WSEAS AIKED, Cambridge, Feature Importance in Bayesian Assessment of Newborn Brain Maturity from EEG Livia Jakaite, Vitaly Schetinin and Carsten Maple Department of Computer Science and Technology University of Bedfordshire
WSEAS AIKED, Cambridge, Outline EEG assessment of brain maturity Why Bayesian Model Averaging (BMA) for the assessment? Problems in using BMA for assessing brain maturity Solution: using posterior information about features Computational Experiments Conclusions
WSEAS AIKED, Cambridge, EEG assessment of brain maturity Newborn brain dismaturity alerts about neurophysiologic abnormality Experts can assess newborn brain maturity by estimating a newborns age from an EEG recording The accuracy of such estimate is usually two weeks Brain maturity is assessed as normal if the newborn’s physical age is within the range of EEG-estimated ages; otherwise the maturity is assessed as abnormal
WSEAS AIKED, Cambridge, … EEG assessment of brain maturity: EEG examples for different ages 28 weeks 36 weeks 40 weeks 20 s
WSEAS AIKED, Cambridge, BMA for brain maturity assessment Bayesian Model Averaging (BMA), in theory, provides the most accurate assessments and estimates of uncertainty In practice, Markov Chain Monte Carlo (MCMC) is used to approximate the posterior distribution by taking random samples
WSEAS AIKED, Cambridge, …BMA for brain maturity assessment: exploring the posterior probability An idea behind BMA is to average over multiple models diverse in their parameters To ensure unbiased estimates, the portions of models sampled from the posterior distribution should be proportional to their likelihoods The assessments will be most accurate, and the variation in models outcomes will be interpreted as the uncertainty in assessment
WSEAS AIKED, Cambridge, Change variable move Change threshold move Combine 2 terminal nodes (death move) Split a terminal node (birth move) …BMA for brain maturity assessment: exploring the posterior probability The exploration is made with moves chosen with predefined probability during a burn-in phase Each move changes the model parameters and is accepted or rejected accordingly to Bayes’ rule During a post burn-in phase, models are collected to be averaged X 1, 1 X 2, 2 X 5, 5 X 3, 3 X 4, 4
WSEAS AIKED, Cambridge, …BMA for brain maturity assessment: lack of prior information causing biased sampling To collect models proportionally, a model parameter space must be explored in detail When the model parameter space is large, possible problem is: Not all areas of PDF are explored, and then the models are disproportionally sampled Prior information about feature importance helps to reduce a model parameter space
WSEAS AIKED, Cambridge, However, in our case, no prior information on feature importance is available The EEG data is represented by spectral features and their statistical characteristics, in total by 72 attributes, some of them make weak contribution To assess the feature importance, we can use Decision Trees (DTs) for BMA …BMA for brain maturity assessment: lack of prior information causing biased sampling
WSEAS AIKED, Cambridge, If an attribute was rarely used in DTs included in the ensemble, we assume that this attribute makes a wear contribution When the number of weak attributes is large, the disproportion in models becomes significant Our hypothesis is that discarding the models using weak EEG attributes will reduce the negative effect of disproportional sampling Solution: using posterior information about features
WSEAS AIKED, Cambridge, Experiments A BMA ensemble was collected from DTs learned from EEG data represented by the 72 attributes We calculated the posterior probability of using each attribute in the DTs We refined the DT ensemble from those DTs which use weak attributes For comparison, we rerun BMA on the EEG without the identified weak attributes
WSEAS AIKED, Cambridge, Performance of the BMA on age groups of 40 – 45 weeks DTs 6 classes Performance 27.4 ± 8.2 Entropy: ± w44 w45 w 40 w41 w42 w
WSEAS AIKED, Cambridge, Posterior feature importance Spectral powersStatistical characteristics DeltaAlphaDeltaAlpha Posterior
WSEAS AIKED, Cambridge, Performance of BMA with discarded attributes Performance, % 29.0± ±8.2 Entropy 478.3± ± Threshold Threshold 25.8±1.7
WSEAS AIKED, Cambridge, Performance of BMA with the refined ensemble 27.4± ±7.9 Performance, % Entropy 478.3± ± Threshold Threshold
WSEAS AIKED, Cambridge, Performance of BMA with the refined ensemble Performance, % Count
WSEAS AIKED, Cambridge, Conclusions The larger the number of weak attributes, the greater the negative impact on BMA performance Reduction of the data dimensionality by discarding of weak attributes enabled improving BMA performance (1.6%) due to reducing a model parameter space The proposed technique provides comparable improvement in performance (1.8%) without the need of rerunning the BMA
WSEAS AIKED, Cambridge, Acknowledgements This research is funded by the Leverhulme Trust