Use of neural network analysis to classify electroencephalographic patterns against depth of midazolam sedation in intensive care unit patients

The electroencephalographic (EEG) analog signal is complex and cannot easily be described by univariate variables. Clear visual changes in the EEG power spectrum can be present with little or no change in univariate variable values. A method that could produce a single value based on the total data available in the EEG power spectrum would be very useful in monitoring EEG changes. Neural network analysis is a technique that can take multiple inputs and produce a single output value using complicated processing patterns that require training to establish. We examined the usefulness of a series of neural network models to classify 63 EEG patterns against sedation level in 26 mechanically ventilated patients requiring midazolam for long-term sedation. During a stable period of sedation, a 4- to 60-minute period of EEG data was obtained concurrently with a sedation level from 1 (follows commands) to 7 (no or gag response to suctioning of the endotracheal tubc). The EEG power spectrum was divided into equal frequency bands, and the log absolute powers in each of these bands were used as inputs for a series of neural network models. The output target was the sedation level associated with each set of EEG data. Networks were trained on a subset of EEG power/sedation score data pairs, and the ability to classify the remaining data pairs was tested. Using at-test comparison with a random set of sedation levels, we found that trained neural network models classified EEG patterns against sedation level successfully (p<0.001). Fifty to sixty percent of EEG patterns were classified to within a sedation level, as oposed to 30% by chance alone, depending on the network model. Correlation coefficients between actual and predicted sedation levels ranged from 0.1 to 0.5. The best efficiency, as measured by training time and classification success, was obtained when the EEG power spectrum was divided into 2-Hz bands. We conclude that (1) neural network analysis can be used to successfully classify EEG patterns against sedation level when long-term midazolam infusions are used; and (2) a trained neural network may be useful as an on-line monitor of EEG changes, producing a single-output value based on the total EEG data set.