300
IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 24, NO. 2, FEBRUARY 2016
An Intelligent Decision System for Intraoperative Somatosensory Evoked Potential Monitoring Bi Fan, Han-Xiong Li, Fellow, IEEE, and Yong Hu, Senior Member, IEEE
Abstract—Somatosensory evoked potential (SEP) is a useful, noninvasive technique widely used for spinal cord monitoring during surgery. One of the main indicators of a spinal cord injury is the drop in amplitude of the SEP signal in comparison to the nominal baseline that is assumed to be constant during the surgery. However, in practice, the real-time baseline is not constant and may vary during the operation due to nonsurgical factors, such as blood pressure, anaesthesia, etc. Thus, a false warning is often generated if the nominal baseline is used for SEP monitoring. In current practice, human experts must be used to prevent this false warning. However, these well-trained human experts are expensive and may not be reliable and consistent due to various reasons like fatigue and emotion. In this paper, an intelligent decision system is proposed to improve SEP monitoring. First, the least squares support vector regression and multi-support vector regression models are trained to construct the dynamic baseline from historical data. Then a control chart is applied to detect abnormalities during surgery. The effectiveness of the intelligent decision system is evaluated by comparing its performance against the nominal baseline model by using the real experimental datasets derived from clinical conditions. Index Terms—Control chart, intelligent decision system, least squares support vector regression (LS-SVR), multi-support vector regression (M-SVR), somatosensory evoked potential (SEP).
I. INTRODUCTION
S
PINAL surgery is an effective way of correcting deformities of the spine, but also entails the risk of damage to the spinal cord. The use of intraoperative spinal cord monitoring can minimize such risk and has become a routine procedure in the spinal surgery [1]–[3]. Previous studies have shown that changes in somatosensory evoked potential (SEP) can indicate neurological deficits in the spinal cord [4]–[8]. To detect the neurological function deficit and prevent further spinal cord injury, the current SEP monitoring technique is widely applied intraoperatively. The amplitude and latency measured in Manuscript received July 15, 2014; revised April 03, 2015 and August 06, 2015; accepted August 30, 2015. Date of publication September 23, 2015; date of current version February 18, 2016. This work was supported in part by the GRF of Hong Kong under Grant CityU: 11207714 and by a seeding fund from the University of Hong Kong. B. Fan is with the Department of System Engineering and Engineering Management, City University of Hong Kong, Hong Kong, and also with State Key Laboratory of High Performance Complex Manufacturing, Central South University, Changsha, Hunan 410083, China (e-mail: fanbia@ gmail.com). H.-X. Li is with the Department of System Engineering and Engineering Management, City University of Hong Kong, Hong Kong (e-mail: [email protected]). Y. Hu is with the Department of Orthopedics, The University of Hong Kong, Hong Kong (e-mail: [email protected]). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TNSRE.2015.2477557
the waveform of SEP signal are compared with the baseline reference to detect existence of a potential spinal cord injury, by monitoring changes in SEP signal such as a reduction in amplitude (more than or equal to 50%) or a delay (more than or equal to 10% increase in latency) [5], [8]–[11]. The SEP signal is not only impacted by surgical factors such as the surgical operations that may impose risks to the spine. It is also impacted by nonsurgical factors [12]–[14], such as anaesthetic agents, physiological parameters, etc. Different concentrations and anaesthetic agents have a significant effect on cortical SEP waveform, leading to an increase in latency and decrease in amplitude, and even causing the loss of the SEP signal [13], [15]. Various physiological parameters, like blood pressures (systolic blood pressure and diastolic blood pressure), heart rate, partial pressure of carbon dioxide in artery, and body temperature, may directly affect the SEP signal [16]–[18]. Furthermore, the false positive results in spinal cord monitoring due to nonsurgical factors have been reported, suggesting that there is an abnormal neurological response when there is actually not any injury to the spinal cord [13], [19], [20]. This can lead to an unnecessary interruption of the surgery. Therefore, the traditional method may not be reliable due to the inappropriate assumption that SEP monitoring is only related to the spine. Considering the effects of the nonsurgical factors, some studies have been made to improve the reliability of SEP monitoring by adjusting the baseline. One suggested that the baseline reference should be determined in relation to the normal variation [12]. The other found the reliability of SEP monitoring was improved by adjusting the baseline reference in anticipation of the SEP changes subject to dosage of anaesthesias and physiological variables [17]. However, the selection of the optimal baseline is very subjective and difficult to command by specialists. Some false positive cases were caused by inexperience of the human monitoring [21]. Until now, this is still an unsolved challenge in clinical settings. Since the amplitude usually presents higher variation than latency during the intraoperative monitoring [10], the amplitude is chosen here for detection. In this paper, an intelligent decision system is proposed to assist human experts for a reliable SEP monitoring. Due to the limited physical knowledge about the volatility of the SEP signal, a data-driven method is firstly employed to model the dynamic baseline by utilizing a small amount of experimental data. The least squares support vector regression (LS-SVR) is trained to be the prediction model [22], [23]. As the clinical data contains noise, a multi-support vector regression (M-SVR) is also proposed to provide a more reliable prediction model than LS-SVR. It is inspired by the probabilistic classification method in [24].
1534-4320 © 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
FAN et al.: INTELLIGENT DECISION SYSTEM FOR INTRAOPERATIVE SEP MONITORING
301
Fig. 1. Human-assisted SEP monitoring.
Fig. 2. Intelligent decision system for SEP monitoring.
After the dynamic baseline model is developed, the control chart [25] is further employed to detect abnormal SEP signals that fall outside the lower control limit. Experiments on the clinical dataset show that the proposed intelligent decision system outperforms the nominal baseline method especially in the false positive dataset. Moreover, the proposed M-SVR is more reliable than the LS-SVR as the prediction model. The rest of this paper is organized as follows. The trauma detection in SEP monitoring is discussed in Section II. The construction of the intelligent decision system is discussed in Section III. Section IV summarizes the experiments. Finally, some conclusions are made in Section V. II. TRAUMA DETECTION IN SEP MONITORING SEP plays an important role in spinal surgery. It can reflect the integrity and functional status of sensory nerve pathways between the stimulated site and the recording site. In traditional SEP monitoring, the nominal baseline is firstly generated from the SEP signals recorded after the spine was exposed but before instrumentation loading and deformity correction. Then SEP signals are monitored continuously during surgery. When the observed SEP signal has a 50% decrease in amplitude compared with the nominal baseline, it indicates the risk of trauma in the surgery. The operation needs to be suspended to check the status of the spinal function by waking up the patient. However, false positive cases are often reported when this traditional SEP monitoring method is used due to the effects of nonsurgical factors as mentioned previously. Human experts are often employed to adjust the baseline manually by experience and to decide if a spinal trauma has occurred in the surgery as shown in Fig. 1. Unfortunately, human experts have two
major disadvantages: high training cost and inconsistency due to human fatigue and emotions. At this point, no one has developed an intelligence based decision system for SEP monitoring. The introduction of an intelligent decision system could significantly assist the human experts to provide more reliable SEP monitoring. Such a system is feasible given the existence of historical data. III. INTELLIGENT DECISION SYSTEM In the proposed intelligent decision system, a prediction model is trained to construct a dynamic baseline that considers the effects of the nonsurgical factors in the noisy environment. Then, the medical decision of whether there exists a trauma can be reliably made with the help of the control chart widely used in quality engineering. The intelligent decision system is developed for the first time for the SEP monitoring. This novel system consists of two parts as shown in Fig. 2: one is the prediction model to estimate the dynamic baseline using the popular data-driven method LS-SVR [27]. Moreover, a new method M-SVR is developed to consider the impact of noisy data when building the prediction model. It is also a novel application to use the control chart method [25], [35] to assist the decision making in SEP monitoring. Though some of these parts are traditional, however, integration of these parts into a new intelligent system to solve difficulties in SEP monitoring is novel. The clinical dataset is recorded as , where is the th input vector to represent the nonsurgical factors and is the th output to represent the amplitude of the current SEP signal. is paired with as the th sample . The number of total samples is .
302
IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 24, NO. 2, FEBRUARY 2016
Fig. 3. Multi-support vector regression model.
A. Least Squares Support Vector Regression To model the dynamic baseline, the prediction model can be established by the least squares support vector regression algorithm as follows [27]:
The aforementioned equations can be equivalently written as the following linear system after elimination of and : (5)
(1) and , is number of dimension. is the mapping function from the primal feature space to the high-dimensional feature space [26], is the weight vector, and is the bias term. Given the training data, function (1) can be estimated by solving the optimization problem as follows [23], [27]: where
subject to
(2)
are assumed to be independent and identically where distributed (i.i.d.) random errors with . denotes the regularization parameter for the tradeoff between approximation accuracy and model complexity [28], [29], and the remaining notation is identical to that of (1). By using Lagrange multipliers, the solutions of (2) can be obtained. The Lagrangian function is given by
with
, , , is the identity matrix of size . Based on Mercer's theorem, one obtains matrix from (6)
where is a kernel function. The radial basis kernel function (RBF) is used with bandwidth . The regularization parameter and bandwidth are selected by grid search on five -fold cross validation accuracy. With and obtained from (5), the result of LS-SVR model for function estimation is as follows: (7) where
is the future input.
B. Multi-Support Vector Regression (3) where denotes Lagrange multipliers. The conditions for optimality are given by
(4)
In clinical environments, strong noise exists and can greatly affect SEP monitoring performance. To handle the noisy data better, an M-SVR is proposed as an alternative prediction model, as shown in Fig. 3. First, data is clustered. Then, subsets are formed by sampling with replacement in all clusters. The number of samplings from each cluster is decided by the cluster size proportionally. It is assumed that the data far away from the cluster center suffers from noise more than others. Then, the data close to the cluster center has a higher chance of being sampled. The subsets formed are calculated by a group of LS-SVR to produce the final output. Each LS-SVR is trained on one subset to be a “local expert” for the corresponding subdomain. The combination of the outputs of all LS-SVRs will lead to a more reliable decision.
FAN et al.: INTELLIGENT DECISION SYSTEM FOR INTRAOPERATIVE SEP MONITORING
a) Clustering: Given dataset , it will be partitioned into clusters by K-means clustering algorithm [30]–[32]. These clusters are denoted by . The data are grouped into different clusters according to their similarity. First, the initial number of clusters is determined by the experts. Then, the centers of these clusters are selected randomly. According to the Nearest-Neighbor rule, each is assigned to its closest cluster center by Euclidean distance (8) is the th cluster center with . Then each where cluster center is updated by calculating the mean of all samples assigned
303
where .. .
..
.. .
.
..
.
and is the size of original dataset. Usually, is a nonsquare matrix as the number of the regression models in (12) is much less than the number of training samples . Therefore, it is not easy to obtain a solution of . However, the solution of (13) can still be found as follows: (14)
(9)
is the Moore-Penrose generalized inverse of matrix where [33], [34]
is the number of samples in cluster . where Next, the clustering cost function is calculated as follows: (10)
Equations (8)–(10) are reiterated several times until the clustering cost function is converged. b) Resampling: There are subsets randomly sampled from the original data with replacement and each has equal size. In each subset, first, the sample size in each cluster is decided by the ratio of the cluster to the whole dataset. Then, resampling is carried out in each cluster according to the distance degree of the data to its cluster center. The distance degree of to the center of cluster is denoted as (11) After repeating this resample procedure times, subsets are obtained as . c) Training of Multimodel: For each subset , the least squares support vector regression model is trained, in the same way as mentioned in Section III-A. A group of regression models is obtained as follows: (12) where , is the number of samples in subset and . d) Decision combination: With these regression functions, all trained LS-SVRs should be integrated in an appropriate way. It is equivalent to estimate the vector for the following linear system: (13)
if if
is nonsingular is nonsingular.
Then we have if if
is nonsingular is nonsingular. (15) And the final output function of M-SVR is if if
is nonsingular is nonsingular. (16)
C. Control Chart Based Decision Making A control chart is applied to identify abnormal cases for SEP monitoring. It is a very useful method in monitoring process capability and performance. This method is based on the assumption that the mean of data (without abnormal things) follows normal distribution [25], [35]. For good and safe process, subsequent data collected should fall within three standard deviations of the mean. If an abnormal result happens, this data will fall outside of control limits (usually called “3 sigma limits” in quality engineering). This kind of control chart has been widely and successfully used in manufacturing industry to detect the potential defects [36], [37]. It is the first time to apply for the SEP monitoring. Since the training dataset contains no abnormal case, the prediction results offered by the dynamic baseline model can be viewed as trustworthy references of the normal status of the SEP signal. Although the dynamic baseline model can provide accurate predictions, residuals between the predicted values and the observed values of SEP signal still exist. Thus, monitoring abnormal cases of SEP has been transformed into monitoring abnormal values of residuals [38]. The control chart is applied in the residual monitoring. The lower bound serves as the threshold for detecting abnormal status during the surgery, which is indicated by the points falling outside the lower bound. If the points
304
IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 24, NO. 2, FEBRUARY 2016
fall outside the upper bound, this often refers to a good status of the spinal cord. The residual between the observed value and the value predicted by the dynamic baseline model is expressed as
TABLE I DESCRIPTION OF FEATURES USED IN EXPERIMENTS
(17) The center line of the control chart is denoted as . The variation of is represented by . The formulations are as follows: (18)
(19) where is the number of data points in the training dataset. The upper and lower limits of the control chart monitoring are constructed as follows: (20) (21) (22)
TABLE II PERFORMANCE COMPARISON OF INTELLIGENT DECISION SYSTEM NOMINAL BASELINE METHOD
AND
, and represent the upper control limit, where center line, and the lower control limit of residuals, respectively; is the number of data points sampled for monitoring [35]. To monitor the values of the future SEP signal, the control chart is modified as follows: (23) (24) (25) where presents the predicted SEP baseline for the test dataset, and other notation is the same as in (20)–(22). Based on this modified control chart, abnormal values of SEP will be detected if the observed values of SEP fall outside the lower control limit. The lower control limit is the decision boundary of the proposed intelligent decision system. IV. EXPERIMENTAL STUDIES The proposed intelligent decision system was evaluated on a clinical dataset. Its performance was compared with the traditional SEP monitoring method that relies on the nominal baseline. Furthermore, under the framework of the intelligent decision system, the performances of least squares support vector regression and M-SVR were also compared. A. Experimental Subjects Three different types of surgeries were included in the experiments. Each surgery contained many samples collected at different time instants and each sample had 23 features as shown in
Table I. Initial SEP signals were recorded after the exposure of the spine. Isoflurane was used for inhalational anesthesia. The three types of surgeries are described as follows. 1) Successful case without any spinal cord injury: Ten successful surgeries that did not result in spinal cord injury were used for training and testing. From these surgeries with total 165 samples, 158 samples were chosen after removing several samples with missing feature values. 2) False-Positive case: False-positive cases were surgeries that were interrupted by human experts due to the occurrence of false-positive samples during the operation. There were four such surgeries, which contributed 72 samples for testing. 3) Trauma case: Trauma cases are due to surgical operation that had to be interrupted by a human expert. One trauma case with 14 samples was used for testing. B. Results Since there were only ten cases of successful surgery, one case was selected for testing and the remaining nine cases were left for training. The experiment was repeated ten trials for a more reliable outcome. At each trial, a different case was chosen for testing and the remaining nine cases were used for training. The testing results of the ten trials were averaged and are shown in Table II. Similarly, all four false-positive cases and the trauma case were tested at each trial.
FAN et al.: INTELLIGENT DECISION SYSTEM FOR INTRAOPERATIVE SEP MONITORING
305
Fig. 4. Comparison of intelligent decision system and nominal baseline method.
These SEP monitoring methods detect the abnormal SEP signal and give the warning. For the nominal baseline method, abnormal SEP signal is defined below the baseline; while for the intelligent method, abnormal SEP signal is below the lower control limit, which is the decision boundary. Warning rate measures the ratio of warning to the total number of samples evaluated. In successful and false-positive cases, since there is no abnormal SEP signal, the lower the warning rate, the greater detection ability of the method. On the contrary, there are many abnormal SEP signals in the trauma case, so the higher warning rate denotes better detection ability in the trauma case. Table II summarizes the testing results of the intelligent decision system and nominal baseline method. Obviously, the proposed intelligent decision system outperforms the nominal baseline method on all datasets when either LS-SVR or M-SVR is employed as the prediction model. Especially, the M-SVR produces the best performance. However, the warning rate of nominal baseline method is close to LS-SVR and M-SVR on successful case. In contrast, the performances of LS-SVR and M-SVR outweigh the nominal baseline method on false-positive and trauma cases. To illustrate their performances, one trail on three selected cases is shown in Fig. 4. There are 11 samples in successful case, 18 samples in false-positive case and 14 samples in trauma
case. These three cases have different sample sizes since different surgery has different surgery time. The vertical axis is the amplitude of SEP signal (in microvolts) and the horizontal axis is the index of the evaluated samples. Firstly, LS-SVR and M-SVR are developed on the same successful training dataset. Then, they are applied to the three selected cases. The decision boundaries of the nominal baseline method are also demonstrated in Fig. 4. As the nominal baseline is decided by the physiological status of the patient, different patients have different nominal baselines. Therefore, the nominal baselines are different in the three cases. Compared with nominal baseline method, the LS-SVR and M-SVR have a better performance on the three cases and present the trend of SEP signal. More importantly, the proposed method greatly outperforms the nominal baseline method in the false positive case. Since the false positive signal causes unnecessary interruption during the clinical practice, the improvement achieved by the proposed intelligent decision system could have a great impact on SEP monitoring during clinical operations. In addition to the effectiveness of both prediction models, the intelligent decision system performs differently when using LS-SVR or M-SVR as the prediction model as shown in Table II and Fig. 4. It is interesting to note that the M-SVR is slightly better than LS-SVR in the successful case and false
306
IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 24, NO. 2, FEBRUARY 2016
V. CONCLUSION
TABLE III PERFORMANCE COMPARISON OF M-SVR AND LS-SVR
positive case. The possible reason is that the influence of noise is taken into consideration by M-SVR. Here, a further step is taken to evaluate these two prediction models. Two common metrics are used to evaluate their performances of modeling accuracy and variability [38], [39]: the mean square error (MSE) and standard deviation of square error (SD of SE), which are defined as follows: MSE
(26)
SD of SE
(27) The results in terms of the above metrics on the successful, false positive and trauma datasets are shown in Table III. The M-SVR produces smaller MSE and SD of SE than LS-SVR on the normal and false positive datasets. Though LS-SVR outweighs M-SVR in the trauma dataset, M-SVR provides better performance in general. Since the control chart is based on the prediction model, the better the prediction model is, the better performance the intelligent decision system will provide. This could explain the better performance of the intelligent decision system with M-SVR than with SVR. The conventional intraoperative SEP monitoring technique is based on a stable baseline to be recorded either during the preoperation or after the surgical area exposure. Because the nonsurgical factors are not taken into consideration, it would be more difficult to determine whether the changes in the SEP actually indicate a surgical related functional change or just are affected by nonsurgical factors in the nominal baseline method [40] for the conventional monitoring. In contrast, the proposed method considers the nonsurgical factors and offers a new intelligent decision system for SEP monitoring. With a clear separation of nonsurgical and surgical effects on changes of SEP, the proposed method can greatly improve the reliability of intraoperative monitoring by decreasing false positive alerts while increasing sensitivity in true positive alerts as shown in Table II. The improvements come from two aspects. One is the dynamic baseline obtained by using the data-driven methods, like LS-SVR or M-SVR. The changes of SEP caused by nonsurgical factors will be captured, which makes the dynamic baseline more reliable as the reference to decrease false positive alerts. The other is the control chart to monitor the changes of SEP caused by surgical factors. It increases the sensitivity to detect the abnormal SEP signal in true positive cases.
An intelligent decision system has been proposed for SEP monitoring in spinal surgeries to reduce false warnings and more accurately detect spinal trauma. According to historical data, a dynamical baseline model was first constructed by two methods: LS-SVR and M-SVR, respectively. With the dynamic baseline model as the prediction model, the control chart was employed to detect abnormal SEP signals that fell outside the lower control limit. The decision system was evaluated on the clinical dataset compared with the nominal baseline method. The results suggest that the proposed decision system has better performance, especially in the false positive cases, and may be more effective in the trauma case. Furthermore, the comparison of the LS-SVR and M-SVR has shown that the decision system with M-SVR performed better than the system with LS-SVR in both successful and false positive datasets. REFERENCES [1] M. R. Nuwer, E. G. Dawson, L. G. Carlson, L. E. Kanim, and J. E. Sherman, “Somatosensory evoked potential spinal cord monitoring reduces neurologic deficits after scoliosis surgery: Results of a large multicenter survey,” Electroencephalogr. Clin. Neurophysiology/Evoked Potentials Section, vol. 96, no. 1, pp. 6–11, Jan. 1995. [2] M. R. Nuwer, “Spinal cord monitoring with somatosensory techniques,” J. Clin. Neurophysiol., vol. 15, no. 3, pp. 183–193, May 1998. [3] T. Hammett, B. Boreham, N. A. Quraishi, and S. M. H. Mehdian, “Intraoperative spinal cord monitoring during the surgical correction of scoliosis due to cerebral palsy and other neuromuscular disorders,” Eur. Spine J., vol. 22, no. 1, pp. 38–41, Jan. 2013. [4] M. Imai, Y. Harada, Y. Atsuta, Y. Takemitsu, and T. Iwahara, “Automated spinal cord monitoring for spinal surgery,” Spinal Cord, vol. 27, no. 3, pp. 204–211, Jun. 1989. [5] A. A. Gonzalez, D. Jeyanandarajan, C. Hansen, G. Zada, and P. C. Hsieh, “Intraoperative neurophysiological monitoring during spine surgery: A review,” Neurosurgical Focus, vol. 27, no. 4, Oct. 2009. [6] T. Hammett, A. Harris, B. Boreham, and S. M. H. Mehdian, “Surgical correction of scoliosis in Rett syndrome: Cord monitoring and complications,” Eur. Spine J., vol. 23, no. 1, pp. 72–75, Apr. 2014. [7] C. D. Glover and N. P. Carling, “Neuromonitoring for scoliosis surgery,” Anesthesiol. Clinics, vol. 32, no. 1, pp. 101–114, Mar. 2014. [8] P. P. Jankowski, R. A. O'Brien, G. B. Cornwall, and W. R. Taylor, “Intraoperative neurophysiology monitoring,” Minimally Invasive Spine Surgery, pp. 43–53, Mar. 2014. [9] I. Chung, J. A. Glow, V. Dimopoulos, M. Sami Walid, H. F. Smisson, K. W. Johnston, J. S. Robinson, and A. A. Grigorian, “Upper-limb somatosensory evoked potential monitoring in lumbosacral spine surgery: A prognostic marker for position-related ulnar nerve injury,” Spine J., vol. 9, no. 4, pp. 287–295, Apr. 2009. [10] Y. Hu, K. D. K. Luk, W. W. Lu, and J. C. Y. Leong, “Application of time-frequency analysis to somatosensory evoked potential for intraoperative spinal cord monitoring,” J. Neurol., Neurosurgery Psychiatry, vol. 74, no. 1, pp. 82–87, 2003. [11] L. Wu, Y. Qiu, W. Ling, and Q. Shen, “Change pattern of somatosensory-evoked potentials after occlusion of segmental vessels: Possible indicator for spinal cord ischemia,” Eur. Spine J., vol. 15, no. 3, pp. 335–340, 2006. [12] J. H. Owen, “Intraoperative stimulation of the spinal cord for prevention of spinal cord injury,” Advances Neurol., vol. 63, pp. 271–288, Feb. 1993. [13] R. W. Keith, J. L. Stambough, and S. H. Awender, “Somatosensory cortical evoked potentials: A review of 100 cases of intraoperative spinal surgery monitoring,” J. Spinal Disorders Techniques, vol. 3, no. 3, pp. 220–226, Sep. 1990. [14] A. R. Møller, “Monitoring Somatosensory Evoked Potentials,” in Intraoperative Neurophysiological Monitoring. New York, NY, USA: Springer, 2011, pp. 93–122. [15] M. Banoub, J. E. Tetzlaff, and A. Schubert, “Pharmacologic and physiologic influences affecting sensory evoked potentials: Implications for perioperative monitoring,” Anesthesiology, vol. 99, no. 3, pp. 716–737, Sep. 2003.
FAN et al.: INTELLIGENT DECISION SYSTEM FOR INTRAOPERATIVE SEP MONITORING
[16] M. Seyal and B. Mull, “Mechanisms of signal change during intraoperative somatosensory evoked potential monitoring of the spinal cord,” J. Clin. Neurophysiol., vol. 19, no. 5, pp. 409–415, Oct. 2002. [17] H. Cui, K. D. K. Luk, and Y. Hu, “Effects of physiological parameters on intraoperative somatosensory evoked potential monitoring: Results from a multifactor analysis,” Med. Sci. Monitor, vol. 15, no. 5, pp. 226–230, Apr. 2009. [18] J. J. Jana, N. Vaid, and J. Shanbhag, “Effect of total intravenous anaesthesia on intraoperative monitoring of Cochlear implant function in paediatric patients,” Cochlear Implants Int., vol. 14, no. 3, pp. 169–173, Jun. 2013. [19] K. J. Noonan, T. Walker, J. R. Feinberg, M. Nagel, W. Didelot, and R. Lindseth, “Factors related to false-versus true-positive neuromonitoring changes in adolescent idiopathic scoliosis surgery,” Spine, vol. 27, no. 8, pp. 825–830, Apr. 2002. [20] H. Wiedemayer, I. E. Sandalcioglu, W. Armbruster, J. Regel, H. Schaefer, and D. Stolke, “False negative findings in intraoperative SEP monitoring: Analysis of 658 consecutive neurosurgical cases and review of published reports,” J. Neurol., Neurosurgery Psychiatry, vol. 75, no. 2, pp. 280–286, Dec. 2004. [21] S. L. Papastefanou, L. M. Henderson, N. J. Smith, A. Hamilton, and J. K. Webb, “Surface electrode somatosensory-evoked potentials in spinal surgery: Implications for indications and practice,” Spine, vol. 25, no. 19, pp. 2467–2472, Oct. 2000. [22] J. A. K. Suykens, T. Van Gestel, J. De Brahanter, B. De Moor, and J. Vandewalle, Least Squares Support Vector Machines. Singapore: World Scientific, 2002. [23] C. K. Qi, H. X. Li, X. X. Zhang, X. C. Zhao, S. Y. Li, and F. Gao, “Time/space-separation-based SVM modeling for nonlinear distributed parameter processes,” Industr. Eng. Chemistry Res., vol. 50, no. 1, pp. 332–341, Nov. 2010. [24] H. X. Li, J. L. Yang, G. Zhang, and B. Fan, “Probabilistic support vector machines for classification of noise affected data,” Inform. Sci., vol. 221, pp. 60–71, Feb. 2013. [25] D. Montgomery, Introduction to Statistical Quality Control. Hoboken, NJ, USA: Wiley, 2009. [26] K. De Brabanter K, J. De Brabanter, J. A. K. Suykens, and B. De Moor, “Approximate confidence and prediction intervals for least squares support vector regression,” IEEE Trans. Neural Networks, vol. 22, no. 1, pp. 110–120, Jan. 2011. [27] I. Goethals, K. Pelckmans, J. A. Suykens, and B. De Moor, “Subspace identification of Hammerstein systems using least squares support vector machines,” IEEE Trans. Automat. Contr., vol. 50, no. 10, pp. 1509–1519, Oct. 2005. [28] T. Hastie, R. Tibshirani, J. Friedman, and J. Franklin, “The elements of statistical learning: Data mining, inference and prediction,” Math. Intelligencer, vol. 27, no. 2, pp. 83–85, Nov. 2005. [29] V. N. Vapnik, “An overview of statistical learning theory,” IEEE Trans. Neural Networks, vol. 10, no. 5, pp. 988–999, Sep. 1999. [30] S. Ben-David, “A framework for statistical clustering with constant time approximation algorithms for k-median and k-means clustering,” Machine Learning, vol. 66, no. 2, pp. 243–257, Mar. 2007. [31] J. A. Hartigan and M. A. Wong, “Algorithm AS 136: A k-means clustering algorithm,” J. Roy. Statistical Soc., Series C (Applied Statistics), vol. 28, no. 1, pp. 100–108, Aug. 1979. [32] K. P. Murphy, Machine Learning: A Probabilistic Perspective. Cambridge, MA, USA: The MIT Press, 2012. [33] G. B. Huang, H. Zhou, X. Ding, and R. Zhang, “Extreme learning machine for regression and multiclass classification,” IEEE Trans. Systems, Man, Cybernetics-Part B: Cybernetics, vol. 42, no. 2, pp. 513–529, Mar. 2012. [34] C. R. Rao and S. K. Mitra, Generalized Inverse of Matrices and Its Applications. New York, NY, USA: Wiley, 1971. [35] T. P. Ryan, Statistical Methods for Quality Improvement. Hoboken, NJ, USA: Wiley, 2011.
307
[36] S. L. Hwang, J. T. Lin, G. F. Liang, Y. J. Yau, T. C. Yenn, and C. C. Hsu, “Application control chart concepts of designing a pre-alarm system in the nuclear power plant control room,” Nuclear Eng. Design, vol. 238, no. 12, pp. 3522–3527, 2008. [37] W. Hachicha and A. Ghorbel, “A survey of control-chart pattern-recognition literature (1991–2010) based on a new conceptual classification scheme,” Computers Industr. Eng., vol. 63, no. 1, pp. 204–222, 2012. [38] Z. J. Zhang and A. Kusiak, “Monitoring wind turbine vibration based on SCADA data,” Trans. ASME-N-J. Solar Energy Eng., vol. 134, no. 2, pp. 1–12, Feb. 2012, 021004. [39] R. Caruana and A. Niculescu-Mizil, “An empirical comparison of supervised learning algorithms,” in Proc. 23rd Int. Conf. Machine Learning ACM, Jun. 2006, pp. 161–168. [40] K. D. K. Luk, Y. Hu, Y. W. Wong, and J. C. Y. Leong, “Variability of somatosensory-evoked potentials in different stages of scoliosis surgery,” Spine, vol. 24, no. 17, pp. 1799–1804, Sep. 1999.
Bi Fan received the B.E. degree in automation from Ocean University of China, in 2002, the M.E. degree in mechatronics from University of Shanghai for Science and Technology, in 2010, and the Ph.D. degree in system engineering and engineering management from City University of Hong Kong, in 2014. His research interests include neural engineering, machine learning and nonlinear system.
Han-Xiong Li (F’10) received the B.E. degree in aerospace engineering from the National University of Defense Technology, China, in 1982, the M.E. degree in electrical engineering from Delft University of Technology, The Netherlands, in 1991, and the Ph.D. degree in electrical engineering from the University of Auckland, New Zealand, in 1997. He is a Professor in the Department of Systems Engineering and Engineering Management, the City University of Hong Kong. His current research interests are in system intelligence and control, process design and control integration, distributed parameter systems. Prof. Li serves as Associate Editor of IEEE TRANSACTIONS ON CYBERNETICS, and IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS (since 2009). He was awarded the Distinguished Young Scholar award (overseas) by the China National Science Foundation in 2004, a Chang Jiang Professorship by the Ministry of Education, China, in 2006, and a National Professorship in China Thousand Talents Program in 2010.
Yong Hu (SM’12) received the B.Sc. and M.Sc degrees in the biomedical engineering from Tianjin University, Tianjin, China, in 1985 and 1988, respectively, and the Ph.D. degree from The University of Hong Kong in 1999. He is currently an Associate Professor and Director of Neural Engineering and Clinical Electrophysiology Laboratory, the Department of Orthopaedics and Traumatology, The University of Hong Kong. His research interests include neural engineering, clinical electrophysiology, biomedical signal measurement and processing.
IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 24, NO. 2, FEBRUARY 2016
An Intelligent Decision System for Intraoperative Somatosensory Evoked Potential Monitoring Bi Fan, Han-Xiong Li, Fellow, IEEE, and Yong Hu, Senior Member, IEEE
Abstract—Somatosensory evoked potential (SEP) is a useful, noninvasive technique widely used for spinal cord monitoring during surgery. One of the main indicators of a spinal cord injury is the drop in amplitude of the SEP signal in comparison to the nominal baseline that is assumed to be constant during the surgery. However, in practice, the real-time baseline is not constant and may vary during the operation due to nonsurgical factors, such as blood pressure, anaesthesia, etc. Thus, a false warning is often generated if the nominal baseline is used for SEP monitoring. In current practice, human experts must be used to prevent this false warning. However, these well-trained human experts are expensive and may not be reliable and consistent due to various reasons like fatigue and emotion. In this paper, an intelligent decision system is proposed to improve SEP monitoring. First, the least squares support vector regression and multi-support vector regression models are trained to construct the dynamic baseline from historical data. Then a control chart is applied to detect abnormalities during surgery. The effectiveness of the intelligent decision system is evaluated by comparing its performance against the nominal baseline model by using the real experimental datasets derived from clinical conditions. Index Terms—Control chart, intelligent decision system, least squares support vector regression (LS-SVR), multi-support vector regression (M-SVR), somatosensory evoked potential (SEP).
I. INTRODUCTION
S
PINAL surgery is an effective way of correcting deformities of the spine, but also entails the risk of damage to the spinal cord. The use of intraoperative spinal cord monitoring can minimize such risk and has become a routine procedure in the spinal surgery [1]–[3]. Previous studies have shown that changes in somatosensory evoked potential (SEP) can indicate neurological deficits in the spinal cord [4]–[8]. To detect the neurological function deficit and prevent further spinal cord injury, the current SEP monitoring technique is widely applied intraoperatively. The amplitude and latency measured in Manuscript received July 15, 2014; revised April 03, 2015 and August 06, 2015; accepted August 30, 2015. Date of publication September 23, 2015; date of current version February 18, 2016. This work was supported in part by the GRF of Hong Kong under Grant CityU: 11207714 and by a seeding fund from the University of Hong Kong. B. Fan is with the Department of System Engineering and Engineering Management, City University of Hong Kong, Hong Kong, and also with State Key Laboratory of High Performance Complex Manufacturing, Central South University, Changsha, Hunan 410083, China (e-mail: fanbia@ gmail.com). H.-X. Li is with the Department of System Engineering and Engineering Management, City University of Hong Kong, Hong Kong (e-mail: [email protected]). Y. Hu is with the Department of Orthopedics, The University of Hong Kong, Hong Kong (e-mail: [email protected]). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TNSRE.2015.2477557
the waveform of SEP signal are compared with the baseline reference to detect existence of a potential spinal cord injury, by monitoring changes in SEP signal such as a reduction in amplitude (more than or equal to 50%) or a delay (more than or equal to 10% increase in latency) [5], [8]–[11]. The SEP signal is not only impacted by surgical factors such as the surgical operations that may impose risks to the spine. It is also impacted by nonsurgical factors [12]–[14], such as anaesthetic agents, physiological parameters, etc. Different concentrations and anaesthetic agents have a significant effect on cortical SEP waveform, leading to an increase in latency and decrease in amplitude, and even causing the loss of the SEP signal [13], [15]. Various physiological parameters, like blood pressures (systolic blood pressure and diastolic blood pressure), heart rate, partial pressure of carbon dioxide in artery, and body temperature, may directly affect the SEP signal [16]–[18]. Furthermore, the false positive results in spinal cord monitoring due to nonsurgical factors have been reported, suggesting that there is an abnormal neurological response when there is actually not any injury to the spinal cord [13], [19], [20]. This can lead to an unnecessary interruption of the surgery. Therefore, the traditional method may not be reliable due to the inappropriate assumption that SEP monitoring is only related to the spine. Considering the effects of the nonsurgical factors, some studies have been made to improve the reliability of SEP monitoring by adjusting the baseline. One suggested that the baseline reference should be determined in relation to the normal variation [12]. The other found the reliability of SEP monitoring was improved by adjusting the baseline reference in anticipation of the SEP changes subject to dosage of anaesthesias and physiological variables [17]. However, the selection of the optimal baseline is very subjective and difficult to command by specialists. Some false positive cases were caused by inexperience of the human monitoring [21]. Until now, this is still an unsolved challenge in clinical settings. Since the amplitude usually presents higher variation than latency during the intraoperative monitoring [10], the amplitude is chosen here for detection. In this paper, an intelligent decision system is proposed to assist human experts for a reliable SEP monitoring. Due to the limited physical knowledge about the volatility of the SEP signal, a data-driven method is firstly employed to model the dynamic baseline by utilizing a small amount of experimental data. The least squares support vector regression (LS-SVR) is trained to be the prediction model [22], [23]. As the clinical data contains noise, a multi-support vector regression (M-SVR) is also proposed to provide a more reliable prediction model than LS-SVR. It is inspired by the probabilistic classification method in [24].
1534-4320 © 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
FAN et al.: INTELLIGENT DECISION SYSTEM FOR INTRAOPERATIVE SEP MONITORING
301
Fig. 1. Human-assisted SEP monitoring.
Fig. 2. Intelligent decision system for SEP monitoring.
After the dynamic baseline model is developed, the control chart [25] is further employed to detect abnormal SEP signals that fall outside the lower control limit. Experiments on the clinical dataset show that the proposed intelligent decision system outperforms the nominal baseline method especially in the false positive dataset. Moreover, the proposed M-SVR is more reliable than the LS-SVR as the prediction model. The rest of this paper is organized as follows. The trauma detection in SEP monitoring is discussed in Section II. The construction of the intelligent decision system is discussed in Section III. Section IV summarizes the experiments. Finally, some conclusions are made in Section V. II. TRAUMA DETECTION IN SEP MONITORING SEP plays an important role in spinal surgery. It can reflect the integrity and functional status of sensory nerve pathways between the stimulated site and the recording site. In traditional SEP monitoring, the nominal baseline is firstly generated from the SEP signals recorded after the spine was exposed but before instrumentation loading and deformity correction. Then SEP signals are monitored continuously during surgery. When the observed SEP signal has a 50% decrease in amplitude compared with the nominal baseline, it indicates the risk of trauma in the surgery. The operation needs to be suspended to check the status of the spinal function by waking up the patient. However, false positive cases are often reported when this traditional SEP monitoring method is used due to the effects of nonsurgical factors as mentioned previously. Human experts are often employed to adjust the baseline manually by experience and to decide if a spinal trauma has occurred in the surgery as shown in Fig. 1. Unfortunately, human experts have two
major disadvantages: high training cost and inconsistency due to human fatigue and emotions. At this point, no one has developed an intelligence based decision system for SEP monitoring. The introduction of an intelligent decision system could significantly assist the human experts to provide more reliable SEP monitoring. Such a system is feasible given the existence of historical data. III. INTELLIGENT DECISION SYSTEM In the proposed intelligent decision system, a prediction model is trained to construct a dynamic baseline that considers the effects of the nonsurgical factors in the noisy environment. Then, the medical decision of whether there exists a trauma can be reliably made with the help of the control chart widely used in quality engineering. The intelligent decision system is developed for the first time for the SEP monitoring. This novel system consists of two parts as shown in Fig. 2: one is the prediction model to estimate the dynamic baseline using the popular data-driven method LS-SVR [27]. Moreover, a new method M-SVR is developed to consider the impact of noisy data when building the prediction model. It is also a novel application to use the control chart method [25], [35] to assist the decision making in SEP monitoring. Though some of these parts are traditional, however, integration of these parts into a new intelligent system to solve difficulties in SEP monitoring is novel. The clinical dataset is recorded as , where is the th input vector to represent the nonsurgical factors and is the th output to represent the amplitude of the current SEP signal. is paired with as the th sample . The number of total samples is .
302
IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 24, NO. 2, FEBRUARY 2016
Fig. 3. Multi-support vector regression model.
A. Least Squares Support Vector Regression To model the dynamic baseline, the prediction model can be established by the least squares support vector regression algorithm as follows [27]:
The aforementioned equations can be equivalently written as the following linear system after elimination of and : (5)
(1) and , is number of dimension. is the mapping function from the primal feature space to the high-dimensional feature space [26], is the weight vector, and is the bias term. Given the training data, function (1) can be estimated by solving the optimization problem as follows [23], [27]: where
subject to
(2)
are assumed to be independent and identically where distributed (i.i.d.) random errors with . denotes the regularization parameter for the tradeoff between approximation accuracy and model complexity [28], [29], and the remaining notation is identical to that of (1). By using Lagrange multipliers, the solutions of (2) can be obtained. The Lagrangian function is given by
with
, , , is the identity matrix of size . Based on Mercer's theorem, one obtains matrix from (6)
where is a kernel function. The radial basis kernel function (RBF) is used with bandwidth . The regularization parameter and bandwidth are selected by grid search on five -fold cross validation accuracy. With and obtained from (5), the result of LS-SVR model for function estimation is as follows: (7) where
is the future input.
B. Multi-Support Vector Regression (3) where denotes Lagrange multipliers. The conditions for optimality are given by
(4)
In clinical environments, strong noise exists and can greatly affect SEP monitoring performance. To handle the noisy data better, an M-SVR is proposed as an alternative prediction model, as shown in Fig. 3. First, data is clustered. Then, subsets are formed by sampling with replacement in all clusters. The number of samplings from each cluster is decided by the cluster size proportionally. It is assumed that the data far away from the cluster center suffers from noise more than others. Then, the data close to the cluster center has a higher chance of being sampled. The subsets formed are calculated by a group of LS-SVR to produce the final output. Each LS-SVR is trained on one subset to be a “local expert” for the corresponding subdomain. The combination of the outputs of all LS-SVRs will lead to a more reliable decision.
FAN et al.: INTELLIGENT DECISION SYSTEM FOR INTRAOPERATIVE SEP MONITORING
a) Clustering: Given dataset , it will be partitioned into clusters by K-means clustering algorithm [30]–[32]. These clusters are denoted by . The data are grouped into different clusters according to their similarity. First, the initial number of clusters is determined by the experts. Then, the centers of these clusters are selected randomly. According to the Nearest-Neighbor rule, each is assigned to its closest cluster center by Euclidean distance (8) is the th cluster center with . Then each where cluster center is updated by calculating the mean of all samples assigned
303
where .. .
..
.. .
.
..
.
and is the size of original dataset. Usually, is a nonsquare matrix as the number of the regression models in (12) is much less than the number of training samples . Therefore, it is not easy to obtain a solution of . However, the solution of (13) can still be found as follows: (14)
(9)
is the Moore-Penrose generalized inverse of matrix where [33], [34]
is the number of samples in cluster . where Next, the clustering cost function is calculated as follows: (10)
Equations (8)–(10) are reiterated several times until the clustering cost function is converged. b) Resampling: There are subsets randomly sampled from the original data with replacement and each has equal size. In each subset, first, the sample size in each cluster is decided by the ratio of the cluster to the whole dataset. Then, resampling is carried out in each cluster according to the distance degree of the data to its cluster center. The distance degree of to the center of cluster is denoted as (11) After repeating this resample procedure times, subsets are obtained as . c) Training of Multimodel: For each subset , the least squares support vector regression model is trained, in the same way as mentioned in Section III-A. A group of regression models is obtained as follows: (12) where , is the number of samples in subset and . d) Decision combination: With these regression functions, all trained LS-SVRs should be integrated in an appropriate way. It is equivalent to estimate the vector for the following linear system: (13)
if if
is nonsingular is nonsingular.
Then we have if if
is nonsingular is nonsingular. (15) And the final output function of M-SVR is if if
is nonsingular is nonsingular. (16)
C. Control Chart Based Decision Making A control chart is applied to identify abnormal cases for SEP monitoring. It is a very useful method in monitoring process capability and performance. This method is based on the assumption that the mean of data (without abnormal things) follows normal distribution [25], [35]. For good and safe process, subsequent data collected should fall within three standard deviations of the mean. If an abnormal result happens, this data will fall outside of control limits (usually called “3 sigma limits” in quality engineering). This kind of control chart has been widely and successfully used in manufacturing industry to detect the potential defects [36], [37]. It is the first time to apply for the SEP monitoring. Since the training dataset contains no abnormal case, the prediction results offered by the dynamic baseline model can be viewed as trustworthy references of the normal status of the SEP signal. Although the dynamic baseline model can provide accurate predictions, residuals between the predicted values and the observed values of SEP signal still exist. Thus, monitoring abnormal cases of SEP has been transformed into monitoring abnormal values of residuals [38]. The control chart is applied in the residual monitoring. The lower bound serves as the threshold for detecting abnormal status during the surgery, which is indicated by the points falling outside the lower bound. If the points
304
IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 24, NO. 2, FEBRUARY 2016
fall outside the upper bound, this often refers to a good status of the spinal cord. The residual between the observed value and the value predicted by the dynamic baseline model is expressed as
TABLE I DESCRIPTION OF FEATURES USED IN EXPERIMENTS
(17) The center line of the control chart is denoted as . The variation of is represented by . The formulations are as follows: (18)
(19) where is the number of data points in the training dataset. The upper and lower limits of the control chart monitoring are constructed as follows: (20) (21) (22)
TABLE II PERFORMANCE COMPARISON OF INTELLIGENT DECISION SYSTEM NOMINAL BASELINE METHOD
AND
, and represent the upper control limit, where center line, and the lower control limit of residuals, respectively; is the number of data points sampled for monitoring [35]. To monitor the values of the future SEP signal, the control chart is modified as follows: (23) (24) (25) where presents the predicted SEP baseline for the test dataset, and other notation is the same as in (20)–(22). Based on this modified control chart, abnormal values of SEP will be detected if the observed values of SEP fall outside the lower control limit. The lower control limit is the decision boundary of the proposed intelligent decision system. IV. EXPERIMENTAL STUDIES The proposed intelligent decision system was evaluated on a clinical dataset. Its performance was compared with the traditional SEP monitoring method that relies on the nominal baseline. Furthermore, under the framework of the intelligent decision system, the performances of least squares support vector regression and M-SVR were also compared. A. Experimental Subjects Three different types of surgeries were included in the experiments. Each surgery contained many samples collected at different time instants and each sample had 23 features as shown in
Table I. Initial SEP signals were recorded after the exposure of the spine. Isoflurane was used for inhalational anesthesia. The three types of surgeries are described as follows. 1) Successful case without any spinal cord injury: Ten successful surgeries that did not result in spinal cord injury were used for training and testing. From these surgeries with total 165 samples, 158 samples were chosen after removing several samples with missing feature values. 2) False-Positive case: False-positive cases were surgeries that were interrupted by human experts due to the occurrence of false-positive samples during the operation. There were four such surgeries, which contributed 72 samples for testing. 3) Trauma case: Trauma cases are due to surgical operation that had to be interrupted by a human expert. One trauma case with 14 samples was used for testing. B. Results Since there were only ten cases of successful surgery, one case was selected for testing and the remaining nine cases were left for training. The experiment was repeated ten trials for a more reliable outcome. At each trial, a different case was chosen for testing and the remaining nine cases were used for training. The testing results of the ten trials were averaged and are shown in Table II. Similarly, all four false-positive cases and the trauma case were tested at each trial.
FAN et al.: INTELLIGENT DECISION SYSTEM FOR INTRAOPERATIVE SEP MONITORING
305
Fig. 4. Comparison of intelligent decision system and nominal baseline method.
These SEP monitoring methods detect the abnormal SEP signal and give the warning. For the nominal baseline method, abnormal SEP signal is defined below the baseline; while for the intelligent method, abnormal SEP signal is below the lower control limit, which is the decision boundary. Warning rate measures the ratio of warning to the total number of samples evaluated. In successful and false-positive cases, since there is no abnormal SEP signal, the lower the warning rate, the greater detection ability of the method. On the contrary, there are many abnormal SEP signals in the trauma case, so the higher warning rate denotes better detection ability in the trauma case. Table II summarizes the testing results of the intelligent decision system and nominal baseline method. Obviously, the proposed intelligent decision system outperforms the nominal baseline method on all datasets when either LS-SVR or M-SVR is employed as the prediction model. Especially, the M-SVR produces the best performance. However, the warning rate of nominal baseline method is close to LS-SVR and M-SVR on successful case. In contrast, the performances of LS-SVR and M-SVR outweigh the nominal baseline method on false-positive and trauma cases. To illustrate their performances, one trail on three selected cases is shown in Fig. 4. There are 11 samples in successful case, 18 samples in false-positive case and 14 samples in trauma
case. These three cases have different sample sizes since different surgery has different surgery time. The vertical axis is the amplitude of SEP signal (in microvolts) and the horizontal axis is the index of the evaluated samples. Firstly, LS-SVR and M-SVR are developed on the same successful training dataset. Then, they are applied to the three selected cases. The decision boundaries of the nominal baseline method are also demonstrated in Fig. 4. As the nominal baseline is decided by the physiological status of the patient, different patients have different nominal baselines. Therefore, the nominal baselines are different in the three cases. Compared with nominal baseline method, the LS-SVR and M-SVR have a better performance on the three cases and present the trend of SEP signal. More importantly, the proposed method greatly outperforms the nominal baseline method in the false positive case. Since the false positive signal causes unnecessary interruption during the clinical practice, the improvement achieved by the proposed intelligent decision system could have a great impact on SEP monitoring during clinical operations. In addition to the effectiveness of both prediction models, the intelligent decision system performs differently when using LS-SVR or M-SVR as the prediction model as shown in Table II and Fig. 4. It is interesting to note that the M-SVR is slightly better than LS-SVR in the successful case and false
306
IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 24, NO. 2, FEBRUARY 2016
V. CONCLUSION
TABLE III PERFORMANCE COMPARISON OF M-SVR AND LS-SVR
positive case. The possible reason is that the influence of noise is taken into consideration by M-SVR. Here, a further step is taken to evaluate these two prediction models. Two common metrics are used to evaluate their performances of modeling accuracy and variability [38], [39]: the mean square error (MSE) and standard deviation of square error (SD of SE), which are defined as follows: MSE
(26)
SD of SE
(27) The results in terms of the above metrics on the successful, false positive and trauma datasets are shown in Table III. The M-SVR produces smaller MSE and SD of SE than LS-SVR on the normal and false positive datasets. Though LS-SVR outweighs M-SVR in the trauma dataset, M-SVR provides better performance in general. Since the control chart is based on the prediction model, the better the prediction model is, the better performance the intelligent decision system will provide. This could explain the better performance of the intelligent decision system with M-SVR than with SVR. The conventional intraoperative SEP monitoring technique is based on a stable baseline to be recorded either during the preoperation or after the surgical area exposure. Because the nonsurgical factors are not taken into consideration, it would be more difficult to determine whether the changes in the SEP actually indicate a surgical related functional change or just are affected by nonsurgical factors in the nominal baseline method [40] for the conventional monitoring. In contrast, the proposed method considers the nonsurgical factors and offers a new intelligent decision system for SEP monitoring. With a clear separation of nonsurgical and surgical effects on changes of SEP, the proposed method can greatly improve the reliability of intraoperative monitoring by decreasing false positive alerts while increasing sensitivity in true positive alerts as shown in Table II. The improvements come from two aspects. One is the dynamic baseline obtained by using the data-driven methods, like LS-SVR or M-SVR. The changes of SEP caused by nonsurgical factors will be captured, which makes the dynamic baseline more reliable as the reference to decrease false positive alerts. The other is the control chart to monitor the changes of SEP caused by surgical factors. It increases the sensitivity to detect the abnormal SEP signal in true positive cases.
An intelligent decision system has been proposed for SEP monitoring in spinal surgeries to reduce false warnings and more accurately detect spinal trauma. According to historical data, a dynamical baseline model was first constructed by two methods: LS-SVR and M-SVR, respectively. With the dynamic baseline model as the prediction model, the control chart was employed to detect abnormal SEP signals that fell outside the lower control limit. The decision system was evaluated on the clinical dataset compared with the nominal baseline method. The results suggest that the proposed decision system has better performance, especially in the false positive cases, and may be more effective in the trauma case. Furthermore, the comparison of the LS-SVR and M-SVR has shown that the decision system with M-SVR performed better than the system with LS-SVR in both successful and false positive datasets. REFERENCES [1] M. R. Nuwer, E. G. Dawson, L. G. Carlson, L. E. Kanim, and J. E. Sherman, “Somatosensory evoked potential spinal cord monitoring reduces neurologic deficits after scoliosis surgery: Results of a large multicenter survey,” Electroencephalogr. Clin. Neurophysiology/Evoked Potentials Section, vol. 96, no. 1, pp. 6–11, Jan. 1995. [2] M. R. Nuwer, “Spinal cord monitoring with somatosensory techniques,” J. Clin. Neurophysiol., vol. 15, no. 3, pp. 183–193, May 1998. [3] T. Hammett, B. Boreham, N. A. Quraishi, and S. M. H. Mehdian, “Intraoperative spinal cord monitoring during the surgical correction of scoliosis due to cerebral palsy and other neuromuscular disorders,” Eur. Spine J., vol. 22, no. 1, pp. 38–41, Jan. 2013. [4] M. Imai, Y. Harada, Y. Atsuta, Y. Takemitsu, and T. Iwahara, “Automated spinal cord monitoring for spinal surgery,” Spinal Cord, vol. 27, no. 3, pp. 204–211, Jun. 1989. [5] A. A. Gonzalez, D. Jeyanandarajan, C. Hansen, G. Zada, and P. C. Hsieh, “Intraoperative neurophysiological monitoring during spine surgery: A review,” Neurosurgical Focus, vol. 27, no. 4, Oct. 2009. [6] T. Hammett, A. Harris, B. Boreham, and S. M. H. Mehdian, “Surgical correction of scoliosis in Rett syndrome: Cord monitoring and complications,” Eur. Spine J., vol. 23, no. 1, pp. 72–75, Apr. 2014. [7] C. D. Glover and N. P. Carling, “Neuromonitoring for scoliosis surgery,” Anesthesiol. Clinics, vol. 32, no. 1, pp. 101–114, Mar. 2014. [8] P. P. Jankowski, R. A. O'Brien, G. B. Cornwall, and W. R. Taylor, “Intraoperative neurophysiology monitoring,” Minimally Invasive Spine Surgery, pp. 43–53, Mar. 2014. [9] I. Chung, J. A. Glow, V. Dimopoulos, M. Sami Walid, H. F. Smisson, K. W. Johnston, J. S. Robinson, and A. A. Grigorian, “Upper-limb somatosensory evoked potential monitoring in lumbosacral spine surgery: A prognostic marker for position-related ulnar nerve injury,” Spine J., vol. 9, no. 4, pp. 287–295, Apr. 2009. [10] Y. Hu, K. D. K. Luk, W. W. Lu, and J. C. Y. Leong, “Application of time-frequency analysis to somatosensory evoked potential for intraoperative spinal cord monitoring,” J. Neurol., Neurosurgery Psychiatry, vol. 74, no. 1, pp. 82–87, 2003. [11] L. Wu, Y. Qiu, W. Ling, and Q. Shen, “Change pattern of somatosensory-evoked potentials after occlusion of segmental vessels: Possible indicator for spinal cord ischemia,” Eur. Spine J., vol. 15, no. 3, pp. 335–340, 2006. [12] J. H. Owen, “Intraoperative stimulation of the spinal cord for prevention of spinal cord injury,” Advances Neurol., vol. 63, pp. 271–288, Feb. 1993. [13] R. W. Keith, J. L. Stambough, and S. H. Awender, “Somatosensory cortical evoked potentials: A review of 100 cases of intraoperative spinal surgery monitoring,” J. Spinal Disorders Techniques, vol. 3, no. 3, pp. 220–226, Sep. 1990. [14] A. R. Møller, “Monitoring Somatosensory Evoked Potentials,” in Intraoperative Neurophysiological Monitoring. New York, NY, USA: Springer, 2011, pp. 93–122. [15] M. Banoub, J. E. Tetzlaff, and A. Schubert, “Pharmacologic and physiologic influences affecting sensory evoked potentials: Implications for perioperative monitoring,” Anesthesiology, vol. 99, no. 3, pp. 716–737, Sep. 2003.
FAN et al.: INTELLIGENT DECISION SYSTEM FOR INTRAOPERATIVE SEP MONITORING
[16] M. Seyal and B. Mull, “Mechanisms of signal change during intraoperative somatosensory evoked potential monitoring of the spinal cord,” J. Clin. Neurophysiol., vol. 19, no. 5, pp. 409–415, Oct. 2002. [17] H. Cui, K. D. K. Luk, and Y. Hu, “Effects of physiological parameters on intraoperative somatosensory evoked potential monitoring: Results from a multifactor analysis,” Med. Sci. Monitor, vol. 15, no. 5, pp. 226–230, Apr. 2009. [18] J. J. Jana, N. Vaid, and J. Shanbhag, “Effect of total intravenous anaesthesia on intraoperative monitoring of Cochlear implant function in paediatric patients,” Cochlear Implants Int., vol. 14, no. 3, pp. 169–173, Jun. 2013. [19] K. J. Noonan, T. Walker, J. R. Feinberg, M. Nagel, W. Didelot, and R. Lindseth, “Factors related to false-versus true-positive neuromonitoring changes in adolescent idiopathic scoliosis surgery,” Spine, vol. 27, no. 8, pp. 825–830, Apr. 2002. [20] H. Wiedemayer, I. E. Sandalcioglu, W. Armbruster, J. Regel, H. Schaefer, and D. Stolke, “False negative findings in intraoperative SEP monitoring: Analysis of 658 consecutive neurosurgical cases and review of published reports,” J. Neurol., Neurosurgery Psychiatry, vol. 75, no. 2, pp. 280–286, Dec. 2004. [21] S. L. Papastefanou, L. M. Henderson, N. J. Smith, A. Hamilton, and J. K. Webb, “Surface electrode somatosensory-evoked potentials in spinal surgery: Implications for indications and practice,” Spine, vol. 25, no. 19, pp. 2467–2472, Oct. 2000. [22] J. A. K. Suykens, T. Van Gestel, J. De Brahanter, B. De Moor, and J. Vandewalle, Least Squares Support Vector Machines. Singapore: World Scientific, 2002. [23] C. K. Qi, H. X. Li, X. X. Zhang, X. C. Zhao, S. Y. Li, and F. Gao, “Time/space-separation-based SVM modeling for nonlinear distributed parameter processes,” Industr. Eng. Chemistry Res., vol. 50, no. 1, pp. 332–341, Nov. 2010. [24] H. X. Li, J. L. Yang, G. Zhang, and B. Fan, “Probabilistic support vector machines for classification of noise affected data,” Inform. Sci., vol. 221, pp. 60–71, Feb. 2013. [25] D. Montgomery, Introduction to Statistical Quality Control. Hoboken, NJ, USA: Wiley, 2009. [26] K. De Brabanter K, J. De Brabanter, J. A. K. Suykens, and B. De Moor, “Approximate confidence and prediction intervals for least squares support vector regression,” IEEE Trans. Neural Networks, vol. 22, no. 1, pp. 110–120, Jan. 2011. [27] I. Goethals, K. Pelckmans, J. A. Suykens, and B. De Moor, “Subspace identification of Hammerstein systems using least squares support vector machines,” IEEE Trans. Automat. Contr., vol. 50, no. 10, pp. 1509–1519, Oct. 2005. [28] T. Hastie, R. Tibshirani, J. Friedman, and J. Franklin, “The elements of statistical learning: Data mining, inference and prediction,” Math. Intelligencer, vol. 27, no. 2, pp. 83–85, Nov. 2005. [29] V. N. Vapnik, “An overview of statistical learning theory,” IEEE Trans. Neural Networks, vol. 10, no. 5, pp. 988–999, Sep. 1999. [30] S. Ben-David, “A framework for statistical clustering with constant time approximation algorithms for k-median and k-means clustering,” Machine Learning, vol. 66, no. 2, pp. 243–257, Mar. 2007. [31] J. A. Hartigan and M. A. Wong, “Algorithm AS 136: A k-means clustering algorithm,” J. Roy. Statistical Soc., Series C (Applied Statistics), vol. 28, no. 1, pp. 100–108, Aug. 1979. [32] K. P. Murphy, Machine Learning: A Probabilistic Perspective. Cambridge, MA, USA: The MIT Press, 2012. [33] G. B. Huang, H. Zhou, X. Ding, and R. Zhang, “Extreme learning machine for regression and multiclass classification,” IEEE Trans. Systems, Man, Cybernetics-Part B: Cybernetics, vol. 42, no. 2, pp. 513–529, Mar. 2012. [34] C. R. Rao and S. K. Mitra, Generalized Inverse of Matrices and Its Applications. New York, NY, USA: Wiley, 1971. [35] T. P. Ryan, Statistical Methods for Quality Improvement. Hoboken, NJ, USA: Wiley, 2011.
307
[36] S. L. Hwang, J. T. Lin, G. F. Liang, Y. J. Yau, T. C. Yenn, and C. C. Hsu, “Application control chart concepts of designing a pre-alarm system in the nuclear power plant control room,” Nuclear Eng. Design, vol. 238, no. 12, pp. 3522–3527, 2008. [37] W. Hachicha and A. Ghorbel, “A survey of control-chart pattern-recognition literature (1991–2010) based on a new conceptual classification scheme,” Computers Industr. Eng., vol. 63, no. 1, pp. 204–222, 2012. [38] Z. J. Zhang and A. Kusiak, “Monitoring wind turbine vibration based on SCADA data,” Trans. ASME-N-J. Solar Energy Eng., vol. 134, no. 2, pp. 1–12, Feb. 2012, 021004. [39] R. Caruana and A. Niculescu-Mizil, “An empirical comparison of supervised learning algorithms,” in Proc. 23rd Int. Conf. Machine Learning ACM, Jun. 2006, pp. 161–168. [40] K. D. K. Luk, Y. Hu, Y. W. Wong, and J. C. Y. Leong, “Variability of somatosensory-evoked potentials in different stages of scoliosis surgery,” Spine, vol. 24, no. 17, pp. 1799–1804, Sep. 1999.
Bi Fan received the B.E. degree in automation from Ocean University of China, in 2002, the M.E. degree in mechatronics from University of Shanghai for Science and Technology, in 2010, and the Ph.D. degree in system engineering and engineering management from City University of Hong Kong, in 2014. His research interests include neural engineering, machine learning and nonlinear system.
Han-Xiong Li (F’10) received the B.E. degree in aerospace engineering from the National University of Defense Technology, China, in 1982, the M.E. degree in electrical engineering from Delft University of Technology, The Netherlands, in 1991, and the Ph.D. degree in electrical engineering from the University of Auckland, New Zealand, in 1997. He is a Professor in the Department of Systems Engineering and Engineering Management, the City University of Hong Kong. His current research interests are in system intelligence and control, process design and control integration, distributed parameter systems. Prof. Li serves as Associate Editor of IEEE TRANSACTIONS ON CYBERNETICS, and IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS (since 2009). He was awarded the Distinguished Young Scholar award (overseas) by the China National Science Foundation in 2004, a Chang Jiang Professorship by the Ministry of Education, China, in 2006, and a National Professorship in China Thousand Talents Program in 2010.
Yong Hu (SM’12) received the B.Sc. and M.Sc degrees in the biomedical engineering from Tianjin University, Tianjin, China, in 1985 and 1988, respectively, and the Ph.D. degree from The University of Hong Kong in 1999. He is currently an Associate Professor and Director of Neural Engineering and Clinical Electrophysiology Laboratory, the Department of Orthopaedics and Traumatology, The University of Hong Kong. His research interests include neural engineering, clinical electrophysiology, biomedical signal measurement and processing.
