TITLE: A SIMPLE ARTIFICIAL NEURAL NETWORK (ANN) CAN HELP PREDICT RESIDUAL NEUROMUSCULAR BLOCKADE

AUTHORS: John G. Laffey, MB, John F. Boylan., MB, Alan J. McShane, MB
AFFILIATION: St. Vincent's University Hospital, Dublin, Ireland.

INTRODUCTION: Residual neuromuscular blockade (RNMB) following surgery remains a significant problem¹, even following the use of shorter acting neuromuscular blockers. Human error in assessing PNS data is very common.² ANN's consist of computer software designed to mimic multiple inputs and nonlinear interactions and are increasingly used to examine complex data.³ We tested the hypothesis that an ANN-based analysis would enhance the prediction of RNMB when compared with human decision making.
METHODS: After IRB approval and informed consent, 40 patients were recruited. Transduced twitch height data and TOF values were measured from induction of anaesthesia up to tracheal extubation, by an independent observer. Care was provided by residents in training, under the direction of a staff anesthesiologist; all were blinded to the data being collected. A TOF of 0.7 at tracheal extubation was defined as residual curarization. Using a back-propagation ANN model (Neuralyst 1.40, Cheshire Engineering, CA) with three layers (an input layer, an output layer and a hidden layer containing four nodes), a predictive model was developed by presenting data as numeric input variables. Analysis was limited to the combined predictive value of two simple, readily available, clinical measurements (1) number of twitch responses at the time of reversal agent administration and (2) time elapsed from neostigmine admini- stration until extubation. Momentum and learning rates were examined over a range of values before finally using rates of 0.05. Training and test tolerances were set at 0.2. Using the jackknife method, the network was trained 40 times using each patient in turn as the test dataset.
RESULTS: Of the 40 patients, 27 had TOF 7 at the time of tracheal extubation. Since all anesthetists believed that patients were adequately reversed at extubation, this indicates a sensitivity of zero (i.e. no RNMB diagnosed in 27) and specificity of 1 (i.e. adequate reversal diagnosed in 13). The performance of the training and test phases of the ANN were essentially identical, with the ANN correctly classifying 38 of 39 during training and 38 of 40 correctly classified during testing. The test phase of the ANN performed with a sensitivity of 0.96 (c2 = 63, P < 0.0001) and a specificity of 0.93 (P = 0.54, Fisher's test). ANN predictive performance was markedly superior to the detection ability of the residents providing anesthesia care.
DISCUSSION: A simple ANN was retrospectivelyable to predict the likelihood of curarization with greater accuracy than individual anaesthetists' clinical assessment. This suggests that given limited and imprecise data, ANN-based prediction of simple drug pharmacodynamic relationships can equal of surpass real-time human assessment.
REFERENCES:

Anesthesiology 1979; 50: 539-41.

Anesthesiology 1985; 63: 440-3. Lancet 1995; 346:1075-9.