The procedure of generalized cross-subject drowsiness predictor analysis is depicted in Figure 24. The EEG data from five subjects were used as the training data, and the remaining subject was reserved as the testing pattern.
Figure 24. Generalized cross-subject drowsiness predictor analysis structure, where Si means the i-th subject.
4.2.1. Boxplots of PPMCC and RMSE
Table 11-12 summarize the averages of PPMCC and RMSE performance in comparison with the actual and estimated RTs. The PPMCC on the training and
40
testing sets using obtained by SVR, MLPNN, RBFNN, and SONFIN are 98.0%, 96.8%, 99.3%, 98.4% and 61.6%, 61.3%, 47.9%, 78.3%, respectively. The RMSE values for training and testing evaluation with SVR, MLPNN, RBFNN, and SONFIN are 0.06 s, 0.04 s, 0.01 s, 0.06 s and 0.37 s, 0.42 s, 1.01 s, 0.36 s, respectively.
Table 11. Correlation coefficients Comparisons for Generalized Cross-Subject Drowsiness Prediction
Subject 1 2 3 4 5 6
Average
(%)SVR Training 96.5% 99.1% 95.5% 98.8% 98.9% 99.0% 98.0±1.4 Testing 58.0% 65.3% 66.4% 51.4% 62.2% 66.5% 61.6±8.6 MLPNN Training 95.4% 99.4% 98.9% 91.6% 98.5% 97.0% 96.8±9.7 Testing 56.5% 66.5% 58.2% 61.2% 63.8% 61.6% 61.3±12.2 RBFNN Training 96.4% 99.6% 98.7% 98.7% 95.4% 96.7% 97.3±1.8
Testing 68.3% 48.5% 62.6% 74.2% 17.1% 16.5% 47.9±23.6 SONFIN Training 98.5% 99.1% 97.9% 97.6% 98.2% 99.0% 98.4±1.3
Testing 78.1% 81.7% 74.6% 82.7% 76.0% 76.4% 78.3±5.7
Table 12. RMSE Comparisons for Generalized Cross-Subject Drowsiness Prediction
Subject 1 2 3 4 5 6 Average (s)
SVR Training 0.092 0.046 0.089 0.049 0.048 0.047
0.060±0.030
Testing 0.222 0.309 0.454 0.267 0.540 0.3370.370±0.110
MLPNN Training 0.069 0.025 0.024 0.060 0.033 0.0410.040±0.070
Testing 0.229 0.448 0.578 0.366 0.443 0.4300.420±0.150
RBFNN Training 0.006 0.006 0.006 0.003 0.005 0.0040.010±0.002
Testing 0.417 0.467 0.798 3.361 0.496 0.5071.010±1.070
SONFIN Training 0.058 0.048 0.060 0.063 0.060 0.0460.060±0.020
Testing 0.153 0.318 0.537 0.371 0.321 0.4880.360±0.140
41
Figure 25 shows the boxplot of the PPMCC for cross-subject drowsiness prediction using SVR, MLPNN, RBFNN, and SONFIN. Take Subject 1 for example, the median, upper and lower quartile, maximum and minimum PPMCC for cross-subject drowsy state predictor with SONFIN are 77.0%, 82.9% and 74.6%, 84.0% and 74.5%, respectively.
Figure 25. Correlation coefficient boxplot comparison of subject‟s drowsy state testing evaluation for generalized cross-subject drowsiness prediction experiment with SVR, MLPNN, RBFNN and SONFIN. The boxes have three lines to present the values for lower quartile (+), median (red line), and upper quartile (++) for column data. Two addition lines at both ends of the whisker indicate the maximum (*) and minimum (**) value of a column data.
4.2.2. Experimental Result Examples
Some experimental results of testing data evaluation with cross-subject drowsiness prediction experiment for Subject 1 to Subject 6 that use SVR, MLPNN, RBFNN and SONFIN were depicted from Figure 26 to Figure 31 respectively.
Figure 26 shows some estimated RT evaluation result samples of testing
42
data for Subject 1 with cross-subject drowsiness prediction infrastructure using (a) SVR (r = 0. 5615), (b) MLPNN (r = 0. 5178), (c) RBFNN (r = 0. 6625) and (d) SONFIN (r = 0. 8352). The red dashed line and blue dash-dot line present the golden testing data and estimated evaluation result respectively. The correlation coefficients of training data validation for SVR, MLPNN, RBFNN and SONFIN in the sample results of Subject 1 are 96.5%, 95.4%, 96.4% and 96.5% respectively.
Figure 27 shows some estimated RT evaluation result samples of testing data for Subject 2 with cross-subject drowsiness prediction infrastructure using (a) SVR (r = 0.7232), (b) MLPNN (r = 0.6989), (c) RBFNN (r = 0.5230) and (d) SONFIN (r = 0.8650)The red dashed line and blue dash-dot line present the golden testing data and estimated evaluation result respectively. The correlation coefficients of training data validation for SVR, MLPNN, RBFNN and SONFIN in the sample results of Subject 2 are 98.1%, 97.9%, 98.6% and 97.8%
respectively.
Figure 28 shows some estimated RT evaluation result samples of testing data for Subject 3 with cross-subject drowsiness prediction infrastructure using (a) SVR (r = 0.6882), (b) MLPNN (r = 0.6841), (c) RBFNN (r = 0.6553) and (d) SONFIN (r = 0.7934). The red dashed line and blue dash-dot line present the golden testing data and estimated evaluation result respectively. The correlation coefficients of training data validation for SVR, MLPNN, RBFNN and SONFIN in the sample results of Subject 3 are 95.5%, 98.9%, 98.7% and 97.9%
respectively.
Figure 29 shows some estimated RT evaluation result samples of testing data for Subject 4 with cross-subject drowsiness prediction infrastructure using (a) SVR (r = 0.5998), (b) MLPNN (r = 0.6790), (c) RBFNN (r = 0.7737) and (d) SONFIN (r = 0.8510). The red dashed line and blue dash-dot line present the golden testing data and estimated evaluation result respectively. The correlation
43
coefficients of training data validation for SVR, MLPNN, RBFNN and SONFIN in the sample results of Subject 4 are 98.8%, 92.5%, 98.7% and 97.6%
respectively.
Figure 30 shows some estimated RT evaluation result samples of testing data for Subject 5 with cross-subject drowsiness prediction infrastructure using (a) SVR (r = 0.6345), (b) MLPNN (r = 0.7370), (c) RBFNN (r = 0.2033) and (d) SONFIN (r = 0.8843). The red dashed line and blue dash-dot line present the golden testing data and estimated evaluation result respectively. The correlation coefficients of training data validation for SVR, MLPNN, RBFNN and SONFIN in the sample results of Subject 5 are 97.9%, 98.5%, 96.4% and 98.2%
respectively.
Figure 31 shows some estimated RT evaluation result samples of testing data for Subject 6 with cross-subject drowsiness prediction infrastructure using (a) SVR (r = 0.6573), (b) MLPNN (r = 0.7219), (c) RBFNN (r = 0.2573) and (d) SONFIN (r = 0.8789). The red dashed line and blue dash-dot line present the golden testing data and estimated evaluation result respectively. The correlation coefficients of training data validation for SVR, MLPNN, RBFNN and SONFIN in the sample results of Subject 6 are 96.9%, 97.0%, 96.7% and 97.9%
respectively.
44
Figure 26. Evaluation result examples of testing data for subject1 with
cross-subject drowsiness prediction infrastructure using (a) SVR (r = 0.5615), (b) MLPNN (r = 0.5178), (c) RBFNN (r = 0.6625) and (d) SONFIN (r = 0.8352).
The red dashed line and blue dash-dot line present the golden testing data and estimated evaluation result respectively.
0 50 100 150 200 250 300
45
Figure 27. Evaluation result examples of testing data for subject 2 with
cross-subject drowsiness prediction infrastructure using (a) SVR (r = 0.7232), (b) MLPNN (r = 0.6989), (c) RBFNN (r = 0.5230) and (d) SONFIN (r = 0.8650).
The red dashed line and blue dash-dot line present the golden testing data and estimated evaluation result respectively.
0 50 100 150 200 250 300
46
Figure 28. Evaluation result examples of testing data for subject 3 with
cross-subject drowsiness prediction infrastructure using (a) SVR (r = 0.6882), (b) MLPNN (r = 0.6841), (c) RBFNN (r = 0.6553) and (d) SONFIN (r = 0.7934).
The red dashed line and blue dash-dot line present the golden testing data and estimated evaluation result respectively.
0 50 100 150 200
47
Figure 29. Evaluation result examples of testing data for subject 4 with
cross-subject drowsiness prediction infrastructure using (a) SVR (r = 0.5998), (b) MLPNN (r = 0.6790), (c) RBFNN (r = 0.7737) and (d) SONFIN (r = 0.8510).
The red dashed line and blue dash-dot line present the golden testing data and estimated evaluation result respectively.
0 50 100 150 200
48
Figure 30. Evaluation result examples of testing data for subject 5 with
cross-subject drowsiness prediction infrastructure using (a) SVR (r = 0.6345), (b) MLPNN (r = 0.7370), (c) RBFNN (r = 0.2033) and (d) SONFIN (r = 0.8843).
The red dashed line and blue dash-dot line present the golden testing data and estimated evaluation result respectively.
0 50 100 150 200
49
Figure 31. Evaluation result examples of testing data for subject 6 with
cross-subject drowsiness prediction infrastructure using (a) SVR (r = 0.6573), (b) MLPNN (r = 0.7219), (c) RBFNN (r = 0.2573) and (d) SONFIN (r = 0.8789).
The red dashed line and blue dash-dot line present the golden testing data and estimated evaluation result respectively.
0 50 100 150 200
50
4.2.3. Derived Parameters for SONFIN
The constructed parameters, such as rules numbers, mean values and variances of membership functions (MF) and weights for the testing data evaluation examples taken in the previous section with the generalized cross-subject drowsiness prediction using SONFIN will be comprehensively described in this section.
The RT estimation rule numbers generated by SONFIN for the generalized cross-subject experimental testing evaluation samples of Subject 1-6 are 11, 11, 12, 10, 12 and 8, as listed in Table 13, respectively. The MFs for Subject1-6 were depicted in Figure 32-37, while the corresponded mean values (mij), variance (σ2ij) and weights (wi) were summarized in Table 14-19 accordingly.
Table 13. Rules Numbers For Sampled Testing Data Evaluation Subjects Derived by Cross-Subject Drowsiness Prediction with SONFIN
Subject 1 2 3 4 5 6
Rules 8 9 8 10 10 14
51
Figure 32. Constructed Membership Functions for Sampled Subject 1 with Subject-Dependent Drowsiness Prediction using SONFIN
Table 14. Constructed Mean Values, Variances and Weights for Sampled Subject 1 with Cross-Subject Drowsiness Prediction using SONFIN
Rules Mean and Variance(mij, σ2ij )
4Hz 5Hz 6Hz 7Hz 8Hz 9Hz 10Hz 11Hz 12Hz
Weight (wi) 1 (17.47,1.94) (15.90,2.00) (15.83,1.96) (16.63,1.34) (18.11,0.82) (19.07,2.37) (19.59,2.47) (18.92,2.17) (17.08,0.93) 1.14 2 (19.20,0.08) (18.19,1.58) (17.88,2.19) (18.87,1.97) (20.54,1.94) (21.96,2.03) (22.54,2.20) (21.77,2.24) (19.60,2.17) 2.89 3 (17.90,2.42) (17.13,2.41) (17.26,2.12) (18.19,2.16) (19.97,1.79) (22.14,0.14) (22.24,1.42) (21.25,2.00) (19.41,2.41) 1.98 4 (18.42,1.21) (17.15,2.06) (16.49,2.46) (16.99,2.66) (18.55,2.24) (19.74,1.15) (20.49,2.08) (19.42,2.37) (17.34,1.20) 1.78 5 (15.53,0.55) (14.70,1.68) (14.90,1.86) (15.59,2.07) (17.28,2.12) (19.16,1.69) (19.65,2.06) (19.25,2.23) (18.70,1.86) 0.85 6 (18.60,2.02) (18.04,0.10) (17.94,1.52) (18.78,1.98) (20.28,2.42) (21.68,2.54) (22.32,2.55) (21.62,2.41) (19.51,2.18) 2.19 7 (17.23,2.01) (15.66,2.04) (15.32,2.00) (16.05,2.00) (17.79,1.80) (19.71,0.74) (19.98,1.73) (19.46,1.88) (18.32,1.97) 0.87 8 (17.26,2.01) (15.82,2.02) (15.47,2.01) (15.89,1.95) (17.14,1.88) (18.54,1.79) (19.43,1.80) (19.29,1.84) (18.51,1.93) 0.66
10 15 20 25
52
Figure 33. Constructed Membership Functions for Sampled Subject 2 with Subject-Dependent Drowsiness Prediction using SONFIN
Table 15. Constructed Mean Values, Variances and Weights for Sampled Subject 2 with Cross-Subject Drowsiness Prediction using SONFIN
Rules Mean and Variance(mij, σ2ij )
4Hz 5Hz 6Hz 7Hz 8Hz 9Hz 10Hz 11Hz 12Hz
Weight (wi) 1 (18.10,0.92) (16.84,1.78) (16.75,1.81) (18.81,0.48) (18.90,2.33) (19.63,3.03) (20.02,3.12) (19.17,3.09) (17.09,2.51) 1.82 2 (18.61,1.95) (17.92,0.69) (17.86,1.22) (18.84,1.81) (20.72,1.80) (22.22,1.96) (22.70,2.09) (21.81,2.18) (19.61,2.15) 2.72 3 (16.78,1.99) (15.21,2.02) (15.07,2.04) (15.85,2.09) (17.26,2.12) (18.49,2.16) (18.92,2.22) (18.08,1.82) (16.38,0.60) 1.22 4 (17.40,2.32) (17.12,2.08) (17.37,1.74) (18.41,1.75) (20.04,1.20) (20.85,1.21) (21.60,0.12) (20.41,1.80) (18.84,2.30) 1.46 5 (17.75,1.35) (16.02,1.86) (15.68,1.82) (16.38,1.49) (18.15,1.97) (19.33,2.32) (20.12,1.78) (19.06,2.25) (16.57,1.32) 1.41 6 (16.61,2.20) (15.45,2.02) (15.47,1.99) (16.52,2.04) (18.21,1.85) (20.90,0.09) (20.71,1.31) (19.80,1.81) (18.15,2.09) 1.05 7 (15.43,0.93) (14.64,1.60) (15.00,1.96) (15.98,2.18) (17.30,2.14) (18.36,2.15) (18.61,2.23) (18.16,2.19) (17.36,2.15) 0.86 8 (18.48,1.96) (17.99,1.43) (18.00,1.40) (18.95,1.64) (20.68,1.76) (21.92,1.98) (22.36,2.03) (21.44,2.08) (19.20,2.07) 1.45 9 (17.06,2.06) (15.79,1.99) (15.73,1.92) (16.16,2.01) (16.74,2.27) (17.65,2.33) (18.43,2.14) (18.27,1.96) (17.22,2.05) 0.58
10 15 20 25
53
Figure 34. Constructed Membership Functions for Sampled Subject 3 with Subject-Dependent Drowsiness Prediction using SONFIN
Table 16. Constructed Mean Values, Variances and Weights for Sampled Subject 3 with Cross-Subject Drowsiness Prediction using SONFIN
Rules Mean and Variance(mij, σ2ij )
4Hz 5Hz 6Hz 7Hz 8Hz 9Hz 10Hz 11Hz 12Hz
Weight (wi) 1 (17.38,2.74) (16.76,3.03) (17.38,2.48) (19.09,0.17) (19.28,2.78) (19.95,3.31) (20.48,3.09) (20.22,2.53) (18.72,1.86) 2.07 2 (18.90,1.73) (18.39,0.26) (18.50,1.15) (19.21,1.77) (20.54,1.86) (21.82,1.87) (22.52,1.92) (22.14,1.94) (20.24,2.00) 2.04 3 (18.46,1.09) (16.14,2.82) (16.45,2.82) (17.93,2.60) (19.68,2.34) (22.09,0.49) (20.88,2.80) (19.64,2.93) (17.67,2.32) 1.45 4 (16.48,2.09) (15.29,2.02) (15.96,2.02) (17.54,1.81) (18.92,1.83) (20.62,1.29) (21.60,0.32) (19.62,1.72) (18.00,1.56) 1.21 5 (17.01,2.52) (16.06,2.72) (16.50,2.57) (17.64,2.50) (19.48,2.24) (21.23,1.89) (22.97,0.11) (21.64,1.26) (19.63,1.79) 1.57 6 (15.72,0.91) (14.65,1.64) (14.96,1.99) (15.86,2.38) (17.09,2.57) (18.10,2.58) (18.31,2.39) (17.72,1.94) (16.77,1.55) 0.95 7 (16.27,2.17) (15.11,2.05) (15.62,2.07) (16.95,2.07) (18.54,1.91) (20.32,1.45) (21.62,0.38) (20.53,1.49) (18.90,1.89) 0.99 8 (15.43,1.42) (14.51,1.66) (14.61,1.78) (15.38,2.01) (16.61,2.19) (17.66,2.22) (17.96,2.12) (17.50,1.90) (16.83,1.88) 0.46
10 15 20 25
54
Figure 35. Constructed Membership Functions for Sampled Subject 4 with Subject-Dependent Drowsiness Prediction using SONFIN
Table 17. Constructed Mean Values, Variances and Weights for Sampled Subject 4 with Cross-Subject Drowsiness Prediction using SONFIN
Rules Mean and Variance(mij, σ2ij )
4Hz 5Hz 6Hz 7Hz 8Hz 9Hz 10Hz 11Hz 12Hz
Weight (wi) 1 (18.42,0.30) (17.52,1.07) (17.53,1.87) (18.55,2.06) (19.46,3.18) (20.20,3.65) (20.71,3.51) (20.39,3.06) (18.67,2.65) 2.72 2 (17.79,1.77) (16.98,0.46) (16.81,1.91) (17.76,1.93) (19.33,1.36) (20.69,0.64) (20.24,2.12) (19.20,2.35) (17.48,1.95) 1.95 3 (17.96,1.19) (16.20,1.79) (15.88,1.81) (16.85,1.50) (18.58,1.45) (19.72,1.78) (20.43,1.30) (18.93,2.04) (16.68,0.55) 1.13 4 (17.13,2.36) (16.14,2.21) (16.27,2.16) (17.35,2.17) (19.33,1.81) (21.44,0.22) (19.44,1.61) (20.79,1.48) (19.03,2.04) 1.33 5 (16.35,1.08) (14.64,1.65) (14.99,1.84) (15.97,2.14) (17.36,2.22) (18.29,2.20) (18.27,2.07) (17.35,1.68) (15.93,1.39) 1.09 6 (17.40,0.73) (16.48,1.86) (16.33,2.05) (17.28,2.03) (18.75,1.85) (20.33,1.16) (21.16,0.56) (19.61,1.76) (17.81,2.01) 1.58 7 (15.06,0.54) (14.41,1.55) (14.63,1.80) (15.50,2.12) (16.77,2.28) (17.73,2.26) (17.90,2.10) (17.28,1.75) (16.40,1.74) 0.64 8 (18.11,1.71) (16.94,1.90) (16.99,1.99) (18.22,1.86) (20.04,1.46) (21.33,1.26) (21.61,1.71) (20.83,1.92) (18.79,2.02) 1.39 9 (16.44,2.39) (15.23,2.17) (15.56,2.01) (16.80,1.90) (18.40,1.85) (19.79,1.80) (21.05,0.72) (20.65,0.99) (18.76,1.96) 0.79 10 (16.10,1.98) (14.94,1.97) (14.88,1.97) (15.47,2.02) (16.60,2.06) (17.57,2.06) (17.95,2.04) (17.60,2.02) (16.84,2.03) 0.64
10 15 20 25
55
Figure 36. Constructed Membership Functions for Sampled Subject 5 with Subject-Dependent Drowsiness Prediction using SONFIN
Table 18. Constructed Mean Values, Variances and Weights for Sampled Subject 5 with Cross-Subject Drowsiness Prediction using SONFIN
Rules Mean and Variance(mij, σ2ij )
4Hz 5Hz 6Hz 7Hz 8Hz 9Hz 10Hz 11Hz 12Hz
Weight (wi) 1 (17.36,0.05) (16.93,1.12) (16.82,1.13) (18.00,1.04) (19.76,0.71) (20.28,1.89) (20.84,2.03) (20.11,2.25) (18.33,2.21) 1.90 2 (18.24,1.35) (17.00,1.21) (17.20,1.36) (18.47,1.32) (20.37,1.53) (22.00,1.51) (22.71,1.48) (22.13,1.42) (20.08,1.42) 2.70 3 (16.98,1.00) (15.47,1.30) (15.36,1.48) (16.41,1.47) (17.86,1.19) (19.29,0.56) (19.71,1.09) (18.61,1.16) (16.85,0.81) 1.06 4 (17.69,0.53) (16.86,0.34) (16.68,1.23) (17.65,1.33) (19.03,1.26) (20.46,1.18) (21.14,1.23) (20.64,1.35) (18.83,1.44) 1.60 5 (15.36,0.90) (13.99,0.14) (14.66,0.97) (16.09,1.27) (17.68,1.32) (18.80,1.37) (18.99,1.46) (18.38,1.43) (17.32,1.28) 0.74 6 (17.15,1.14) (15.75,0.86) (15.77,0.60) (16.82,1.12) (18.27,0.98) (19.60,0.87) (20.12,0.98) (19.36,1.34) (17.60,1.37) 1.41 7 (16.04,1.38) (15.28,1.23) (15.92,1.20) (17.28,1.29) (19.11,0.56) (20.24,0.57) (20.39,1.27) (19.89,1.66) (18.44,1.82) 0.89 8 (15.79,1.13) (14.68,1.04) (14.63,0.62) (15.63,0.88) (16.88,1.18) (18.06,1.27) (18.39,1.26) (17.94,1.21) (17.03,1.16) 0.71 9 (16.45,1.57) (15.59,1.56) (16.04,1.47) (17.02,1.33) (18.10,1.32) (18.97,1.50) (19.59,1.37) (19.41,1.24) (18.04,1.35) 0.72 10 (18.12,1.18) (16.61,1.14) (16.77,1.15) (18.17,1.16) (19.93,1.16) (21.74,1.13) (22.58,1.08) (22.14,1.03) (20.08,1.06) 0.40
10 15 20 25
56
Figure 37. Constructed Membership Functions for Sampled Subject 6 with Subject-Dependent Drowsiness Prediction using SONFIN
Table 19. Constructed Mean Values, Variances and Weights for Sampled Subject 6 with Cross-Subject Drowsiness Prediction using SONFIN
Rules Mean and Variance(mij, σ2ij )
4Hz 5Hz 6Hz 7Hz 8Hz 9Hz 10Hz 11Hz 12Hz
Weight (wi) 1 (17.75,0.22) (16.74,1.16) (16.38,1.28) (17.23,1.29) (18.29,0.80) (19.78,0.47) (20.28,1.38) (19.59,1.68) (17.86,1.74) 1.72 2 (17.36,0.74) (16.93,1.22) (17.00,1.49) (17.51,1.22) (19.92,1.56) (21.56,1.55) (22.03,1.47) (21.22,1.30) (18.99,1.15) 2.54 3 (16.48,1.00) (15.15,1.38) (15.52,1.49) (16.52,1.53) (17.79,1.55) (19.11,1.25) (20.06,0.10) (18.35,1.25) (16.30,0.91) 1.10 4 (16.86,1.31) (15.64,0.30) (15.89,0.94) (16.85,1.14) (18.22,0.97) (19.37,0.33) (19.79,1.21) (19.31,1.72) (17.65,1.84) 1.37 5 (18.00,1.28) (17.03,0.12) (17.00,1.00) (18.05,1.19) (19.70,1.40) (21.16,1.48) (21.75,1.49) (21.17,1.46) (19.21,1.36) 1.76 6 (16.79,0.80) (15.69,0.97) (15.96,1.09) (16.74,1.29) (17.82,1.39) (18.75,1.47) (18.83,1.47) (18.01,1.20) (16.14,0.32) 1.01 7 (15.50,0.44) (14.73,1.05) (15.37,1.15) (16.62,1.16) (17.84,1.27) (18.75,1.31) (18.90,1.28) (18.02,1.13) (16.57,1.14) 0.92 8 (17.84,1.33) (17.33,0.33) (16.97,1.18) (17.86,1.32) (19.44,1.59) (20.82,1.61) (21.33,1.51) (20.70,1.27) (18.91,0.91) 1.65 9 (17.37,0.60) (15.87,1.00) (15.78,1.08) (16.70,1.09) (18.09,1.20) (19.42,1.24) (20.31,0.52) (19.76,0.43) (17.70,1.24) 1.17 10 (15.71,0.99) (14.53,0.88) (14.50,0.76) (15.22,0.70) (16.60,1.15) (17.68,1.24) (18.19,1.20) (17.80,1.07) (16.69,0.93) 0.68 11 (15.65,1.17) (14.91,1.16) (15.44,1.23) (16.56,1.35) (18.28,1.25) (20.07,0.78) (21.03,0.15) (20.37,1.34) (18.86,1.52) 0.76 12 (17.73,1.20) (16.09,1.22) (16.01,1.17) (17.34,1.16) (19.35,1.32) (21.11,1.39) (21.76,1.36) (21.26,1.31) (19.52,1.29) 0.91 13 (16.66,1.61) (15.51,1.37) (15.84,1.12) (17.27,0.73) (18.54,0.75) (19.73,0.40) (19.66,0.93) (18.69,1.00) (17.24,1.23) 0.85 14 (17.17,1.20) (15.84,1.20) (15.69,1.20) (16.51,1.20) (17.47,1.21) (18.48,1.21) (18.98,1.20) (18.71,1.21) (17.67,1.23) 0.67
10 15 20 25
57
4.2.4. Section Discussion
Compared to the subject-dependent drowsiness results, the averaged PPMCC between the actual and estimated RTs on testing data with these four predictors maintained sound results.
However, the PPMCC obtained by the generalized cross-subject drowsiness prediction showed a significant performance decline on the test data (p-value <
0.038). Only SONFIN still maintained a better PPMCC between actual and estimated RTs at 78.3% than other predictors. Furthermore, the SONFIN produced the lowest RMSE (0.36 s) on the testing data in this experiment.
According to safety distance between vehicles reported by CEDR [61] and RSA [62], a rule thumb of 2-s braking distance under dry ground conditions with additional reaction distance of 18.3 m at a 100 km/hr car speed is recommended.
The RMSE of proposed cross-subject drowsiness predictor with SONFIN is 0.36 s or 10 m at a 100 km/hr car speed in average, which does not violate the recommended reaction distance requirement of 18.3 m. Therefore, the proposed cross-subject drowsy state predictor with SONFIN showed a promising model for real-life applications.