Background and objective: The aim was to train artificial neural nets to predict the recovery of a neuromuscular block during general anaesthesia. It was assumed that the initial/early neuromuscular recovery data with the simultaneously measured physical variables as inputs into a well-trained back-propagation neural net would enable the net to predict a rough estimate of the remaining recovery time.
Methods: Spontaneous recovery from neuromuscular block (electrically evoked electromyographic train-of-four responses) were recorded with the following variables known to affect the block: multiple minimum alveolar concentration, end-tidal CO2 concentration, and peripheral and central temperature.
Results: The mean prediction errors, mean absolute prediction errors, root-mean-squared prediction errors and correlation coefficients of all the nets were significantly better than those of average-based predictions used in the study. The root-mean-squared prediction error of the net – employing minimum alveolar concentrations from the whole recovery period (the recovery time from E2/E1 = 0.30 to E4/E1 = 0.75; E1 = first response of train-of-four, E2 = second response of train-of-four, etc.) – were significantly smaller than those of other nets, or the same net employing minimum alveolar concentrations only from the initial recovery period (from E2/E1 = 0.30 to E4/E1 = 0.25).
Conclusions: Neural nets could predict individual recovery times from the neuromuscular block significantly better than the average-based method used here, which was supposed to be more accurate than guesses by any clinician. The minimum alveolar concentration was the only monitored variable that influenced the recovery rate, but it did not aid neural net prediction.
Email your librarian or administrator to recommend adding this journal to your organisation's collection.