HRV derived time domain index of data serials similarity
to measure anesthetic depth based on ensemble neural network
Quan Liu1, 2, Li Ma1, 2, Ren-Chun Chiu5, Shou-Zen Fan3, Maysam F. Abbod4, Jiann-Shing Shieh 5, 6,*
1 School of Information Engineering, Wuhan University of Technology, Wuhan 430070, China;
2 Key Laboratory of Fiber Optic Sensing Technology and Information Processing, Ministry of Education, Wuhan 430070, China
E-Mail: (Q.L.); (L.M.)
3 Department of Anesthesiology, College of Medicine, National Taiwan University, Taipei 100, Taiwan; E-Mail: ;
4 Department of Electronic and Computer Engineering, Brunel University London, UB8 3PH, UK;
E-Mail: ;
5 Department of Mechanical Engineering, and Innovation Center for Big Data and Digital Convergence, Yuan Ze University, 135, Yuan-Tung Road, Chung-Li 32003, Taiwan;
Email:
6 Center for Dynamical Biomarkers and Translational Medicine, National Central University, Chung-Li 32001, Taiwan.
* Author to whom correspondence should be addressed; E-Mail: ;
Tel.: +886-3-4638800 ext. 2470; Fax: +886-3-4558013.
Abstract: Evaluation of depth of anesthesia (DoA) accurately is always critical in clinical surgery. Traditionally, index derived from electroencephalogram (EEG) plays the dominant role to measure DoA. For lack of ideal approach to quantify the consciousness level when drugs are used like ketamine, nitrous oxide and so on, much many efforts are devoted to optimize the DoA measurement methods. In this study, 110 cases of physiological data are analyzed to predict DoA. We propose a short term index generated by heart rate variability (HRV) of electrocardiogram (ECG) called similarity index (SI). It represents the data difference complexity by observing two consecutive 32s HRV data segments. Compared with expert assessment of consciousness level (EACL) of DoA, it shows strong correlation with anesthetic depth. In order to optimize measure thise effect, artificial neural network (ANN) models are constructed to fit model SI. We also conduct tThe ANN model is developed absed on blind cross validation to overcome the random error of neural network. The results show that Furthermore, the ensemble ANN (EANN) presents better capability accuracy of DoA assessment. Our This research shows thatis HRV related SI parameter can be another an effective method for DoA evaluation. We It is believed that it is possible and meaningful to incorporate the SI to measure the DoA with other methods together if suitablywhen conditions allow.
Keywords: HRV, DoA, similarity index, artificial neural network, EACL
1. Introduction
Anesthesia has been the essentially important procedure in almost all surgeries [1, 2]. General anesthesia is a kind of “artificial sleep” — actually a drug-induced, reversible condition that shows specific behavioral and physiological features like unconsciousness, analgesia and akinesia. Clinically and practically, the routine observations, for example, heart rate, respiration, blood pressure, lacrimation, sweating and so on, mainly assist the doctors to control the anesthetic management smoothly and safely. Nevertheless, clinical post-operation challenges including airway and oxygenation problems, emergence delirium [3] and cognitive dysfunction [4] still exist when patients recover from general anesthesia, especially for the elderly with risk of even stroke and heart attack [5]. Therefore, accurate monitoring of depth of anesthesia (DoA) can contribute to improvements in safety and quality of anesthesia, thus fulfilling patients’ satisfactions.
Because the state of general anesthesia is aroused by the anesthetics functioning in the spinal cord, stem and cortex in brain [6, 7], it is reasonable to get inspiration from electroencephalogram (EEG) patterns [8]. Two main indices derived from EEG are the bispectral index (BIS) (Aspect Medical Systems, Newton, MA, USA) [9] and Entropy family (GE Healthcare, Helsinki, Finland) [10]. The former one is obtained by calculating the adjusted weights on the spectral power, the burst suppression features and the bispectrum of the EEG data, while the latter index is constructed by relating the data degree of disorder (entropy) with consciousness state of patients [10, 11]. Although the EEG-based indices have been applied commercially for nearly 20 years, it is still not yet part the standard anesthesiology practice [12]. Reasons for this situation may be as follows. Firstly, these indices are developed from adult patient cohorts resulting in less accuracy in infants or the younger [13]. Secondly, it cannot give anthe accurate DoA for some specific drug occasions, especially when patients are induced by ketamine and nitrous dioxide [14]. Finally, the EEG signal is sensitive to noise and weak to be acquired purely for real-time computing. Besides, the electrode sensors for EEG data is much more expensive for patients, which may be another main reason to keep it from becoming the regular tool in surgery.
Since EEG has so many disadvantages above [15], all features above pose an urgent need of new ideas to compensate the mainstream methods. Actually, electrocardiogram (ECG) is another very important kind of clinical physiological signal for patients and it is highly recommended to be continuously monitored as international standards for a safe practice of anesthesia [16]. Differential anesthetics affect QT interval of ECG during anesthetic induction [17]. Rhythmic-to-non-rhythmic observations from ECG can also provide the anesthetic information [18]. Due to the heart rate variability (HRV) which varies with the anesthetic procedure[19, 20], HRV is associated with autonomic regulation and highly influenced by general anesthesia [21]. Heartbeat dynamics is significantly correlated with loss of consciousness [22]. As we know, the ECG data has more stability than EEG signal, and it means more resistant to noise with even cheap electrode sensors. So HRV can be one of potential indicator for DoA. Moreover, the inter-individual variations occur normally among people influenced by age, weight, life habits, etc. This effect makes the ECG derived analysis index reflect the individual anesthetic state more specifically, not like EEG-based indices assuming that same index value indicates the same consciousness level for all anesthetics and patients [12]. So it is much worthwhile to undertake some DoA research based on the HRV.
As it iswe known, artificial neural network is a very significant modeling tool in statistics, machine learning and cognitive science [23, 24]. It is similar to biological neural networks in the performing by its units of functions collectively and in parallel, rather than by a clear delineation of subtasks to which individual units are assigned. Neural network models which command the central nervous system and the rest of the brain are part of theoretical neuroscience and computational neuroscience, thus making it optimal for non-linear, distributed, parallel and local processing and adaptation. Ensemble artificial neural network (EANN) is the process of creating multiple models and combining them to produce a desired output, as opposed to creating just one model [25-27]. Normally, an ensemble of models performs better than any individual model because of the ensemble modelsformer average effects [28, 29]. In summary, neural network is a powerful and effective method for data regression and model optimization for non-stationary data. In biomedical fields, it plays an important role for the complex physiological data analysis [30].
In this study, we optimize an indicator index named similarity index (SI)derived from HRV [31].This time domain index is calculated by evaluating the similarity of the statistical distribution of R–R interval measurements in consecutive data blocks.And we use it to do the comparisons with expert assessment of consciousness level (EACL), which are determined by five expert anesthetists’ average evaluation after their observation for patients. Also, optimization is conducted by applying the EANN to build a model for estimation of DoA. Through the time domain SI extraction and EANN modeling by targeting the EACL, we find the result can predict the DoA of whole surgery. The rest of this paper is divided into four sections. Section 2 describes the general anesthesia knowledge and patients participants as well as the data analysis methods. Section 3 presents the processing results and some comparisons with EACL. Section 4 gives the discussion and limitations. TheAnd conclusion is givenprovided in Section 5.
2. Materials and methods
1.
2.
2.1. Ethnic statement
All studies were approved by the Institutional Review Board within ethics committee from the participated hospitals and written informed consent was also obtained from the permission of the patients. All other efforts were made to keep the regular hospital surgeries running smoothly.
2.2. Standard anesthetic procedure
For surgeries, anesthesia is essential and significant. Anesthetic procedures were as described below [32] and are outlined in Fig. 1. It simply utilizes the end-tide gas concentration vs. time to show the anesthesia steps. The routine anesthetic practice consist of four stages: awake, induction, maintenance, emergence (recovery) [33]. General anesthesia performs all stages with monitoring the physiological signals like EEG, ECG, photoplethysmography (PPG) and also the intermittent vital signs of blood pressure (BP), heart rate (HR), pulse rate (PR), oxyhemoglobin saturation by pulse oximetry (SPO2), etc. After electrodes are placed, the medical data can be collected for the whole operation. For analysis of data, data segments of different stages can be obtained to undertake the analysis.
2.3. Data recording
ECG data acquired in this study is collected from patients through chest mounted electrode sensors by MP60 anesthetic monitor machine (Philip, IntelliVue, US). This machine is connected with a recording computer within a real-time software developed by our research team using Borland C++ Builder 6 developing environment kit (Borland company, C++ version 6) to collect the data with sampling rate of 500 Hz. The recording rate of intermittent vital signs like heart rate and blood pressure is one point every 5 seconds. Patients who provide the electrophysiological data come from National Taiwan University Hospital (NTUH) in Taiwan.
2.4. Clinical data collection
Patients’ demographics and clinical information, including height, weight, age, gender, operation time, surgical procedure and anesthetic management are acquired from the anesthetic recording sheets by hospital staff. And also some issues related with the research procedure like body movement, electrotome operation are recorded by research team. Before doing these information collection, the research team will ask the signature consent from the patients. Both the hospital regular recording and the research specific notes are integrated to serve as auxiliary clinical information. Large data mining can be conducted through these valuable referenced data.
2.5. Patient participants
Patients who were scheduled for an elective surgical procedure were recruited from the pre-operative clinic at NTUH in 2015. Eligibility criteria consisted of age, personal willingness and specific operation type. People were not eligible for inclusion in the study if they were (1) under 22 years old, (2) diagnosed with a neurological or cardiovascular disorder, (3) under the local anesthesia, not general. The selective method isare shown in Fig. 2. According to the criteria, more than 122 patients were eligible. However, dozens of cases were unexpectedly not able to be collected. Here, it should be pointed out that when it comes to the failure collection, the cases arewe just ignored them, leading to the number value loss of unsuccessful collection cases. So 122 cases of data were obtained finally. And then 12 cases data were abandoned in next data preprocessing part. All the patients have 7-8 hours of limosis before operations. The general parameters information covers all the 110 patients, but the anesthetic management drugs for individual differ practically. For example, the propofol and fentanyl induction are implemented for patients mostly (n = 100). Details characteristics of subjects are provided in Table 1.
2.6. ECG data preprocessing
(1) Data conditioning
Data conditioning is critical in signal analysis for the DoA, mostly called preprocessing. Because it can overcome the compatibility and non-analysis trouble in advance. It generally consists of data format conversion, artefactartifact removal, data rearrangement. Due to the data collection storage limitation, we employ the txt file format. Before applying the following algorithm analysis, we transformed all the patients’ data type available to MATLAB analysis with appropriate specific name. Moreover, the intact data series needs to be checked visually. It means some cases may have technique failure and clinic problems for data recording. Besides, during the induction stage of the anesthesia, the noise might be strong enough to be impossible to extract the R peak of ECG wave in some cases. 12 cases like this are rejected. All the data rearrangement work should be undertaken to promote the analysis.
(2) Expert Assessment of Consciousness Level
As a whole, the EACL means the patients’ DoA quantifictation derived from anesthetic experts with rich clinical anesthetic experiences. It is well known that there is no absolutely accurate standard index for symbolizing the patients’ anesthetic state and in clinical surgery the anesthetist usually control the anesthesia procedure based on the experience by observing the ECG, BP, HR, SPO2, etc. So we discuss to accept the assessment quantifictation from the experienced anesthesiologists. The procedure in Fig. 3 is described as follows: Firstly, while acquiring the physiological data, research nurses keep observing the state of patients to record the events and signs in detail and carefully, which possibly have relationship with ‘the state of anesthetic depth, for example, time of the induction and extubation, drugs administered time and their dose, minimum alveolar concentration (MAC) values and so on. Then, five experienced anesthesiologists need to make a continuous curve individually and independently to draw the changes of ‘the state of anesthetic depth’ of patients for the whole duration of operation based on the hospital formal anesthesia sheets and our research extra records mentioned previously. In order to be consistent with BIS, the curve is predefined the range 0 - 100 from unconsciousbrain death to fully awake (40 - 60 represents an appropriate anesthesia level during surgery). Finally, because the original curve was completed by hand drawing, it is digitalized by a web-software (ANSYS, webplotdigtilzer, US) [34] and resampled with a frequency of 0.2 Hz with MATLAB interpolation method to be same to BIS index. Thus, the result should be treated as expert assessment of conscious level. However, each anesthesiologist with different experience may have a different standard on EACL, therefore, in order to eliminate consciousness level error to the least, the assessmentwe can average the data from five anesthesiologists is averaged. Fig. 4 gives aone case example of EACL from five doctors. It is more convinced to use the mean value as a real gold standard of DoA.
2.7. Data analysis
A. Similarity Index Algorithm of HRV