OBJECTIVE: Awareness with recall is a rare but serious complication of general anaesthesia with an incidence ranging from 0.1%-0.7%. In the absence of a reliable depth-of-anaesthesia monitor, attempts have been made to predict awareness from intraoperative haemodynamic monitoring data, with little success. Artificial neural networks can sometimes detect relationships between input and output variables even when conventional methods fail. Therefore, we subjected standard intraoperative monitoring data to both artificial neural models and conventional statistical methods in an attempt to predict awareness with recall. METHODS: Anaesthesia records from 33 patients with awareness and 510 patients without awareness were collected. Summary data (mean, maximum, and minimum) of end-tidal carbon dioxide concentration, arterial blood oxygen saturation, systolic and diastolic blood pressure, and heart rate were calculated for each patient. These data were subjected to an analysis by artificial neural networks and by Poisson regression. RESULTS: The two best neural models both had sensitivity and specificity of 23% and 98%, respectively. The models have high specificity, and in view of the low incidence of awareness, a high negative predictive value. The prediction probabilities P(k) (SE) for the best neural models were 0.66 (0.08) and 0.60 (0.10), respectively. In the Poisson regression, there were significant differences in systolic and diastolic blood pressures and heart rate between patients with and without awareness. CONCLUSIONS: A prediction indicating awareness by the network is very suggestive of true awareness and recall. Blood pressure and heart rate are significantly higher on average in patients with awareness than in patients without. In an individual patient, however, none of our artificial neural models can detect awareness sufficiently reliably.
OBJECTIVE: Awareness with recall is a rare but serious complication of general anaesthesia with an incidence ranging from 0.1%-0.7%. In the absence of a reliable depth-of-anaesthesia monitor, attempts have been made to predict awareness from intraoperative haemodynamic monitoring data, with little success. Artificial neural networks can sometimes detect relationships between input and output variables even when conventional methods fail. Therefore, we subjected standard intraoperative monitoring data to both artificial neural models and conventional statistical methods in an attempt to predict awareness with recall. METHODS: Anaesthesia records from 33 patients with awareness and 510 patients without awareness were collected. Summary data (mean, maximum, and minimum) of end-tidal carbon dioxide concentration, arterial blood oxygen saturation, systolic and diastolic blood pressure, and heart rate were calculated for each patient. These data were subjected to an analysis by artificial neural networks and by Poisson regression. RESULTS: The two best neural models both had sensitivity and specificity of 23% and 98%, respectively. The models have high specificity, and in view of the low incidence of awareness, a high negative predictive value. The prediction probabilities P(k) (SE) for the best neural models were 0.66 (0.08) and 0.60 (0.10), respectively. In the Poisson regression, there were significant differences in systolic and diastolic blood pressures and heart rate between patients with and without awareness. CONCLUSIONS: A prediction indicating awareness by the network is very suggestive of true awareness and recall. Blood pressure and heart rate are significantly higher on average in patients with awareness than in patients without. In an individual patient, however, none of our artificial neural models can detect awareness sufficiently reliably.