The methodology of Bayesian Model Averaging (BMA) is applied for assessment of newborn brain maturity from sleep EEG. In theory this methodology provides the most accurate assessments of uncertainty in decisions. However, the existing BMA techniques have been shown providing biased assessments in the absence of some prior information enabling to explore model parameter space in details within a reasonable time. The lack in details leads to disproportional sampling from the posterior distribution. In case of the EEG assessment of brain maturity, BMA results can be biased because of the absence of information about EEG feature importance. In this paper we explore how the posterior information about EEG features can be used in order to reduce a negative impact of disproportional sampling on BMA performance. We use EEG data recorded from sleeping newborns to test the efficiency of the proposed BMA technique.
Assessment of brain maturity can be obtained by estimating newborn’s age from sleep EEG  - . This approach is based on the clinical evidences that the post-conceptional and EEG estimated ages of healthy newborns typically match each other, and the newborn’s brain maturity is most likely abnormal if the ages mismatch , . Thus, the mismatch alerts about abnormal brain development.
The established assessment methodologies are based on learning models from EEGs recorded from sleeping newborns whose brain maturity was already assessed by clinicians. The regression models are made capable of mapping the brain maturity into EEG based index . The classification models are made capable of distinguishing maturity levels: at least one with normal and other with abnormal brain maturity , . The established methodologies are based on learning a single model from a given set of data and they cannot ensure that a model will not be overfitted to the data
Probabilistic reasoning, based on the Bayesian methodology of averaging over decision models, enables to evaluate the uncertainty in decision making  – . The use of the Bayesian Model Averaging (BMA) over Decision Trees (DTs) enables to select features which make the most significant contribution to outcomes, and the resultant DT ensemble can be interpreted by clinicians as shown in . However, the success in implementation of BMA is critically dependent on the diversity and proportion of models sampled for averaging. The models should be diverse in parameters and structure. The portion of models whose likelihood is high should be largest to ensure unbiased estimates. The use of a priori information provides better conditions for achieving these requirements. In many practical cases when a priori information about feature importance is absent, the provision of the diversity and proportion of models becomes problematic , .
In our previous research , we attempted to overcome the above BMA problems and developed a new feature selection strategy for...