Wavelet-based Gaussian-mixture hidden Markov model for the detection of multistage seizure dynamics: A proof-of-concept study

Table 5 A summary of HMM performance measures

Features	Wavelet Coefficients Only		Wavelet Coefficients + Rate of Change
	Validation Set	mRMR and ΔAICc	Validation Set	mRMR and ΔAICc
	7-D (Q = 5, M = 4)	2-D (Q = 3, M = 2)	14-D (Q = 8, M = 3)	4-D (Q = 5, M = 3)
Sensitivity (TP)	69.8 ± 20.3%	80.9 ± 34.4%	86.7 ± 27.2%	95.7 ± 14.0%
Specificity (TN)	88.1 ± 20.9%	94.8 ± 15.6%	98.6 ± 7.7%	98.9 ± 6.5%
Detection delay (ΔT)	8.30 ± 15.33 s	3.79 ± 8.67 s	-0.68 ± 10.07 s	-2.03 ± 7.10 s
Optimality index (O)	0.665 ± 0.260	0.813 ± 0.247	0.915 ± 0.302	0.995 ± 0.129

The optimal HMM model topologies were determined using two methods. The first method involved the use of validation data set consisting of 10 SLEs to find the HMM topology that gave the highest optimality index. The second method used the mRMR feature selection method and AICc to reduce the feature space dimension and the complexity of the HMM. The HMM obtained using mRMR and AICc has the highest optimality index.

ISSN: 1475-925X