Mining the hidden dysrhythmia ‐ can machines get smarter at defining the anaesthetised state?

Before the inception of commercial depth-of-anaesthesia monitors using electrophysiological parameters from the brain, such as processed electroencephalography (EEG) and evoked potentials, we used to evaluate the balance between anaesthesia and surgical stress using multimodal information derived from several physiological monitors, including the electrocardiogram (ECG), indirect or direct blood pressure, pulse oximetry, body temperature, peripheral nerve stimulation, as well as pharmacokinetic assumptions. In addition, we used our own senses to see the colour of the skin and the blood; to touch the body and feel its temperature and whether it was sweating, and to assess peripheral circulation; to assess the pupils for anaesthetic/narcotic effects; and even to talk to a patient to determine response by body movement and postoperative recall 1, 2. Although a major part of this evaluation is based on scientific logic, the rest rather belongs to some kind of ‘art’. Put another way, such art might be rephrased as ‘fuzzy logic’ 3, which enables clinical decision-making using a mixture of positive, negative, and grey-zone signs of various physiological phenomena. Even in the present era of high-tech, computer-intensive EEG monitoring, we are still routinely required to make such decisions by integrating heterogeneous information during our clinical practice. Such an art remains indispensable in modern anaesthetic management, where multiple drugs are often administered to attempt to attain a balanced anaesthetic state, consisting of unconsciousness (or amnesia), antinociception, immobility, suppression of the stress response, and preserved homeostasis. Part of the art of managing general anaesthesia or sedation may be easily explained, but some may not, and cannot even be documented in any detail: clinicians can therefore be said to be using their ‘sixth sense’. As practitioners, we usually estimate the balance of anaesthesia ourselves, and titrate anaesthetic agents and related drugs routinely, often without using objective, validated parameters from a modern EEG-based depth-of-anaesthesia monitor. This sixth sense tries to make the most of the multimodal information displayed by various monitors, mostly circulatory, respiratory, and other autonomic variables. The apparent absence of stress-related autonomic and escape-like motor responses to surgical stimulus makes us assume that an anaesthetised patient should probably be unconscious and pain-free. However, such experience-based decisions, and even an EEG monitor, cannot ever exclude possible ‘dysanaesthesia’, i.e. preserved responsiveness under apparently sufficient surgical anaesthesia and amnesia 4. Both the sixth sense and EEG remain imperfect in titrating anaesthesia. In this issue of Anaesthesia, Kenwrite et al. 5 try to put some evidence-based shape to the sixth sense. They programmed a computer to memorise unique patterns among several kinds of processed physiological parameters from the heart, lungs, skin, and the peripheral pulses during the awake and anaesthetised state. Then the computer, using a kind of newly developed ‘machine-learning’ algorithm, succeeded in distinguishing between the awake and the anaesthetised state, and between propofol and sevoflurane anaesthesia, with high precision (above 90%). Specifically, they recruited 27 surgical patients and gave either propofol (n = 12; targeted plasma concentration at 3 μg.ml−1) or sevoflurane (n = 15; end-tidal concentration 2%) to patients who were breathing spontaneously. Before and during anaesthesia, they recorded 1) ECG R-peaks, 2) chest cage movement, 3) temperatures at the right hand and the right foot, 4) skin conductivity at the right thumb, and 5) the pressure-generated pulse wave at the right index finger, using a custom-made monitoring device at a high sampling rate of 1200 Hz. They subsequently dissected each rhythm into multiple ‘wavelets’ and calculated power distribution across different frequency ranges. They further analysed patterns of interaction among different modalities by examining phase synchronisation of wavelets. The wavelets showed unique patterns of power spectrum and synchronisation among different modalities, such as the cardiac and respiratory rhythms, depending on the three states: wakefulness, propofol- and sevoflurane-anaesthesia. Such information from wavelet analysis was used to define the unique ‘fingerprint’ of each state using a ‘distance-based classification technique’ 5. The question that needs to be asked is: what would such fingerprints mean from a physiological point of view? As we already know, the heart and lungs are functionally inseparable. During spontaneous breathing, inspiration and expiration cause rhythmic changes in the preload of the heart, which in turn affects the cardiac output and rhythm by a centrally mediated autonomic reflex by way of afferents from the carotid sinus baroreceptor and efferent autonomic innervations of the heart. As such, all the different systems, including circulatory, respiratory, and autonomic systems, interact or cross-talk with each other in an inseparable way. Part of such interaction could potentially be estimated by analysing wavelets across various frequency ranges 6, 7. General anaesthesia might potentially affect such inseparable relationships in a certain way that our senses, or even the sixth sense, cannot even notice. Can the computers do better than our art, the sixth sense, in estimating anaesthetic states by detection of such ‘dysrhythmias’? The fingerprints measured in this study included much more information than we could usually glean from standard monitors; that is, the extents of variability and synchronisation of the heart and respiratory rates. The cardiac rhythm has been known to show decreased variability during anaesthesia 8, as well as in cardiac disease and other critical illness 9. The authors found decreased variability in heart and respiratory rates, and skin conductivity changes during general anaesthesia compared with those during wakefulness. That is, general anaesthesia seems to have flattened the natural fluctuations commonly observed in the human vegetative system. Decreased variability might be a common signature belonging to the anaesthetised state, in contrast with wakefulness. In addition, general anaesthesia decreased the complexity or randomness of the interaction between the circulatory and respiratory systems usually observed under physiological conditions, i.e. increased cardiopulmonary synchronisation. Such changes in bodily rhythms might be regarded as ‘hidden dysrhythmias’ induced by general anaesthesia, which might potentially turn out to be a ‘biomarker’ of the anaesthetised state. The authors’ success in developing an algorithm to classify such unnoticed, multi-faceted fingerprints should be one pioneering example in developing an ideal depth-of-anaesthesia monitor in the future. This is amongst the first studies published that have attempted to adapt machine learning to develop an algorithm to estimate the anaesthetised state using multimodal information from the whole body, with the exception of the electrical activity of the brain. The question arises why the authors opted not to use an EEG as part of their algorithm? One reason may be that EEG monitors may not yet reliably provide parameters to indicate ‘depth of anaesthesia’, which should involve much more than the bivariate states of ‘conscious or unconscious’, or ‘probability of awareness’ 10. Another group in Germany have adopted a similar approach, but included the EEG. Schneider et al. 11 tried to establish a depth-of-anaesthesia monitor using multimodal information from the EEG and other standard physiological monitoring. They developed a computational algorithm using an ‘Adaptive Neuro Fuzzy Inference System’ for the computers to distinguish between various anaesthetic regimens and estimate depth of anaesthesia. They reported that their novel algorithm, named ‘Anaesthesia Multimodal Index of Consciousness’, differentiated multiple behavioural states of anaesthesia: wakefulness; unconsciousness; surgical anaesthesia; and deep anaesthesia characterised by burst suppression on the EEG. They also showed that this was more efficient than either standard or bispectral index monitoring. On the other hand, Kenwrite et al. 5 did not use either EEG or behavioural evaluation in their study. It is possible that they designed this to be a feasibility study to explore the possibility of evaluating anaesthesia using information from the body alone. In this sense, their study remains only preliminary. Real-life general anaesthesia utilises multiple drugs whose actions are balanced, including narcotic, local anaesthetic, neuromuscular blocking, and cardiovascular agents that potentially affect autonomic tone. Respiration is also frequently controlled by mechanical ventilation, which may totally mask its spontaneous trigger. Behavioural planes ranging from wakefulness to deep anaesthesia, calibrated with EEG analysis, should also result in various patterns of autonomic wavelets. In the future, as the authors are aware 5, the present algorithm should be tested and calibrated under numerous different conditions. All the data were derived from control awake and single anaesthetic agent conditions, which can only be comparable with sedation or hypnosis with opioid/regional analgesia but without neuromuscular blockade. Most actual anaesthetic regimens, however, are made up with a balanced mixture of several agents with different roles, i.e. hypnotics, analgesics, and immobilisers, and may often require controlled mechanical ventilation. Such anaesthesia in the real world would highly complicate the coherent interactions among the different systems of the body. The state of general anaesthesia does not necessarily implicate the binary states of consciousness or a linear measure of anaesthetic depth, but rather encompasses biologically heterogeneous conditions where multiple agents act on various receptors in multiple organs and systems of the body to a variable extent 12. The ‘discriminatory classification’ algorithm was developed from data recorded from 27 patients, and it was also validated on the same dataset 5. Therefore, it is not surprising that the authors demonstrated a high success rate. A separate validation study is required, performed on a different dataset from a different cohort of subjects/patients, who are administered multiple, different doses of anaesthetics. It would also be important to make it a double-blind study to both patients and the scientists analysing the data. No doubt, the authors might also want to test the dose-response and behaviour-response characteristics of the algorithm. The anaesthetised state should also be described using behavioural parameters as well as calculated effect-site or end-tidal concentrations of agents. Without such validation, the authors cannot conclude the true feasibility of the presented methodology. Such further studies for this pioneering work should be keenly awaited. The present study raises an important issue in the current era of EEG-weighted monitoring. Loss of consciousness or memory is no doubt a definite prerequisite of general anaesthesia. Although a general anaesthetic principally targets the brain, it affects all the other organ systems of the body. Modern balanced anaesthetic techniques mandate not only loss of consciousness but also protection of the patient's body and mind from surgical stress and its consequences, that are mediated through multiple bodily systems, including the nervous, circulatory, hormonal systems, cytokines, and even anaesthetics themselves. Therefore, it is entirely conceivable that we should indeed analyse and integrate multimodal information from the whole body, consisting of multiple organs and systems, to determine the most appropriate anaesthetised state for the best surgical outcome, and whole life after surgery, of patients. The present study by Kenwrite et al. 5 surely indicates there is a need for such an approach for the next generation of depth-of-anaesthesia monitors. The present work was supported by Grants-in-Aid for Scientific Research (no. 26460695 to J.K.), Japan. No conflicts of interest declared.