Improved Seizure Prediction Using Discrete Hidden Markov Model and Wilks' Lambda Analysis of the Electroencephalographic Signals

Background: In this paper, the Stationary Wavelet Transform (SWT) together with Hidden Markov Model (HMM) were utilized for accurate prediction of epileptic seizure patterns. Tests using popular intracranial Electroencephalograph (iEEG) recordings involving 18 seizure patients of different sex, age and seizure's type indicated that the algorithm performs reasonably well by three major iEEG rhythms. Methods: Three different seizure states were considered in the investigation - (i) ictal, (ii) and preictal and, (iii) interictal. A sliding window approach with data averaging was implemented in order to avoid overlapping and ensuring balanced datasets. Meanwhile the 4th order Daubechies wavelet was utilized in signal decomposition, while machine learning was established by means of 5-state HMM classifier. During training the Wilks' lambda algorithm was invoked in order to reduce correlationship between variables by selecting those with high discriminant power. Results: The algorithm took forty-seven steps to converge, producing a subset containing 44 variables from 2560 available. Results from this study reveal that the classification after Wilks' lambda analysis was more precise compared to direct classification. Prediction analysis performed on all principle components yielded a correct classification rate of 95.1%, 95.2% sensitivity, and 97.6% specificity. Results demonstrate that the proposed method were more accurate compared to the existing methods. Conclusion: It is shown in this paper that HMM with Wilks' lambda analysis were capable of escalating the correct classification decisions compared to direct approach. The difficulty in separating preictal from ictal rhythms is evident from the canonical plot and proven by classification analysis.