Using convolutional neural network online estimators for predicting pain-level variability enables predictive control of anesthesia

This work presents the results of convolutional neural networks trained in various procedures on data recordings from conscious, communicative patients recovering in post-surgery units. A prototype device measures non-invasively the bioimpedance by placing 3M patch electrodes in the hand palm of the patient. From the time signals acquired, a time-frequency algorithm is applied to extract the spectrogram 3D graphical model every second, during 4-6 hours. The aim was to detect whether artificial intelligence (AI) tools may offer a possible solution to obtain prediction of increasing pain levels in an objective way as to avoid the subjectivity of self-evaluation (currently golden standard in practice). The results obtained indicate a 73% accuracy with 86% successful detection of pain correlation to self-reported pain level. Specific training procedures have shown they are able to predict changes in pain levels which could be subsequently used in control algorithms, especially those prediction-based, which could greatly benefit from such complementary information. The discussion in this paper also puts the use of AI tools within the context of depth of anesthesia regulation, as AI tools greatly aid to minimize uncertainty in the loop dynamics.