Improved EEG Event Classification Using Differential Energy A.
Improved EEG Event Classification Using Differential Energy A. Harati, M. Golmohammadi, S. Lopez, I. Obeid and J. Picone Neural Engineering Data Consortium Temple University The TUH EEG Corpus The TUH EEG Corpus contains over 28,000 sessions collected from 15,000+ patients over a period of 14 years at Temple University Hospital. M. Golmohammadi: Improved EEG Event Classification Using Differential Energy December 12, 20152
AutoEEG: Automatic Interpretation of EEGs AutoEEG is a hybrid system that uses three levels of processing to achieve high performance event detection on clinical data: P1: hidden Markov models P2: Deep Learning P3: Language Model M. Golmohammadi: Improved EEG Event Classification Using Differential Energy December 12, 20153 AutoEEG: Clinically-Relevant Signal Events Six events of interest based on multiple iterations with board certified neurologists: Clinical data contains many artifacts that can easily be interpreted as spikes, such as patient movement.
Manual review of the data is a time-consuming process, especially for long-term EEG recordings. In ICU applications, real-time response is critical. Performance requirements for this application is a detection rate above 95% with a false alarm rate below 5%. For ICU applications, the false alarm rate must be extremely low (e.g., 1/8 FA/hr.). M. Golmohammadi: Improved EEG Event Classification Using Differential Energy December 12, 20154 AutoEEG: Feature Extraction Goal: Calibrate traditional approaches to feature extraction based on cepstral coefficients, energy and derivatives. Mel Frequency Cepstral Coefficients (MFCCs): A linear frequency scale is used for EEGs.
We have shown wavelets have very little advantage over MFCCs on the TUH EEG (which contradicts the literature). Introduced a new differential energy term that accentuates the differences between quasi-periodic signals and spikes. M. Golmohammadi: Improved EEG Event Classification Using Differential Energy December 12, 20155 AutoEEG: Feature Extraction Time Domain energy an overlapping analysis window (a 50% overlap was used here) to ensure a smooth trajectory
Frequency Domain energy provides a smoother, more stable estimate of the energy that leverages the cepstral representation of the signal. Derivatives of features: A typical value for N is 9 (corresponding to 0.9 secs) for the first derivative in EEG processing, and 3 for the second derivative. M. Golmohammadi: Improved EEG Event Classification Using Differential Energy December 12, 20156 AutoEEG: Feature Extraction Differential energy:
To improve differentiation between transient pulse-like events (e.g., SPSW events) and stationary background noise This simple feature has proven to be surprisingly effective. We typically use a 0.9 sec window for this calculation. M. Golmohammadi: Improved EEG Event Classification Using Differential Energy December 12, 20157 Experiments: The Curse of Dimensionality The use of differential features raises the dimension of a typical feature vector from 9 (e.g., 7 cepstral coefficients, frequency domain energy and differential energy) to 27. The training set contains segments from 359 sessions while the evaluation set was drawn from 159 sessions.
M. Golmohammadi: Improved EEG Event Classification Using Differential Energy December 12, 20158 Experiments: Absolute Features 6-way classification: SPSW, GPED, PLED, EYEM, ARTF, BCKG 4-way classification: collapse the 3 background classes into one category, resulting in 4 classes SPSW, GPED, PLED and BACKG. 2-way classification: collapse SPSW, GPED and PLED into a target class (TARG) and (ARTF, EYEM and BCKG) into a background class (BCKG). M. Golmohammadi: Improved EEG Event Classification Using Differential Energy December 12, 20159 Experiments: Differential Features
Derivatives reduced the error rate on the 6-way task by 4% absolute (systems no. 1, 6 and 11). However, the improvement for a system using differential energy is much less pronounced (systems no. 5, 10 and 15). Eliminate the second derivative for differential energy (no. 16): M. Golmohammadi: Improved EEG Event Classification Using Differential Energy December 12, 2015 10 Analysis: DET Curves DET curves for systems nos. 1, 5, 10, and 15 are shown below. The addition of first derivatives provides a measurable improvement in performance while second derivatives are less beneficial. Differential energy provides a substantial improvement over the base cepstral features.
M. Golmohammadi: Improved EEG Event Classification Using Differential Energy December 12, 2015 11 Analysis: Deep Learning and Language Modeling System no. 15: 7 cepstral coefficients 1 f_energy 1 d_energy 9 deltas for all 9 delta-deltas for all Total: 27 features System no. 16: 7 cepstral coefficients 1 f_energy
1 d_energy 9 deltas for all 8 delta-deltas (cepstral + f_energy only) Total: 26 features M. Golmohammadi: Improved EEG Event Classification Using Differential Energy December 12, 2015 12 Summary Established valuable baselines for standard feature extraction approaches for EEG signals. Traditional feature extraction methods used in speech recognition are relevant to EEGs. The use of a novel differential energy feature improved performance for
absolute features, but that benefit diminishes as first and second order derivatives are included. There is benefit to using derivatives and there is a small advantage to using frequency domain energy. There are no significant benefits to using wavelets and other timefrequency representations. Future work will focus on new feature extraction methods based on principles of deep learning, discriminative training and nonparametric Bayesian models. The database will be significantly expanded over the next year and evolved into new tasks (e.g., real-time seizure detection). M. Golmohammadi: Improved EEG Event Classification Using Differential Energy December 12, 2015 13 Brief Bibliography
 T. Yamada and E. Meng, Practical Guide for Clinical Neurophysiologic Testing: EEG. Philadelphia, Pennsylvania, USA: Lippincott Williams & Wilkins, 2009.  A. Harati, S. Lopez, I. Obeid, M. Jacobson, S. Tobochnik, and J. Picone, THE TUH EEG CORPUS: A Big Data Resource for Automated EEG Interpretation, in Proceedings of the IEEE SPMB, 2014, pp. 15.  J. Picone, Continuous Speech Recognition Using Hidden Markov Models, IEEE ASSP Mag., vol. 7, no. 3, pp. 2641, Jul. 1990. 
S. Sanei and J. A. Chambers, EEG signal processing. Hoboken, New Jersey, USA: WileyInterscience, 2008.  J. Lyons, Mel Frequency Cepstral Coefficient (MFCC) tutorial, Practical Cryptography, 2015 (available: http://practicalcryptography. com/miscellaneous/machine-learning/guide-melfrequency-cepstral-coefficients-mfccs/.  S. Davis and P. Mermelstein, Comparison of Parametric Representations for Monosyllabic Word Recognition in Continuously Spoken Sentences, IEEE Trans. ASSP, vol. 28, no. 4, pp. 357366, 1980.  P. Garrit, et al., Wavelet Analysis for Feature Extraction on EEG Signals, presented at
Temple University CoE Res. Exp. for UG Conf., 2015 (available at http://www.isip.piconepress.com/publications/ unpublished/conferences/2015/summer_of_code/ wavelets/). M. Golmohammadi: Improved EEG Event Classification Using Differential Energy December 12, 2015 14
Well-being websites published on VLE, and student diary, well-being noticeboard in the sixth form. Year 13 students will adopt a Year 12 student as a buddy! Our counsellor - a trained expert who you can talk to or be referred...
Risk adjustment is a method of determining and balancing the cost associated with caring for patients with varying severities of illness. Risk Adjustment is current trend in payment methodology for leveling the payment and cost of caring for a patient...
Recognizing that victims of domestic violence often seek medical care before calling the police, and that many physicians lack adequate training and guidelines to address domestic violence, AB 890 was introduced to the California legislature in 1993 and passed into...
Characteristics of functions describe specific attributes of the function. They tell us things like, for example, how tall/short, slim/wide, x-values we can & can not use, etc. We are going to focus on the following characteristics of radical functions:
An approximate but computationally efficient numerical approach is also developed to simulate reverberation over wide areas for operational sonar syatems. The numerical model employs the Navy standard range-dependent propagation model based on the parabolic equation, RAM, and can be readily...
Shape Definitions: Symmetric: if the right and left sides of the graph are approximately mirror images of each other. Skewed to the right (right-skewed) if the right side of the graph is much longer than the left side.
Ribose sugar. Phosphate group. Nitrogenous Base. A, U, C, G. Uracil replaces Thymine. DNA vs. RNA. DNA. RNA. NUMBEROF STRANDS. LOCATION. SMALLEST SUBUNIT. TYPE OF SUGAR. NITROGEN BASES. DNA vs. RNA. Types of RNA. mRNA (Messenger RNA) created from DNA;...
Ready to download the document? Go ahead and hit continue!