focused on upper limb or hand motions, studies on inferring dexterous finger motions have not been done so much. Nishikawa et al. (1999) used the Gabor transform and absolute mean value to extract the features and classify four wrist motions and six finger motions in real-time learning based on NN. Nagata et al. (2007) used absolute sum analysis, canonical component analysis, and minimum Euclidean distance to classify four wrist motions and five finger motions. Uchida et al. (1992) used FFT analysis and NN to classify four finger motions. Chen et al. (2007) used mean absolute values (MAV), the ratio of the MAVs, autoregressive (AR) model, and linear Bayesian classifier to classify 5~16 finger motions. However, NN classifiers (Hudgins et al., 1993; Chen et al., 2007; Nagata et al., 2007) have strong dependency on initial parameter conditions. Also, linear classifiers (Uchida et al., 1992; Englehart et al., 2001; Englehart and Hudgins, 2003; Momen et al., 2007; Nagata et al., 2007) are simple but show low accuracy compared with nonlinear classifiers.

In this paper, we propose a novel method to infer the finger motions using surface EMG signals. We recorded the EMG signals during finger flexing motions. To infer the finger motions, we used the maximum likelihood method. The likelihood was obtained by building probabilistic models of EMG activities. The information entropy of the EMG magnitude was calculated to quantify the EMG activities. Experimentally we demonstrated that the proposed method could infer the finger flexing motions including multi-fingers as high as 97.75%.

In addition, we examined the influence of inference accuracies with various arm postures.

A cross-validation is used to estimate the inference accuracies. The number of training data is 25 and that of the test data is 25. All test data are excluded from the training data. The results are averaged over 500 combinations.

The accuracy for six motions (F₁, F₂, F₃, F₄, and F₈) using the proposed method was 100.00% without any confusions as shown in Table 1. The proposed method can infer the motions more accurately than the studies of Uchida et al. (1992) and Chen et al. (2007)

Fig. 6 shows Likelihood for various candidates of finger motions (each value is normalized so that the maximum was equal to one. k and k̂ denote the actual and the inferred motions, respectively).

(a) k = F₁, k̂ = F₁ (b) k = F₂, k̂ = F₂

(c) k = F₃, k̂ = F₃ (d) k = F₄, k̂ = F₄

(e) k = F₅, k̂ = F₅ (f) k = F₆, k̂ = F₆

(g) k = F₇, k̂ = F₇ (h) k = F₈, k̂ = F₈

The average inference accuracy for four motions (flexion of thumb, flexion of index finger, flexion of middle finger, flexion of all fingers) in a study (Uchida et al., 1992) was 86% and the accuracy for six motions (extension of thumb, extension of index finger, extension of middle finger, extension of ring finger, extension of little finger, and 'hook' gesture) was 96.83% (Chen et al., 2007).

The inference results for eight motions (F₁~F₈) are summarized in Table 2 showing the high inference accuracy. The inference accuracies on F₁, F₄, F₅, and F₈ are 100% but the slight decreases on F₂, F₃, F₆, and F₇ arise from false inference. Fig. 7 shows the confusion matrix which visualizes the false inference among the performed motions and Table 3 shows the false inference rate which indicates the rate when performed motion is inferred incorrectly. The false inference rate on F₂ confused with F₆ is 2.00±1.41%. The false inference rate on F₃ confused with F₆ is 5.00±2.20% while that on F₆ confused with F₃ is 8.00±2.74%. The false inference rate on F7 confused with F₃ is 3.00±1.72%.

We have proposed a new method to infer finger flexing motions based on the entropy and the maximum likelihood estimation. The average recognition accuracy for six finger motions was 100.00% and the accuracy for eight finger motions was 97.75%. Also we have shown the quantitative interpretation of the need of avoidance of use of reference dataset cross arm posture because that influences the correlations between the EMG signals and the performed finger flexing motions. This study may trigger a new type of human-computer interface with user's intuitive hand motions which could control electronic devices.

Table 1. Inference accuracy (%) for six motions (subject A, k: actual motion, k̂: inferred motion)

Table 2. Inference accuracy (subject A, test data: 25, training data: 25)

Table 3. False table (subject A, k: actual motion, k̂: inferred motion)

Table 4. Average Inference accuracy