ORIGINAL ARTICLE
Machine learning on encephalographic activity may predict opioid analgesia M. Gram1, C. Graversen1, A.E. Olesen1,2, A.M. Drewes1,3 1 Mech-Sense, Department of Gastroenterology and Hepatology, Aalborg University Hospital, Denmark 2 Department of Drug Design and Pharmacology, Faculty of Health and Medical Sciences, University of Copenhagen, Denmark 3 Clinical Institute, Aalborg University Hospital, Denmark
Correspondence A.M. Drewes E-mail: [email protected] Funding sources This work was supported by funding from the Danish Council for Strategic Research (grant number 10-092786). Conflicts of interest None declared.
Accepted for publication 23 April 2015 doi:10.1002/ejp.734
Abstract Background: Opioids are used for the treatment of pain. However, 30–50% of patients have insufficient effect to the opioid initially selected by the physician, and there is an urgent need for biomarkers to select responders to the most appropriate drug. Since opioids mediate their effect in the central nervous system, this study aimed to investigate if electroencephalography (EEG) during rest or pain before treatment could predict the analgesic response. Methods: EEG from 62 channels was recorded in volunteers during rest and tonic pain (cold pressor test). Morphine (30 mg) or placebo was then administered, and the pain test repeated 60 min after. Washout period between drugs was 7 days. Based on pain ratings, subjects were stratified into responders and non-responders. Spectral analysis was performed on the EEG. Conventional statistics on group basis were used and, furthermore, the most discriminative EEG features were subjected to support vector machine classification to predict the response for the individual subjects. Results: Conventional statistics on the frequency bands revealed no differences between responders and non-responders. On the individual basis, no differences between groups were found using resting EEG. However, EEG during cold pain was able to classify responders with an accuracy of 72% (p = 0.01) and the result was reproducible using baseline data from both study days. Conclusions: Machine learning based on EEG before treatment enabled separation between responders and non-responders. This study represents the first step towards the prediction of opioid analgesia based on EEG features prior to drug administration, and advocates for the use of machine learning in future studies. For this article, a commentary is available at the Wiley Online Library.
1. Introduction Opioids are the drug of choice for moderate-tosevere pain (Liu and Wu, 2007). Despite major advances in the understanding of pain, treatment of acute pain such as in the postoperative conditions still remains unsatisfactory (Dolin et al., 2002; Sommer et al., 2008). The inadequate treatment has a 1552 Eur J Pain 19 (2015) 1552--1561
negative impact on patients both immediately following surgery, but also in the development of persistent long-term pain (VanDenKerkhof et al., 2012). On an individual level, there is a difference in the analgesic response to a given opioid. Various factors such as gender, age, and genetic variation can affect the analgesic response. The genetic variation can influence pharmacokinetics (e.g. drug transporters, drug-metabolizing enzymes and tolerance) and/or © 2015 European Pain Federation - EFICâ
M. Gram et al.
What’s already known about this topic • Pain patients often experience insufficient pain relief from opioids. • Response to opioids varies highly between individuals. • Accurate prediction of analgesic effect from opioids would enable more efficient pain therapy. What does this study add
• Electroencephalography shows promise for separating morphine responders from nonresponders. • Machine learning methods are required to accurately assess the complex relationships in the EEG.
pharmacodynamics (e.g. opioid receptor and catechol-O-methyltransferase enzymes) (Nielsen et al., 2014). Hence, in a study treating cancer pain with morphine, 74% of patients were well-treated initially with morphine (Riley et al., 2006). However, when switching the nonresponsive patients to other opioids, 96% experienced successful pain control (Riley et al., 2006). Therefore, personalizing pain treatment and thus initially enabling choice of the correct opioid shows great promise in optimizing pain treatment. Several attempts to predict analgesic efficacy has already been made using quantitative sensory testing (QST) (Grosen et al., 2013; Olesen et al., 2013). However, results have been contradictory between different QST modalities and study groups, indicating that while QST measures may be valuable in personalized medicine, there is still need for improvement and/or other methods to predict the effect of treatment (Grosen et al., 2013). Pain is a highly complex and individual experience, influenced by many factors which together has been shown to form a cerebral “signature”, responsible for producing the unique experience of pain (Tracey, 2008). This cerebral signature can be assessed using positron emission tomography and magnetic resonance imaging (Ahmedzai, 2013), and a better understanding of the inter-individual differences in brain function is thought to pave the way for personalized treatment of pain (Bruehl et al., 2013). A more clinically feasible method for assessing cerebral biomarkers could be electroencephalography (EEG) which also has considerable lower costs compared to imaging
© 2015 European Pain Federation - EFICâ
Machine learning on encephalographic activity may predict analgesia
solutions. EEG has been shown to predict the experience of pain within single subjects, underlining its ability to assess individual differences in cerebral function (Schulz et al., 2011). Therefore, the EEG might be used to differentiate between potential responders and non-responders to opioid treatment. The EEG is a composite signal reflecting the sum of a variety of neuronal activities. Therefore, the response to pain has traditionally been directed in a more specific way using painful stimulations followed by recordings of evoked brain potentials (Frøkjær et al., 2011). Although the “signal to noise ratio” of phasic stimulations are likely better than the EEG obtained during tonic pain, such brief stimuli may be less unsuited to simulate natural pain conditions dominated by hyperalgesia and affective feelings. This has moved quantitative sensory testing towards more tonic pain models (Nir et al., 2012). Especially, the cold pressor test has been widely used since it has been shown to be highly reliable and to mimic clinical pain (Rainville et al., 1992; Lewis et al., 2012). Furthermore, endogenous opioids play a major role in activation of the descending noxious inhibitory control system. Hence, the cold pressor test which activates this system was used to assess the effect of morphine. Due to the complex nature of pain perception and analgesics, it is likely that true prediction can only be achieved using a combination of several features, which requires advanced methods of analysis such as machine learning techniques (Khodayari-Rostamabad et al., 2010). Such methods have previously been used to detect individual differences in pain experiences (Brown et al., 2011) and in contrast to conventional statistics on group basis outcome of the individual subject is predicted by the model. We hypothesized that classification of pre-treatment EEG during tonic pain with a support vector machine (SVM) can be used to predict the response to morphine treatment. Thus, this study aimed to investigate if (1) SVM classification based on EEG recorded prior to morphine administration would be superior to conventional group-based statistics to differentiate between morphine responders and non-responders and (2) to investigate reproducibility and robustness of the model using data in the same subjects from another recording day.
2. Methodology The study was conducted between November 2010 and April 2012 in the research laboratory at
Eur J Pain 19 (2015) 1552--1561
1553
M. Gram et al.
2.5 Cold pressor test The cold pressor test was performed using circulated water bath (Grant, Fischer Scientific, Slangerup, Denmark). The water was cooled to 2 °C and the subjects immersed their left hand up to the wrist for 2 min while water was circulated. Subjects rated their pain continuously on a handheld electronic device with a visual analogue scale (VAS) going from no feeling of pain (0) to the worst imaginable pain (10). The pain ratings were electronically sampled continuously with a frequency of 10 Hz.
2.6 Stratification Subjects were stratified into responders and nonresponders based on their relative change in cold pain perception from before to after morphine administration. The area under the VAS curve (AUC) was calculated before and after drug administration. Afterwards, the relative change from baseline (in percentage) to post-drug administration was calculated. Subjects were labelled as responders if their AUC decreased by 5% or more after administration of morphine.
2.7 EEG recordings before treatment On both days, EEG was recorded in a dimly lit room, first during a resting period and then during the cold pressor test. EEG was recorded from a standard 62-channel cap (Quick-Cap International, Neuroscan, El Paso, TX, USA), amplified digitally on a Synamps 2 system (Neuroscan Compumedics, El Paso, TX, USA) and recorded for later analysis (Neuroscan 4.3.1, Neuroscan, El Paso, TX, USA). During the resting EEG recordings, subjects were instructed to keep their eyes open while minimizing eye blinking during the 2.5 min period. EEG during cold pain was recorded by starting the recording simultaneously as the subject submerged their hand in cold water.
2.8 Pre-processing The data were first pre-processed in the Neuroscan EEG software in the following steps: (1) zero-phase shift notch filtering (49–51 Hz) using a finite impulse-response filter with a slope of 24 dB/octave; (2) zero-phase shift band-pass filtering (1–70 Hz) using a finite impulse-response filter with a slope of 12 dB/octave; (3) blinded visual inspection of data quality for all channels using linked-ear reference. Channels with abnormal signals were discarded and © 2015 European Pain Federation - EFICâ
Machine learning on encephalographic activity may predict analgesia
replaced by signals interpolated from neighbouring electrodes; (4) Re-referencing to the average electrode; (5) Finally, resting EEG was cleaned by visually selecting 2 min of artefact-free EEG from the 2.5 min recording where the investigator was blinded to the origin of the recording.
2.9 Spectral analysis of EEG dynamics Spectral analysis of EEG amplitudes was carried out using Matlab 2012a (The Matworks, Inc., Natick, MA, USA) to calculate the relative EEG amplitude. The continuous wavelet transform was applied to the EEG signals from all channels. The complex Morlet wavelet was chosen for analysis with a bandwidth of 10 Hz and a centre frequency of 1 Hz. Scales for the transform were chosen to match frequencies ranging from 1 to 70 Hz, with a 0.5 Hz between-scale frequency interval. The absolute value of the obtained wavelet coefficients were used in the following analysis. To assess the static EEG spectral indices, the wavelet coefficients were divided into the following five standardized bands: delta (1–4 Hz), theta (4–8 Hz), alpha (8–12 Hz), beta (12–32 Hz) and gamma (32–70 Hz). The wavelet coefficients were averaged over time and scales contained within each frequency band were summed together to yield the absolute activity within each frequency band. The relative activity was calculated separately for each channel by dividing each frequency band with the total energy of all bands and multiplying by 100. The values then represent the percentage of total amplitude contained in each frequency band.
2.10 Conventional group-based analysis Data are presented as average SD. The EEG data from the morphine day prior to drug administration were first analysed using two-way repeated measures analysis of variance (ANOVA) using the group (responder or non-responder) as between-subject factor and the channel as within-subject factor. A p-value below 0.05 for the group factor was considered statistically significant. If statistical differences were found, the analysis was repeated using data from the placebo day.
2.11 Machine learning analysis To avoid over-fitting, the most discriminative features were selected using the criteria for joint mutual information, as this criterion has been found to provide the best selection for data sets with a limited Eur J Pain 19 (2015) 1552--1561
1555
Machine learning on encephalographic activity may predict analgesia
M. Gram et al.
Table 1 Results from the spectral analysis from each study day. Results are shown for responders and non-responders to morphine for each channel selected by “joint mutual information”. All values are shown as the percentage of total amplitude for each electrode. The electrodes are ranked in order of importance for the classification. Morphine day Selected channel
Responders
1. 2. 3. 4. 5. 6. 7.
31.2 44.6 38.4 32.9 31.9 31.4 16.1
Delta FT8 Delta FPZ Delta AF4 Delta FT7 Delta FC5 Delta FC3 Alpha C2
7.1 6.3 6.3 8.0 7.6 7.5 4.0
Placebo day Non-responders 34.4 46.4 41.3 32.3 29.7 29.3 16.3
7.0 6.2 6.8 7.6 6.5 6.3 3.8
Table 2 Confusion matrix from the classification using data prior to the administration of morphine and placebo.
Morphine day Classified responder Classified non-responder Placebo day Classified responder Classified non-responder
Responders
Non-responders
13 4
5 10
12 5
4 11
SVM classification using pre-treatment EEG during cold pain for the morphine day data was significant (p = 0.01) with 71.9% accuracy (positive predictive value = 70.0% and negative predictive value = 75.0%). Table 2 shows the 2 9 2 confusion matrix for the classification. Likewise, classification obtained with the data recorded on the other day prior to placebo drug administration provided a similar significant separation of the groups (p = 0.01) with 71.9% accuracy (positive predictive value = 75% and negative predictive value = 68.8%). The 2 9 2 confusion matrix is seen in Table 2.
4. Discussion To our knowledge, this is the first attempt to utilize EEG to identify specific brain patterns with advanced machine learning techniques in order to predict the individual response to opioid analgesia. The study showed that the support vector machine approach indeed was able to predict a positive response with 72% accuracy using EEG before treatment during tonic cold pain while conventional group-based analysis was insufficient. Furthermore, this result was reliable using data recorded on both study days. The primary features used for separation was the delta band from frontally and centrally placed electrodes. 1558 Eur J Pain 19 (2015) 1552--1561
Selected channel
Responders
1. 2. 3. 4. 5. 6. 7.
21.7 28.3 13.2 42.0 24.5 15.1 35.2
Delta T8 Delta FC1 Alpha T7 Delta FP2 Delta C6 Alpha TP7 Beta FT7
7.2 5.4 3.1 7.1 5.9 4.0 6.6
Non-responders 27.2 29.8 12.7 45.5 26.7 14.5 30.3
7.5 6.0 3.8 7.7 6.2 4.0 6.0
4.1 Prediction of morphine analgesia Resting EEG did not discriminate between responders and non-responders in this study, while this was possible using EEG during cold pain. The reason for this is unknown; however, reliability for resting EEG has been shown to be less than for EEG during cold pain (Gram et al., 2014). It could be speculated that the intense, painful stimuli increased attention levels to pain while simultaneously suppressing some conscious and non-conscious processes, which could confound the analysis (Gram et al., 2014). EEG during cold pain provided a reliable result, both in terms of accuracy and in the selected channels using data from both study days. For a prediction method to be clinically relevant, reliability is highly desirable, making EEG during cold pain better suited for prediction of analgesic treatment outcome. The spectral EEG indices recorded during rest and cold pressor test has previously been shown to be reliable between days (Gram et al., 2014). Furthermore, the current analysis yielded very similar results using data from both study days, both in terms of classification performance and in terms of the channels selected, indicating that the method is reliable. There are minor differences in the channels selected between the 2 days, mainly in the alpha and beta bands. However, some variation is to be expected in the feature selection procedure, since features are selected based on relevance to the classifier. However, features sharing too much information are excluded to prevent redundancy. Furthermore, the alpha and beta band features were generally selected last, meaning that they are of lesser relevance to the classification. As such, the variation in selected channels seems well within the expected variation for the method. The accuracy of the method in its current form, while statistically significant, is likely still too low to be considered usable in clinical practice. Further work should focus on increasing the accuracy of the © 2015 European Pain Federation - EFICâ
Machine learning on encephalographic activity may predict analgesia
number of samples (Brown et al., 2012). For both resting EEG and EEG during cold pain recorded immediately before administration of morphine, features were selected using the joint mutual information criterion and subsequently used for SVM classification. The number of features to be used was determined by investigating the accuracy of the classifier by gradually increasing the number of features up to 15. Accuracy was defined as the ratio between correctly classified subjects and total number of subjects in percentage. The number of features that yielded the highest accuracy on both study days was chosen for the final analysis due to the importance of reliability in the classifier. Classification was performed using the libSVM toolbox (version 3.20) for Matlab (Chang and Lin, 2011). A linear kernel function was used to avoid overfitting of the data (Gong et al., 2011). The cost parameter C of the SVM was set to 1 which is in the right order of magnitude (Auffarth et al., 2010). Classification accuracy was determined using leaveone-out cross-validation which consisted of removing one subject before training the SVM classifier using all remaining subjects. The last subject was then used to test the predictive capability of the classifier. This was repeated until all subjects had been left out, to assess the overall accuracy of the classifier (Gong et al., 2011). Accuracy, positive predictive value and negative predictive value of the classification were calculated as well as the statistical significance of the classification using chi-square where a p-value below 0.05 was used to indicate statistical significance.
M. Gram et al.
3. Results Thirty-nine subjects were enrolled in the study. Five subjects were excluded as they could not tolerate the cold pressor test for the full 2 min and thus had incomplete EEG recordings and two subjects were excluded due to poor EEG data quality. The subjects’ course through the study is summarized in Fig. 2 as a CONSORT flowchart. This resulted in 32 subjects in the final analysis (15 females and 17 males, aged 27.2 7.1 years).
3.1 Stratification Seventeen subjects decreased their AUC during the cold pressor test more than 5% during morphine treatment (AUC: 14.7% 6.9), and as such were labelled as responders. The remaining 15 subjects were labelled as non-responders (AUC: 1.1% 2.3). Response to placebo medication was not different between the responder and the non-responder group (2.2% 6.7 vs. 0.4% 3.5; p = 0.4).
3.2 Conventional group-based analysis The topographical representation of the spectral indices during resting EEG is shown on Fig. 3, while the spectral indices for EEG during cold pain is shown on Fig. 4. There were no differences between responders and non-responders for resting EEG spectral indices (all frequency bands F < 1.9; all p > 0.15). The same was true for EEG during cold pain (all F < 0.3; all p > 0.4).
Screened N = 93
Excluded N = 44
Included N = 49
Completed N = 39
Analyzed N = 32
Not eligible (N = 10) Not willing to participate (N = 34)
Drop outs N = 10
Decided to leave the study (N = 3) Did not tolerate experiment (N = 3) Due to side effects (N = 3) Due to accidental knee injury (N = 1)
Removed N = 7
Due to data quality (N = 2) Did not complete cold pressor test (N = 5)
1556 Eur J Pain 19 (2015) 1552--1561
Figure 2 CONSORT flow chart showing the subjects’ course during the study.
© 2015 European Pain Federation - EFICâ
M. Gram et al.
Machine learning on encephalographic activity may predict analgesia
Delta
Theta
Alpha
Beta
Gamma
Responders
Non-responders
Figure 3 Topographical plot of the spectral EEG indices from resting EEG for responders and non-responders to morphine analgesia. All values are the relative amplitude within the frequency band in percentage of total amplitude.
Delta
Theta
Alpha
Beta
Gamma
Responders
Non-responders
Figure 4 Topographical plot of the spectral EEG indices from cold pressor EEG for responders and non-responders to morphine analgesia. All values are the relative amplitude within the frequency band in percentage of total amplitude.
3.3 Machine learning analysis Resting EEG did not provide significant classification using any number of features (maximum accuracy: 12 features; 66% accuracy, p = 0.08); as such no further analysis was done on resting EEG. For EEG during cold pain the optimal number of features was found to be 7 since the accuracy was high using data from both days.
The seven features used in the classification were found in the delta band (FPZ, AF4, FC3, FC5, FT7 and FT8) and alpha band (C2) in the morphine arm (Fig. 5 shows the topographical distribution of selected channels). In the placebo arm, the features were found in the delta band (FP2, FC1, C6 and T8), the alpha band (TP7 and T7) and the beta band (FT7). Table 1 shows the values for all selected features, for each group.
Delta
Alpha
FPZ AF4
FT8
Day 1
FT7 FC5 FC3
C2
Delta
Alpha
Beta
Figure 5 Topographical representation of the selected channels in each electroencephalographic frequency band during cold pain using data from day 1 (prior to morphine administration) and day 2 (prior to placebo administration).
© 2015 European Pain Federation - EFICâ
Day 2
AF4
FT7
FC1 C6
T8
T7 TP7
Eur J Pain 19 (2015) 1552--1561
1557
Machine learning on encephalographic activity may predict analgesia
M. Gram et al.
Table 1 Results from the spectral analysis from each study day. Results are shown for responders and non-responders to morphine for each channel selected by “joint mutual information”. All values are shown as the percentage of total amplitude for each electrode. The electrodes are ranked in order of importance for the classification. Morphine day Selected channel
Responders
1. 2. 3. 4. 5. 6. 7.
31.2 44.6 38.4 32.9 31.9 31.4 16.1
Delta FT8 Delta FPZ Delta AF4 Delta FT7 Delta FC5 Delta FC3 Alpha C2
7.1 6.3 6.3 8.0 7.6 7.5 4.0
Placebo day Non-responders 34.4 46.4 41.3 32.3 29.7 29.3 16.3
7.0 6.2 6.8 7.6 6.5 6.3 3.8
Table 2 Confusion matrix from the classification using data prior to the administration of morphine and placebo.
Morphine day Classified responder Classified non-responder Placebo day Classified responder Classified non-responder
Responders
Non-responders
13 4
5 10
12 5
4 11
SVM classification using pre-treatment EEG during cold pain for the morphine day data was significant (p = 0.01) with 71.9% accuracy (positive predictive value = 70.0% and negative predictive value = 75.0%). Table 2 shows the 2 9 2 confusion matrix for the classification. Likewise, classification obtained with the data recorded on the other day prior to placebo drug administration provided a similar significant separation of the groups (p = 0.01) with 71.9% accuracy (positive predictive value = 75% and negative predictive value = 68.8%). The 2 9 2 confusion matrix is seen in Table 2.
4. Discussion To our knowledge, this is the first attempt to utilize EEG to identify specific brain patterns with advanced machine learning techniques in order to predict the individual response to opioid analgesia. The study showed that the support vector machine approach indeed was able to predict a positive response with 72% accuracy using EEG before treatment during tonic cold pain while conventional group-based analysis was insufficient. Furthermore, this result was reliable using data recorded on both study days. The primary features used for separation was the delta band from frontally and centrally placed electrodes. 1558 Eur J Pain 19 (2015) 1552--1561
Selected channel
Responders
1. 2. 3. 4. 5. 6. 7.
21.7 28.3 13.2 42.0 24.5 15.1 35.2
Delta T8 Delta FC1 Alpha T7 Delta FP2 Delta C6 Alpha TP7 Beta FT7
7.2 5.4 3.1 7.1 5.9 4.0 6.6
Non-responders 27.2 29.8 12.7 45.5 26.7 14.5 30.3
7.5 6.0 3.8 7.7 6.2 4.0 6.0
4.1 Prediction of morphine analgesia Resting EEG did not discriminate between responders and non-responders in this study, while this was possible using EEG during cold pain. The reason for this is unknown; however, reliability for resting EEG has been shown to be less than for EEG during cold pain (Gram et al., 2014). It could be speculated that the intense, painful stimuli increased attention levels to pain while simultaneously suppressing some conscious and non-conscious processes, which could confound the analysis (Gram et al., 2014). EEG during cold pain provided a reliable result, both in terms of accuracy and in the selected channels using data from both study days. For a prediction method to be clinically relevant, reliability is highly desirable, making EEG during cold pain better suited for prediction of analgesic treatment outcome. The spectral EEG indices recorded during rest and cold pressor test has previously been shown to be reliable between days (Gram et al., 2014). Furthermore, the current analysis yielded very similar results using data from both study days, both in terms of classification performance and in terms of the channels selected, indicating that the method is reliable. There are minor differences in the channels selected between the 2 days, mainly in the alpha and beta bands. However, some variation is to be expected in the feature selection procedure, since features are selected based on relevance to the classifier. However, features sharing too much information are excluded to prevent redundancy. Furthermore, the alpha and beta band features were generally selected last, meaning that they are of lesser relevance to the classification. As such, the variation in selected channels seems well within the expected variation for the method. The accuracy of the method in its current form, while statistically significant, is likely still too low to be considered usable in clinical practice. Further work should focus on increasing the accuracy of the © 2015 European Pain Federation - EFICâ
M. Gram et al.
classification. It is possible that the discussed limitations of the current study are also limiting the accuracy of the prediction and simply addressing these limitations in future investigations alone will improve the rate of correct classifications. Furthermore, the EEG features identified in this study can be combined in machine learning with other features such as pain thresholds, gender and age to further increase the accuracy of the algorithm. Several studies have already investigated prediction of opioid analgesia using QST assessments, although with varying results, for review see (Grosen et al., 2013). Genetic factors also affect the efficacy of opioids (Maier et al., 2002; Kalso et al., 2004; Tremblay and Hamet, 2010). Features could be extracted from the genetic profile of subjects such as the presence of specific mutations of genes coding for opioid receptors. This could have implications for the acute pain treatment such as in postoperative pain, where many patients experience inadequate pain relief (Dolin et al., 2002; Sommer et al., 2008). If investigated further, the method could potentially offer a low-cost and simple decision-aid to determine if a patient should go ahead with conventional morphine treatment or be switched to another opioid.
4.2 Methodological considerations Morphine was administered as a single dose of 30 mg as it is a clinically relevant dose. Subsequently subjects were tested after 60 min to determine effect. Classification accuracy could possibly improve if subjects were treated over a longer treatment period to achieve steady-states of drug concentrations where testing could be performed. This would possibly account for individual differences in morphine absorption and ensure high levels of drug concentration in the brain. However, in volunteers higher doses can hardly be used due to side effects (Olesen et al., 2014). We stratified subjects into two groups based on their response to morphine in the cold pressor test. Since the placebo response in the cold pressor test was low in general in healthy volunteers and there was no difference in the placebo response between responders and non-responders, the findings likely represent differences in morphine analgesia. Thus, as such, in the stratification process the individual’s response to placebo medication was not considered. Response level is difficult to determine for experimental pain models such as the cold pressor test. Clinically, a 30% decrease in clinical pain scores has traditionally been taken as an indicator for response (Olesen et al., 2013). Had this threshold been © 2015 European Pain Federation - EFICâ
Machine learning on encephalographic activity may predict analgesia
applied in this study, only a single subject would have been considered a responder to morphine, which is not likely with a clinically relevant dose of 30 mg. Lower reductions in pain thresholds are typically seen in quantitative sensory testing models (Olesen et al., 2014). Furthermore, the effect of morphine in volunteers typically differ from that in patients, where the pain is more complex and involve a variety of affective and cognitive components different from pain intensity in the experimental settings (Olesen et al., 2012). Therefore, a value of 5% reduction in AUC following morphine administration was used. This resulted in groups of similar size, which is preferable when using machine learning on a small data samples. EEG has previously been shown to predict the later experience of pain both in laser evoked brain potentials (Babiloni et al., 2008) and in tonic heat pain (Nir et al., 2012). As such it is clear that the immediate state of the brain prior to a painful stimulus affects the experience. Hence, it does not only vary between individuals, but subsequent painful experiences can vary for the individual as well. In the current study results were repeatable with more than a week between the recordings. As such the analgesic effect speaks to a more general trait for the individual (or a long-term state), which is of interest for predicting the analgesic effect in a reliable way. EEG has previously been shown to be sensitive to morphine analgesia (Olesen et al., 2012). On a physiological level it has been suggested that morphine has been shown to attenuate pain by interaction with the medial pain system, especially the anterior cingulate cortex and insula (Lelic et al., 2014). Studies show that morphine affects the brain by reorganizing the pain networks (Lelic et al., 2014). Individual differences in the brain could affect this interaction and increase or decrease the response to morphine. EEG measures the dynamic activity of the brain on a millisecond scale and therefore seems like an appropriate choice for investigating these individual differences. Other methods such as imaging techniques could also be used for investigation, but will incur higher costs for investigation. However, the machine learning approach allows features from other methods to be combined with features from the EEG for further increases in classification accuracy. The study was carried out using experimental pain models to assess the level of analgesia in order to minimize the influence of confounders by allowing accurate control of the painful stimulus (Staahl and Drewes, 2004; Olesen et al., 2012). Previously cold pressor tolerance has been used to predict effect of Eur J Pain 19 (2015) 1552--1561
1559
Machine learning on encephalographic activity may predict analgesia
oxycodone (Eisenberg et al., 2010), and it has been shown to be very sensitive to opioid analgesia (Jones et al., 1988; Koltzenburg et al., 2006; Olesen et al., 2012). Looking solely at EEG spectral indices it was not possible to discriminate on group level (e.g. using conventional group-based analysis). However, the current study demonstrated the possibility of selecting features from the EEG recorded during the cold pressor test using joint mutual information and machine learning to predict response or non-response to morphine for single individuals. The machine learning technique is a new approach compared to previous studies investigating the prediction of analgesic effect, where linear regression statistics are usually employed (Bares et al., 2008; Eisenberg et al., 2010). Using conventional group-based analysis for prediction of treatment outcome imposes several limitations; firstly the analysis does not directly assess the individual subject and assigns them a label such as machine learning methods. These studies utilize methods suitable only for detecting differences among groups, which is of little interest from a clinical perspective. The individual response to treatment is of more interest, and for this purpose machine learning methods are more suitable (Khodayari-Rostamabad et al., 2010). Secondly and most importantly, machine learning algorithms are able to assess multiple features simultaneously, which is beneficial for complex signals such as EEG, where the differences may only be explained by the simultaneous activity in several frequency bands and channels. A previous study was able to assess the analgesic effect of pregabalin by utilizing a SVM classifier and assessing all frequency bands simultaneously (Graversen et al., 2012). This could explain why in this study the group-based statistics failed to discover differences in all frequency bands, while feature extraction using joint mutual information and subsequent SVM classification provided significant classifications using data from both study days. With this in mind, machine learning algorithms seem preferential for future studies to predict treatment response of drugs due to the ability to assess complex interactions between features as well as enabling assessment on the individual level, such as has been shown within personalized medicine in psychiatric disorders (Khodayari-Rostamabad et al., 2010).
5. Conclusions Machine learning based on encephalographic activity before treatment enabled separation between responders and non-responders to morphine analge1560 Eur J Pain 19 (2015) 1552--1561
M. Gram et al.
sia on the individual basis, when conventional group-wise statistical methods failed. The study represents the first step towards prediction of opioid analgesia based on the recorded electrical brain activity prior to drug administration, and advocate for the use of machine learning in future studies. However, more studies with less and more focused electrodes and with larger populations are needed to prove the value of the method in a clinical setting. Author contributions A.M.D., A.E.O. and C.G. initiated the studies, carried out the pain testings and EEG recordings. M.G and C.G. designed and carried out the EEG spectral analysis. M.G. carried out the machine learning analysis. M.G., C.G., A.E.O. and A.M.D. evaluated results and drafted the manuscript. A.M.D., A.E.O. and C.G. conceived the study and participated in its coordination. All authors were involved in reading and approving the final manuscript.
References Ahmedzai, S.H. (2013). Personalized medicine–one size fits one: Tailoring pain therapy to individuals’ needs. J Pain Palliat Care Pharmacother 27, 83–85. Auffarth, B., L opez, M., Cerquides, J. (2010). Comparison of redundancy and relevance measures for feature selection in tissue classification of CT images. Adv Data Mining Appl 248–262. Babiloni, C., Del Percio, C., Brancucci, A., Capotosto, P., Le Pera, D., Marzano, N., Valeriani, M., Romani, G.L., Arendt-Nielsen, L., Rossini, P.M. (2008). Pre-stimulus alpha power affects vertex N2-P2 potentials evoked by noxious stimuli. Brain Res Bull 75, 581–590. Bares, M., Brunovsky, M., Kopecek, M., Novak, T., Stopkova, P., Kozeny, J., Sos, P., Krajca, V., H€ oschl, C. (2008). Early reduction in prefrontal theta QEEG cordance value predicts response to venlafaxine treatment in patients with resistant depressive disorder. Eur Psychiatry 23, 350–355. Brown, G., Pocock, A., Zhao, M.-J., Lujan, M. (2012). Conditional likelihood maximisation: A unifying framework for information theoretic feature selection. J Mach Learn Res 13, 27–66. Brown, J.E., Chatterjee, N., Younger, J., Mackey, S. (2011). Towards a physiology-based measure of pain: Patterns of human brain activity distinguish painful from non-painful thermal stimulation. PLoS ONE 6, e24124. Bruehl, S., Apkarian, A.V., Ballantyne, J.C., Berger, A., Borsook, D., Chen, W.G., Farrar, J.T., Haythornthwaite, J.A., Horn, S.D., Iadarola, M.J., Inturrisi, C.E., Lao, L., Mackey, S., Mao, J., Sawczuk, A., Uhl, G.R., Witter, J., Woolf, C.J., Zubieta, J.-K., Lin, Y. (2013). Personalized medicine and opioid analgesic prescribing for chronic pain: Opportunities and challenges. J Pain 14, 103–113. Chang, C.-C., Lin, C.-J. (2011). LIBSVM. ACM Trans Intell Syst Technol 2, 1–27. Dolin, S.J., Cashman, J.N., Bland, J.M. (2002). Effectiveness of acute postoperative pain management: I. Evidence from published data. Br J Anaesth 89, 409–423. Eisenberg, E., Midbari, A., Haddad, M., Pud, D. (2010). Predicting the analgesic effect to oxycodone by “static” and “dynamic” quantitative sensory testing in healthy subjects. Pain 151, 104–109. Frøkjær, J.B., Egsgaard, L.L., Graversen, C., Søfteland, E., Dimcevski, G., Blauenfeldt, R.A., Drewes, A.M. (2011). Gastrointestinal symptoms in type-1 diabetes: Is it all about brain plasticity? Eur J Pain 15, 249–257.
© 2015 European Pain Federation - EFICâ
M. Gram et al.
Gong, Q., Wu, Q., Scarpazza, C., Lui, S., Jia, Z., Marquand, A., Huang, X., McGuire, P., Mechelli, A. (2011). Prognostic prediction of therapeutic response in depression using high-field MR imaging. NeuroImage 55, 1497–1503. Gram, M., Graversen, C., Olesen, S.S., Drewes, A.M. (2014). Dynamic spectral indices of the electroencephalogram provide new insights into tonic pain. Clin Neurophysiol 126, 763–771. Graversen, C., Olesen, S.S., Olesen, A.E., Steimle, K., Farina, D., Wilder-Smith, O.H.G., Bouwense, S.A.W., van Goor, H.Drewes, A.M. (2012). The analgesic effect of pregabalin in patients with chronic pain is reflected by changes in pharmaco-EEG spectral indices. Br J Clin Pharmacol 73, 363–372. Grosen, K., Fischer, I.W.D., Olesen, A.E., Drewes, A.M. (2013). Can quantitative sensory testing predict responses to analgesic treatment? Eur J Pain 17, 1267–1280. Jones, S.F., McQuay, H.J., Moore, R.A., Hand, C.W. (1988). Morphine and ibuprofen compared using the cold pressor test. Pain 34, 117–122. Kalso, E., Edwards, J.E., Moore, R.A., McQuay, H.J. (2004). Opioids in chronic non-cancer pain: Systematic review of efficacy and safety. Pain 112, 372–380. Khodayari-Rostamabad, A., Hasey, G.M., Maccrimmon, D.J., Reilly, J.P., de Bruin, H. (2010). A pilot study to determine whether machine learning methodologies using pre-treatment electroencephalography can predict the symptomatic response to clozapine therapy. Clin Neurophysiol 121, 1998–2006. Koltzenburg, M., Pokorny, R., Gasser, U.E., Richarz, U. (2006). Differential sensitivity of three experimental pain models in detecting the analgesic effects of transdermal fentanyl and buprenorphine. Pain 126, 165–174. Lelic, D., Olesen, A.E., Gregersen, H., Dahan, A., Kolesnikov, Y., Drewes, A.M. (2014). Morphine modifies the cingulate-operculum network underlying painful rectal evoked potentials. Neuropharmacology 77, 422–427. Lewis, G.N., Heales, L., Rice, D.A., Rome, K., McNair, P.J. (2012). Reliability of the conditioned pain modulation paradigm to assess endogenous inhibitory pain pathways. Pain Res Manag 17, 98–102. Liu, S.S., Wu, C.L. (2007). The effect of analgesic technique on postoperative patient-reported outcomes including analgesia: A systematic review. Anesth Analg 105, 789–808. Maier, C., Hildebrandt, J., Klinger, R., Henrich-Eberl, C., Lindena, G. (2002). Morphine responsiveness, efficacy and tolerability in patients with chronic non-tumor associated pain – Results of a double-blind placebo-controlled trial (MONTAS). Pain 97, 223–233. Nielsen, L.M., Olesen, A.E., Branford, R., Christrup, L.L., Sato, H., Drewes, A.M. (2014). Association between human pain-related genotypes and variability in opioid analgesia: An updated review. Pain Pract http://onlinelibrary.wiley.com/doi/10.1111/papr.12232/ references.
© 2015 European Pain Federation - EFICâ
Machine learning on encephalographic activity may predict analgesia
Nir, R.-R., Sinai, A., Moont, R., Harari, E., Yarnitsky, D. (2012). Tonic pain and continuous EEG: Prediction of subjective pain perception by alpha-1 power during stimulation and at rest. Clin Neurophysiol 123, 605–612. Olesen, A., Andresen, T., Staahl, C., Drewes, A. (2012). Human experimental pain models for assessing the therapeutic efficacy of analgesic drugs. Pharmacol Rev 64, 722–779. Olesen, A.E., Brock, C., Sverrisd ottir, E., Larsen, I.M., Drewes, A.M. (2014). Sensitivity of quantitative sensory models to morphine analgesia in humans. J Pain Res 7, 717–726. Olesen, A.E., Staahl, C., Arendt-Nielsen, L., Drewes, A.M. (2010). Different effects of morphine and oxycodone in experimentally evoked hyperalgesia: A human translational study. Br J Clin Pharmacol 70, 189–200. Olesen, S.S., Graversen, C., Bouwense, S.A.W., van Goor, H., WilderSmith, O.H.G., Drewes, A.M. (2013). Quantitative sensory testing predicts pregabalin efficacy in painful chronic pancreatitis. PLoS ONE 8, e57963. Rainville, P., Feine, J.S., Bushnell, M.C., Duncan, G.H. (1992). A psychophysical comparison of sensory and affective responses to four modalities of experimental pain. Somatosens Mot Res 9, 265–277. Riley, J., Ross, J.R., Rutter, D., Wells, A.U., Goller, K., du Bois, R., Welsh, K. (2006). No pain relief from morphine? Individual variation in sensitivity to morphine and the need to switch to an alternative opioid in cancer patients. Support Care Cancer 14, 56–64. Schulz, E., Tiemann, L., Schuster, T., Gross, J., Ploner, M. (2011). Neurophysiological coding of traits and states in the perception of pain. Cereb Cortex 21, 2408–2414. Sommer, M., de Rijke, J.M., van Kleef, M., Kessels, A.G.H., Peters, M.L., Geurts, J.W.J.M., Gramke, H.-F., Marcus, M.A.E. (2008). The prevalence of postoperative pain in a sample of 1490 surgical inpatients. Eur J Anaesthesiol 25, 267–274. Staahl, C., Christrup, L.L., Andersen, S.D., Arendt-Nielsen, L., Drewes, A.M. (2006). A comparative study of oxycodone and morphine in a multi-modal, tissue-differentiated experimental pain model. Pain 123, 28–36. Staahl, C., Drewes, A.M. (2004). Experimental human pain models: A review of standardised methods for preclinical testing of analgesics. Basic Clin Pharmacol Toxicol 95, 97–111. Tracey, I. (2008). Imaging pain. Br J Anaesth 101, 32–39. Tremblay, J., Hamet, P. (2010). Genetics of pain, opioids, and opioid responsiveness. Metabolism 59(Suppl 1), S5–S8. VanDenKerkhof, E.G., Hopman, W.M., Goldstein, D.H., Wilson, R.A., Towheed, T.E., Lam, M., Harrison, M.B., Reitsma, M.L., Johnston, S.L., Medd, J.D., Gilron, I. (2012). Impact of perioperative pain intensity, pain qualities, and opioid use on chronic pain after surgery: A prospective cohort study. Reg Anesth Pain Med 37, 19–27.
Eur J Pain 19 (2015) 1552--1561
1561
Machine learning on encephalographic activity may predict opioid analgesia M. Gram1, C. Graversen1, A.E. Olesen1,2, A.M. Drewes1,3 1 Mech-Sense, Department of Gastroenterology and Hepatology, Aalborg University Hospital, Denmark 2 Department of Drug Design and Pharmacology, Faculty of Health and Medical Sciences, University of Copenhagen, Denmark 3 Clinical Institute, Aalborg University Hospital, Denmark
Correspondence A.M. Drewes E-mail: [email protected] Funding sources This work was supported by funding from the Danish Council for Strategic Research (grant number 10-092786). Conflicts of interest None declared.
Accepted for publication 23 April 2015 doi:10.1002/ejp.734
Abstract Background: Opioids are used for the treatment of pain. However, 30–50% of patients have insufficient effect to the opioid initially selected by the physician, and there is an urgent need for biomarkers to select responders to the most appropriate drug. Since opioids mediate their effect in the central nervous system, this study aimed to investigate if electroencephalography (EEG) during rest or pain before treatment could predict the analgesic response. Methods: EEG from 62 channels was recorded in volunteers during rest and tonic pain (cold pressor test). Morphine (30 mg) or placebo was then administered, and the pain test repeated 60 min after. Washout period between drugs was 7 days. Based on pain ratings, subjects were stratified into responders and non-responders. Spectral analysis was performed on the EEG. Conventional statistics on group basis were used and, furthermore, the most discriminative EEG features were subjected to support vector machine classification to predict the response for the individual subjects. Results: Conventional statistics on the frequency bands revealed no differences between responders and non-responders. On the individual basis, no differences between groups were found using resting EEG. However, EEG during cold pain was able to classify responders with an accuracy of 72% (p = 0.01) and the result was reproducible using baseline data from both study days. Conclusions: Machine learning based on EEG before treatment enabled separation between responders and non-responders. This study represents the first step towards the prediction of opioid analgesia based on EEG features prior to drug administration, and advocates for the use of machine learning in future studies. For this article, a commentary is available at the Wiley Online Library.
1. Introduction Opioids are the drug of choice for moderate-tosevere pain (Liu and Wu, 2007). Despite major advances in the understanding of pain, treatment of acute pain such as in the postoperative conditions still remains unsatisfactory (Dolin et al., 2002; Sommer et al., 2008). The inadequate treatment has a 1552 Eur J Pain 19 (2015) 1552--1561
negative impact on patients both immediately following surgery, but also in the development of persistent long-term pain (VanDenKerkhof et al., 2012). On an individual level, there is a difference in the analgesic response to a given opioid. Various factors such as gender, age, and genetic variation can affect the analgesic response. The genetic variation can influence pharmacokinetics (e.g. drug transporters, drug-metabolizing enzymes and tolerance) and/or © 2015 European Pain Federation - EFICâ
M. Gram et al.
What’s already known about this topic • Pain patients often experience insufficient pain relief from opioids. • Response to opioids varies highly between individuals. • Accurate prediction of analgesic effect from opioids would enable more efficient pain therapy. What does this study add
• Electroencephalography shows promise for separating morphine responders from nonresponders. • Machine learning methods are required to accurately assess the complex relationships in the EEG.
pharmacodynamics (e.g. opioid receptor and catechol-O-methyltransferase enzymes) (Nielsen et al., 2014). Hence, in a study treating cancer pain with morphine, 74% of patients were well-treated initially with morphine (Riley et al., 2006). However, when switching the nonresponsive patients to other opioids, 96% experienced successful pain control (Riley et al., 2006). Therefore, personalizing pain treatment and thus initially enabling choice of the correct opioid shows great promise in optimizing pain treatment. Several attempts to predict analgesic efficacy has already been made using quantitative sensory testing (QST) (Grosen et al., 2013; Olesen et al., 2013). However, results have been contradictory between different QST modalities and study groups, indicating that while QST measures may be valuable in personalized medicine, there is still need for improvement and/or other methods to predict the effect of treatment (Grosen et al., 2013). Pain is a highly complex and individual experience, influenced by many factors which together has been shown to form a cerebral “signature”, responsible for producing the unique experience of pain (Tracey, 2008). This cerebral signature can be assessed using positron emission tomography and magnetic resonance imaging (Ahmedzai, 2013), and a better understanding of the inter-individual differences in brain function is thought to pave the way for personalized treatment of pain (Bruehl et al., 2013). A more clinically feasible method for assessing cerebral biomarkers could be electroencephalography (EEG) which also has considerable lower costs compared to imaging
© 2015 European Pain Federation - EFICâ
Machine learning on encephalographic activity may predict analgesia
solutions. EEG has been shown to predict the experience of pain within single subjects, underlining its ability to assess individual differences in cerebral function (Schulz et al., 2011). Therefore, the EEG might be used to differentiate between potential responders and non-responders to opioid treatment. The EEG is a composite signal reflecting the sum of a variety of neuronal activities. Therefore, the response to pain has traditionally been directed in a more specific way using painful stimulations followed by recordings of evoked brain potentials (Frøkjær et al., 2011). Although the “signal to noise ratio” of phasic stimulations are likely better than the EEG obtained during tonic pain, such brief stimuli may be less unsuited to simulate natural pain conditions dominated by hyperalgesia and affective feelings. This has moved quantitative sensory testing towards more tonic pain models (Nir et al., 2012). Especially, the cold pressor test has been widely used since it has been shown to be highly reliable and to mimic clinical pain (Rainville et al., 1992; Lewis et al., 2012). Furthermore, endogenous opioids play a major role in activation of the descending noxious inhibitory control system. Hence, the cold pressor test which activates this system was used to assess the effect of morphine. Due to the complex nature of pain perception and analgesics, it is likely that true prediction can only be achieved using a combination of several features, which requires advanced methods of analysis such as machine learning techniques (Khodayari-Rostamabad et al., 2010). Such methods have previously been used to detect individual differences in pain experiences (Brown et al., 2011) and in contrast to conventional statistics on group basis outcome of the individual subject is predicted by the model. We hypothesized that classification of pre-treatment EEG during tonic pain with a support vector machine (SVM) can be used to predict the response to morphine treatment. Thus, this study aimed to investigate if (1) SVM classification based on EEG recorded prior to morphine administration would be superior to conventional group-based statistics to differentiate between morphine responders and non-responders and (2) to investigate reproducibility and robustness of the model using data in the same subjects from another recording day.
2. Methodology The study was conducted between November 2010 and April 2012 in the research laboratory at
Eur J Pain 19 (2015) 1552--1561
1553
M. Gram et al.
2.5 Cold pressor test The cold pressor test was performed using circulated water bath (Grant, Fischer Scientific, Slangerup, Denmark). The water was cooled to 2 °C and the subjects immersed their left hand up to the wrist for 2 min while water was circulated. Subjects rated their pain continuously on a handheld electronic device with a visual analogue scale (VAS) going from no feeling of pain (0) to the worst imaginable pain (10). The pain ratings were electronically sampled continuously with a frequency of 10 Hz.
2.6 Stratification Subjects were stratified into responders and nonresponders based on their relative change in cold pain perception from before to after morphine administration. The area under the VAS curve (AUC) was calculated before and after drug administration. Afterwards, the relative change from baseline (in percentage) to post-drug administration was calculated. Subjects were labelled as responders if their AUC decreased by 5% or more after administration of morphine.
2.7 EEG recordings before treatment On both days, EEG was recorded in a dimly lit room, first during a resting period and then during the cold pressor test. EEG was recorded from a standard 62-channel cap (Quick-Cap International, Neuroscan, El Paso, TX, USA), amplified digitally on a Synamps 2 system (Neuroscan Compumedics, El Paso, TX, USA) and recorded for later analysis (Neuroscan 4.3.1, Neuroscan, El Paso, TX, USA). During the resting EEG recordings, subjects were instructed to keep their eyes open while minimizing eye blinking during the 2.5 min period. EEG during cold pain was recorded by starting the recording simultaneously as the subject submerged their hand in cold water.
2.8 Pre-processing The data were first pre-processed in the Neuroscan EEG software in the following steps: (1) zero-phase shift notch filtering (49–51 Hz) using a finite impulse-response filter with a slope of 24 dB/octave; (2) zero-phase shift band-pass filtering (1–70 Hz) using a finite impulse-response filter with a slope of 12 dB/octave; (3) blinded visual inspection of data quality for all channels using linked-ear reference. Channels with abnormal signals were discarded and © 2015 European Pain Federation - EFICâ
Machine learning on encephalographic activity may predict analgesia
replaced by signals interpolated from neighbouring electrodes; (4) Re-referencing to the average electrode; (5) Finally, resting EEG was cleaned by visually selecting 2 min of artefact-free EEG from the 2.5 min recording where the investigator was blinded to the origin of the recording.
2.9 Spectral analysis of EEG dynamics Spectral analysis of EEG amplitudes was carried out using Matlab 2012a (The Matworks, Inc., Natick, MA, USA) to calculate the relative EEG amplitude. The continuous wavelet transform was applied to the EEG signals from all channels. The complex Morlet wavelet was chosen for analysis with a bandwidth of 10 Hz and a centre frequency of 1 Hz. Scales for the transform were chosen to match frequencies ranging from 1 to 70 Hz, with a 0.5 Hz between-scale frequency interval. The absolute value of the obtained wavelet coefficients were used in the following analysis. To assess the static EEG spectral indices, the wavelet coefficients were divided into the following five standardized bands: delta (1–4 Hz), theta (4–8 Hz), alpha (8–12 Hz), beta (12–32 Hz) and gamma (32–70 Hz). The wavelet coefficients were averaged over time and scales contained within each frequency band were summed together to yield the absolute activity within each frequency band. The relative activity was calculated separately for each channel by dividing each frequency band with the total energy of all bands and multiplying by 100. The values then represent the percentage of total amplitude contained in each frequency band.
2.10 Conventional group-based analysis Data are presented as average SD. The EEG data from the morphine day prior to drug administration were first analysed using two-way repeated measures analysis of variance (ANOVA) using the group (responder or non-responder) as between-subject factor and the channel as within-subject factor. A p-value below 0.05 for the group factor was considered statistically significant. If statistical differences were found, the analysis was repeated using data from the placebo day.
2.11 Machine learning analysis To avoid over-fitting, the most discriminative features were selected using the criteria for joint mutual information, as this criterion has been found to provide the best selection for data sets with a limited Eur J Pain 19 (2015) 1552--1561
1555
Machine learning on encephalographic activity may predict analgesia
M. Gram et al.
Table 1 Results from the spectral analysis from each study day. Results are shown for responders and non-responders to morphine for each channel selected by “joint mutual information”. All values are shown as the percentage of total amplitude for each electrode. The electrodes are ranked in order of importance for the classification. Morphine day Selected channel
Responders
1. 2. 3. 4. 5. 6. 7.
31.2 44.6 38.4 32.9 31.9 31.4 16.1
Delta FT8 Delta FPZ Delta AF4 Delta FT7 Delta FC5 Delta FC3 Alpha C2
7.1 6.3 6.3 8.0 7.6 7.5 4.0
Placebo day Non-responders 34.4 46.4 41.3 32.3 29.7 29.3 16.3
7.0 6.2 6.8 7.6 6.5 6.3 3.8
Table 2 Confusion matrix from the classification using data prior to the administration of morphine and placebo.
Morphine day Classified responder Classified non-responder Placebo day Classified responder Classified non-responder
Responders
Non-responders
13 4
5 10
12 5
4 11
SVM classification using pre-treatment EEG during cold pain for the morphine day data was significant (p = 0.01) with 71.9% accuracy (positive predictive value = 70.0% and negative predictive value = 75.0%). Table 2 shows the 2 9 2 confusion matrix for the classification. Likewise, classification obtained with the data recorded on the other day prior to placebo drug administration provided a similar significant separation of the groups (p = 0.01) with 71.9% accuracy (positive predictive value = 75% and negative predictive value = 68.8%). The 2 9 2 confusion matrix is seen in Table 2.
4. Discussion To our knowledge, this is the first attempt to utilize EEG to identify specific brain patterns with advanced machine learning techniques in order to predict the individual response to opioid analgesia. The study showed that the support vector machine approach indeed was able to predict a positive response with 72% accuracy using EEG before treatment during tonic cold pain while conventional group-based analysis was insufficient. Furthermore, this result was reliable using data recorded on both study days. The primary features used for separation was the delta band from frontally and centrally placed electrodes. 1558 Eur J Pain 19 (2015) 1552--1561
Selected channel
Responders
1. 2. 3. 4. 5. 6. 7.
21.7 28.3 13.2 42.0 24.5 15.1 35.2
Delta T8 Delta FC1 Alpha T7 Delta FP2 Delta C6 Alpha TP7 Beta FT7
7.2 5.4 3.1 7.1 5.9 4.0 6.6
Non-responders 27.2 29.8 12.7 45.5 26.7 14.5 30.3
7.5 6.0 3.8 7.7 6.2 4.0 6.0
4.1 Prediction of morphine analgesia Resting EEG did not discriminate between responders and non-responders in this study, while this was possible using EEG during cold pain. The reason for this is unknown; however, reliability for resting EEG has been shown to be less than for EEG during cold pain (Gram et al., 2014). It could be speculated that the intense, painful stimuli increased attention levels to pain while simultaneously suppressing some conscious and non-conscious processes, which could confound the analysis (Gram et al., 2014). EEG during cold pain provided a reliable result, both in terms of accuracy and in the selected channels using data from both study days. For a prediction method to be clinically relevant, reliability is highly desirable, making EEG during cold pain better suited for prediction of analgesic treatment outcome. The spectral EEG indices recorded during rest and cold pressor test has previously been shown to be reliable between days (Gram et al., 2014). Furthermore, the current analysis yielded very similar results using data from both study days, both in terms of classification performance and in terms of the channels selected, indicating that the method is reliable. There are minor differences in the channels selected between the 2 days, mainly in the alpha and beta bands. However, some variation is to be expected in the feature selection procedure, since features are selected based on relevance to the classifier. However, features sharing too much information are excluded to prevent redundancy. Furthermore, the alpha and beta band features were generally selected last, meaning that they are of lesser relevance to the classification. As such, the variation in selected channels seems well within the expected variation for the method. The accuracy of the method in its current form, while statistically significant, is likely still too low to be considered usable in clinical practice. Further work should focus on increasing the accuracy of the © 2015 European Pain Federation - EFICâ
Machine learning on encephalographic activity may predict analgesia
number of samples (Brown et al., 2012). For both resting EEG and EEG during cold pain recorded immediately before administration of morphine, features were selected using the joint mutual information criterion and subsequently used for SVM classification. The number of features to be used was determined by investigating the accuracy of the classifier by gradually increasing the number of features up to 15. Accuracy was defined as the ratio between correctly classified subjects and total number of subjects in percentage. The number of features that yielded the highest accuracy on both study days was chosen for the final analysis due to the importance of reliability in the classifier. Classification was performed using the libSVM toolbox (version 3.20) for Matlab (Chang and Lin, 2011). A linear kernel function was used to avoid overfitting of the data (Gong et al., 2011). The cost parameter C of the SVM was set to 1 which is in the right order of magnitude (Auffarth et al., 2010). Classification accuracy was determined using leaveone-out cross-validation which consisted of removing one subject before training the SVM classifier using all remaining subjects. The last subject was then used to test the predictive capability of the classifier. This was repeated until all subjects had been left out, to assess the overall accuracy of the classifier (Gong et al., 2011). Accuracy, positive predictive value and negative predictive value of the classification were calculated as well as the statistical significance of the classification using chi-square where a p-value below 0.05 was used to indicate statistical significance.
M. Gram et al.
3. Results Thirty-nine subjects were enrolled in the study. Five subjects were excluded as they could not tolerate the cold pressor test for the full 2 min and thus had incomplete EEG recordings and two subjects were excluded due to poor EEG data quality. The subjects’ course through the study is summarized in Fig. 2 as a CONSORT flowchart. This resulted in 32 subjects in the final analysis (15 females and 17 males, aged 27.2 7.1 years).
3.1 Stratification Seventeen subjects decreased their AUC during the cold pressor test more than 5% during morphine treatment (AUC: 14.7% 6.9), and as such were labelled as responders. The remaining 15 subjects were labelled as non-responders (AUC: 1.1% 2.3). Response to placebo medication was not different between the responder and the non-responder group (2.2% 6.7 vs. 0.4% 3.5; p = 0.4).
3.2 Conventional group-based analysis The topographical representation of the spectral indices during resting EEG is shown on Fig. 3, while the spectral indices for EEG during cold pain is shown on Fig. 4. There were no differences between responders and non-responders for resting EEG spectral indices (all frequency bands F < 1.9; all p > 0.15). The same was true for EEG during cold pain (all F < 0.3; all p > 0.4).
Screened N = 93
Excluded N = 44
Included N = 49
Completed N = 39
Analyzed N = 32
Not eligible (N = 10) Not willing to participate (N = 34)
Drop outs N = 10
Decided to leave the study (N = 3) Did not tolerate experiment (N = 3) Due to side effects (N = 3) Due to accidental knee injury (N = 1)
Removed N = 7
Due to data quality (N = 2) Did not complete cold pressor test (N = 5)
1556 Eur J Pain 19 (2015) 1552--1561
Figure 2 CONSORT flow chart showing the subjects’ course during the study.
© 2015 European Pain Federation - EFICâ
M. Gram et al.
Machine learning on encephalographic activity may predict analgesia
Delta
Theta
Alpha
Beta
Gamma
Responders
Non-responders
Figure 3 Topographical plot of the spectral EEG indices from resting EEG for responders and non-responders to morphine analgesia. All values are the relative amplitude within the frequency band in percentage of total amplitude.
Delta
Theta
Alpha
Beta
Gamma
Responders
Non-responders
Figure 4 Topographical plot of the spectral EEG indices from cold pressor EEG for responders and non-responders to morphine analgesia. All values are the relative amplitude within the frequency band in percentage of total amplitude.
3.3 Machine learning analysis Resting EEG did not provide significant classification using any number of features (maximum accuracy: 12 features; 66% accuracy, p = 0.08); as such no further analysis was done on resting EEG. For EEG during cold pain the optimal number of features was found to be 7 since the accuracy was high using data from both days.
The seven features used in the classification were found in the delta band (FPZ, AF4, FC3, FC5, FT7 and FT8) and alpha band (C2) in the morphine arm (Fig. 5 shows the topographical distribution of selected channels). In the placebo arm, the features were found in the delta band (FP2, FC1, C6 and T8), the alpha band (TP7 and T7) and the beta band (FT7). Table 1 shows the values for all selected features, for each group.
Delta
Alpha
FPZ AF4
FT8
Day 1
FT7 FC5 FC3
C2
Delta
Alpha
Beta
Figure 5 Topographical representation of the selected channels in each electroencephalographic frequency band during cold pain using data from day 1 (prior to morphine administration) and day 2 (prior to placebo administration).
© 2015 European Pain Federation - EFICâ
Day 2
AF4
FT7
FC1 C6
T8
T7 TP7
Eur J Pain 19 (2015) 1552--1561
1557
Machine learning on encephalographic activity may predict analgesia
M. Gram et al.
Table 1 Results from the spectral analysis from each study day. Results are shown for responders and non-responders to morphine for each channel selected by “joint mutual information”. All values are shown as the percentage of total amplitude for each electrode. The electrodes are ranked in order of importance for the classification. Morphine day Selected channel
Responders
1. 2. 3. 4. 5. 6. 7.
31.2 44.6 38.4 32.9 31.9 31.4 16.1
Delta FT8 Delta FPZ Delta AF4 Delta FT7 Delta FC5 Delta FC3 Alpha C2
7.1 6.3 6.3 8.0 7.6 7.5 4.0
Placebo day Non-responders 34.4 46.4 41.3 32.3 29.7 29.3 16.3
7.0 6.2 6.8 7.6 6.5 6.3 3.8
Table 2 Confusion matrix from the classification using data prior to the administration of morphine and placebo.
Morphine day Classified responder Classified non-responder Placebo day Classified responder Classified non-responder
Responders
Non-responders
13 4
5 10
12 5
4 11
SVM classification using pre-treatment EEG during cold pain for the morphine day data was significant (p = 0.01) with 71.9% accuracy (positive predictive value = 70.0% and negative predictive value = 75.0%). Table 2 shows the 2 9 2 confusion matrix for the classification. Likewise, classification obtained with the data recorded on the other day prior to placebo drug administration provided a similar significant separation of the groups (p = 0.01) with 71.9% accuracy (positive predictive value = 75% and negative predictive value = 68.8%). The 2 9 2 confusion matrix is seen in Table 2.
4. Discussion To our knowledge, this is the first attempt to utilize EEG to identify specific brain patterns with advanced machine learning techniques in order to predict the individual response to opioid analgesia. The study showed that the support vector machine approach indeed was able to predict a positive response with 72% accuracy using EEG before treatment during tonic cold pain while conventional group-based analysis was insufficient. Furthermore, this result was reliable using data recorded on both study days. The primary features used for separation was the delta band from frontally and centrally placed electrodes. 1558 Eur J Pain 19 (2015) 1552--1561
Selected channel
Responders
1. 2. 3. 4. 5. 6. 7.
21.7 28.3 13.2 42.0 24.5 15.1 35.2
Delta T8 Delta FC1 Alpha T7 Delta FP2 Delta C6 Alpha TP7 Beta FT7
7.2 5.4 3.1 7.1 5.9 4.0 6.6
Non-responders 27.2 29.8 12.7 45.5 26.7 14.5 30.3
7.5 6.0 3.8 7.7 6.2 4.0 6.0
4.1 Prediction of morphine analgesia Resting EEG did not discriminate between responders and non-responders in this study, while this was possible using EEG during cold pain. The reason for this is unknown; however, reliability for resting EEG has been shown to be less than for EEG during cold pain (Gram et al., 2014). It could be speculated that the intense, painful stimuli increased attention levels to pain while simultaneously suppressing some conscious and non-conscious processes, which could confound the analysis (Gram et al., 2014). EEG during cold pain provided a reliable result, both in terms of accuracy and in the selected channels using data from both study days. For a prediction method to be clinically relevant, reliability is highly desirable, making EEG during cold pain better suited for prediction of analgesic treatment outcome. The spectral EEG indices recorded during rest and cold pressor test has previously been shown to be reliable between days (Gram et al., 2014). Furthermore, the current analysis yielded very similar results using data from both study days, both in terms of classification performance and in terms of the channels selected, indicating that the method is reliable. There are minor differences in the channels selected between the 2 days, mainly in the alpha and beta bands. However, some variation is to be expected in the feature selection procedure, since features are selected based on relevance to the classifier. However, features sharing too much information are excluded to prevent redundancy. Furthermore, the alpha and beta band features were generally selected last, meaning that they are of lesser relevance to the classification. As such, the variation in selected channels seems well within the expected variation for the method. The accuracy of the method in its current form, while statistically significant, is likely still too low to be considered usable in clinical practice. Further work should focus on increasing the accuracy of the © 2015 European Pain Federation - EFICâ
M. Gram et al.
classification. It is possible that the discussed limitations of the current study are also limiting the accuracy of the prediction and simply addressing these limitations in future investigations alone will improve the rate of correct classifications. Furthermore, the EEG features identified in this study can be combined in machine learning with other features such as pain thresholds, gender and age to further increase the accuracy of the algorithm. Several studies have already investigated prediction of opioid analgesia using QST assessments, although with varying results, for review see (Grosen et al., 2013). Genetic factors also affect the efficacy of opioids (Maier et al., 2002; Kalso et al., 2004; Tremblay and Hamet, 2010). Features could be extracted from the genetic profile of subjects such as the presence of specific mutations of genes coding for opioid receptors. This could have implications for the acute pain treatment such as in postoperative pain, where many patients experience inadequate pain relief (Dolin et al., 2002; Sommer et al., 2008). If investigated further, the method could potentially offer a low-cost and simple decision-aid to determine if a patient should go ahead with conventional morphine treatment or be switched to another opioid.
4.2 Methodological considerations Morphine was administered as a single dose of 30 mg as it is a clinically relevant dose. Subsequently subjects were tested after 60 min to determine effect. Classification accuracy could possibly improve if subjects were treated over a longer treatment period to achieve steady-states of drug concentrations where testing could be performed. This would possibly account for individual differences in morphine absorption and ensure high levels of drug concentration in the brain. However, in volunteers higher doses can hardly be used due to side effects (Olesen et al., 2014). We stratified subjects into two groups based on their response to morphine in the cold pressor test. Since the placebo response in the cold pressor test was low in general in healthy volunteers and there was no difference in the placebo response between responders and non-responders, the findings likely represent differences in morphine analgesia. Thus, as such, in the stratification process the individual’s response to placebo medication was not considered. Response level is difficult to determine for experimental pain models such as the cold pressor test. Clinically, a 30% decrease in clinical pain scores has traditionally been taken as an indicator for response (Olesen et al., 2013). Had this threshold been © 2015 European Pain Federation - EFICâ
Machine learning on encephalographic activity may predict analgesia
applied in this study, only a single subject would have been considered a responder to morphine, which is not likely with a clinically relevant dose of 30 mg. Lower reductions in pain thresholds are typically seen in quantitative sensory testing models (Olesen et al., 2014). Furthermore, the effect of morphine in volunteers typically differ from that in patients, where the pain is more complex and involve a variety of affective and cognitive components different from pain intensity in the experimental settings (Olesen et al., 2012). Therefore, a value of 5% reduction in AUC following morphine administration was used. This resulted in groups of similar size, which is preferable when using machine learning on a small data samples. EEG has previously been shown to predict the later experience of pain both in laser evoked brain potentials (Babiloni et al., 2008) and in tonic heat pain (Nir et al., 2012). As such it is clear that the immediate state of the brain prior to a painful stimulus affects the experience. Hence, it does not only vary between individuals, but subsequent painful experiences can vary for the individual as well. In the current study results were repeatable with more than a week between the recordings. As such the analgesic effect speaks to a more general trait for the individual (or a long-term state), which is of interest for predicting the analgesic effect in a reliable way. EEG has previously been shown to be sensitive to morphine analgesia (Olesen et al., 2012). On a physiological level it has been suggested that morphine has been shown to attenuate pain by interaction with the medial pain system, especially the anterior cingulate cortex and insula (Lelic et al., 2014). Studies show that morphine affects the brain by reorganizing the pain networks (Lelic et al., 2014). Individual differences in the brain could affect this interaction and increase or decrease the response to morphine. EEG measures the dynamic activity of the brain on a millisecond scale and therefore seems like an appropriate choice for investigating these individual differences. Other methods such as imaging techniques could also be used for investigation, but will incur higher costs for investigation. However, the machine learning approach allows features from other methods to be combined with features from the EEG for further increases in classification accuracy. The study was carried out using experimental pain models to assess the level of analgesia in order to minimize the influence of confounders by allowing accurate control of the painful stimulus (Staahl and Drewes, 2004; Olesen et al., 2012). Previously cold pressor tolerance has been used to predict effect of Eur J Pain 19 (2015) 1552--1561
1559
Machine learning on encephalographic activity may predict analgesia
oxycodone (Eisenberg et al., 2010), and it has been shown to be very sensitive to opioid analgesia (Jones et al., 1988; Koltzenburg et al., 2006; Olesen et al., 2012). Looking solely at EEG spectral indices it was not possible to discriminate on group level (e.g. using conventional group-based analysis). However, the current study demonstrated the possibility of selecting features from the EEG recorded during the cold pressor test using joint mutual information and machine learning to predict response or non-response to morphine for single individuals. The machine learning technique is a new approach compared to previous studies investigating the prediction of analgesic effect, where linear regression statistics are usually employed (Bares et al., 2008; Eisenberg et al., 2010). Using conventional group-based analysis for prediction of treatment outcome imposes several limitations; firstly the analysis does not directly assess the individual subject and assigns them a label such as machine learning methods. These studies utilize methods suitable only for detecting differences among groups, which is of little interest from a clinical perspective. The individual response to treatment is of more interest, and for this purpose machine learning methods are more suitable (Khodayari-Rostamabad et al., 2010). Secondly and most importantly, machine learning algorithms are able to assess multiple features simultaneously, which is beneficial for complex signals such as EEG, where the differences may only be explained by the simultaneous activity in several frequency bands and channels. A previous study was able to assess the analgesic effect of pregabalin by utilizing a SVM classifier and assessing all frequency bands simultaneously (Graversen et al., 2012). This could explain why in this study the group-based statistics failed to discover differences in all frequency bands, while feature extraction using joint mutual information and subsequent SVM classification provided significant classifications using data from both study days. With this in mind, machine learning algorithms seem preferential for future studies to predict treatment response of drugs due to the ability to assess complex interactions between features as well as enabling assessment on the individual level, such as has been shown within personalized medicine in psychiatric disorders (Khodayari-Rostamabad et al., 2010).
5. Conclusions Machine learning based on encephalographic activity before treatment enabled separation between responders and non-responders to morphine analge1560 Eur J Pain 19 (2015) 1552--1561
M. Gram et al.
sia on the individual basis, when conventional group-wise statistical methods failed. The study represents the first step towards prediction of opioid analgesia based on the recorded electrical brain activity prior to drug administration, and advocate for the use of machine learning in future studies. However, more studies with less and more focused electrodes and with larger populations are needed to prove the value of the method in a clinical setting. Author contributions A.M.D., A.E.O. and C.G. initiated the studies, carried out the pain testings and EEG recordings. M.G and C.G. designed and carried out the EEG spectral analysis. M.G. carried out the machine learning analysis. M.G., C.G., A.E.O. and A.M.D. evaluated results and drafted the manuscript. A.M.D., A.E.O. and C.G. conceived the study and participated in its coordination. All authors were involved in reading and approving the final manuscript.
References Ahmedzai, S.H. (2013). Personalized medicine–one size fits one: Tailoring pain therapy to individuals’ needs. J Pain Palliat Care Pharmacother 27, 83–85. Auffarth, B., L opez, M., Cerquides, J. (2010). Comparison of redundancy and relevance measures for feature selection in tissue classification of CT images. Adv Data Mining Appl 248–262. Babiloni, C., Del Percio, C., Brancucci, A., Capotosto, P., Le Pera, D., Marzano, N., Valeriani, M., Romani, G.L., Arendt-Nielsen, L., Rossini, P.M. (2008). Pre-stimulus alpha power affects vertex N2-P2 potentials evoked by noxious stimuli. Brain Res Bull 75, 581–590. Bares, M., Brunovsky, M., Kopecek, M., Novak, T., Stopkova, P., Kozeny, J., Sos, P., Krajca, V., H€ oschl, C. (2008). Early reduction in prefrontal theta QEEG cordance value predicts response to venlafaxine treatment in patients with resistant depressive disorder. Eur Psychiatry 23, 350–355. Brown, G., Pocock, A., Zhao, M.-J., Lujan, M. (2012). Conditional likelihood maximisation: A unifying framework for information theoretic feature selection. J Mach Learn Res 13, 27–66. Brown, J.E., Chatterjee, N., Younger, J., Mackey, S. (2011). Towards a physiology-based measure of pain: Patterns of human brain activity distinguish painful from non-painful thermal stimulation. PLoS ONE 6, e24124. Bruehl, S., Apkarian, A.V., Ballantyne, J.C., Berger, A., Borsook, D., Chen, W.G., Farrar, J.T., Haythornthwaite, J.A., Horn, S.D., Iadarola, M.J., Inturrisi, C.E., Lao, L., Mackey, S., Mao, J., Sawczuk, A., Uhl, G.R., Witter, J., Woolf, C.J., Zubieta, J.-K., Lin, Y. (2013). Personalized medicine and opioid analgesic prescribing for chronic pain: Opportunities and challenges. J Pain 14, 103–113. Chang, C.-C., Lin, C.-J. (2011). LIBSVM. ACM Trans Intell Syst Technol 2, 1–27. Dolin, S.J., Cashman, J.N., Bland, J.M. (2002). Effectiveness of acute postoperative pain management: I. Evidence from published data. Br J Anaesth 89, 409–423. Eisenberg, E., Midbari, A., Haddad, M., Pud, D. (2010). Predicting the analgesic effect to oxycodone by “static” and “dynamic” quantitative sensory testing in healthy subjects. Pain 151, 104–109. Frøkjær, J.B., Egsgaard, L.L., Graversen, C., Søfteland, E., Dimcevski, G., Blauenfeldt, R.A., Drewes, A.M. (2011). Gastrointestinal symptoms in type-1 diabetes: Is it all about brain plasticity? Eur J Pain 15, 249–257.
© 2015 European Pain Federation - EFICâ
M. Gram et al.
Gong, Q., Wu, Q., Scarpazza, C., Lui, S., Jia, Z., Marquand, A., Huang, X., McGuire, P., Mechelli, A. (2011). Prognostic prediction of therapeutic response in depression using high-field MR imaging. NeuroImage 55, 1497–1503. Gram, M., Graversen, C., Olesen, S.S., Drewes, A.M. (2014). Dynamic spectral indices of the electroencephalogram provide new insights into tonic pain. Clin Neurophysiol 126, 763–771. Graversen, C., Olesen, S.S., Olesen, A.E., Steimle, K., Farina, D., Wilder-Smith, O.H.G., Bouwense, S.A.W., van Goor, H.Drewes, A.M. (2012). The analgesic effect of pregabalin in patients with chronic pain is reflected by changes in pharmaco-EEG spectral indices. Br J Clin Pharmacol 73, 363–372. Grosen, K., Fischer, I.W.D., Olesen, A.E., Drewes, A.M. (2013). Can quantitative sensory testing predict responses to analgesic treatment? Eur J Pain 17, 1267–1280. Jones, S.F., McQuay, H.J., Moore, R.A., Hand, C.W. (1988). Morphine and ibuprofen compared using the cold pressor test. Pain 34, 117–122. Kalso, E., Edwards, J.E., Moore, R.A., McQuay, H.J. (2004). Opioids in chronic non-cancer pain: Systematic review of efficacy and safety. Pain 112, 372–380. Khodayari-Rostamabad, A., Hasey, G.M., Maccrimmon, D.J., Reilly, J.P., de Bruin, H. (2010). A pilot study to determine whether machine learning methodologies using pre-treatment electroencephalography can predict the symptomatic response to clozapine therapy. Clin Neurophysiol 121, 1998–2006. Koltzenburg, M., Pokorny, R., Gasser, U.E., Richarz, U. (2006). Differential sensitivity of three experimental pain models in detecting the analgesic effects of transdermal fentanyl and buprenorphine. Pain 126, 165–174. Lelic, D., Olesen, A.E., Gregersen, H., Dahan, A., Kolesnikov, Y., Drewes, A.M. (2014). Morphine modifies the cingulate-operculum network underlying painful rectal evoked potentials. Neuropharmacology 77, 422–427. Lewis, G.N., Heales, L., Rice, D.A., Rome, K., McNair, P.J. (2012). Reliability of the conditioned pain modulation paradigm to assess endogenous inhibitory pain pathways. Pain Res Manag 17, 98–102. Liu, S.S., Wu, C.L. (2007). The effect of analgesic technique on postoperative patient-reported outcomes including analgesia: A systematic review. Anesth Analg 105, 789–808. Maier, C., Hildebrandt, J., Klinger, R., Henrich-Eberl, C., Lindena, G. (2002). Morphine responsiveness, efficacy and tolerability in patients with chronic non-tumor associated pain – Results of a double-blind placebo-controlled trial (MONTAS). Pain 97, 223–233. Nielsen, L.M., Olesen, A.E., Branford, R., Christrup, L.L., Sato, H., Drewes, A.M. (2014). Association between human pain-related genotypes and variability in opioid analgesia: An updated review. Pain Pract http://onlinelibrary.wiley.com/doi/10.1111/papr.12232/ references.
© 2015 European Pain Federation - EFICâ
Machine learning on encephalographic activity may predict analgesia
Nir, R.-R., Sinai, A., Moont, R., Harari, E., Yarnitsky, D. (2012). Tonic pain and continuous EEG: Prediction of subjective pain perception by alpha-1 power during stimulation and at rest. Clin Neurophysiol 123, 605–612. Olesen, A., Andresen, T., Staahl, C., Drewes, A. (2012). Human experimental pain models for assessing the therapeutic efficacy of analgesic drugs. Pharmacol Rev 64, 722–779. Olesen, A.E., Brock, C., Sverrisd ottir, E., Larsen, I.M., Drewes, A.M. (2014). Sensitivity of quantitative sensory models to morphine analgesia in humans. J Pain Res 7, 717–726. Olesen, A.E., Staahl, C., Arendt-Nielsen, L., Drewes, A.M. (2010). Different effects of morphine and oxycodone in experimentally evoked hyperalgesia: A human translational study. Br J Clin Pharmacol 70, 189–200. Olesen, S.S., Graversen, C., Bouwense, S.A.W., van Goor, H., WilderSmith, O.H.G., Drewes, A.M. (2013). Quantitative sensory testing predicts pregabalin efficacy in painful chronic pancreatitis. PLoS ONE 8, e57963. Rainville, P., Feine, J.S., Bushnell, M.C., Duncan, G.H. (1992). A psychophysical comparison of sensory and affective responses to four modalities of experimental pain. Somatosens Mot Res 9, 265–277. Riley, J., Ross, J.R., Rutter, D., Wells, A.U., Goller, K., du Bois, R., Welsh, K. (2006). No pain relief from morphine? Individual variation in sensitivity to morphine and the need to switch to an alternative opioid in cancer patients. Support Care Cancer 14, 56–64. Schulz, E., Tiemann, L., Schuster, T., Gross, J., Ploner, M. (2011). Neurophysiological coding of traits and states in the perception of pain. Cereb Cortex 21, 2408–2414. Sommer, M., de Rijke, J.M., van Kleef, M., Kessels, A.G.H., Peters, M.L., Geurts, J.W.J.M., Gramke, H.-F., Marcus, M.A.E. (2008). The prevalence of postoperative pain in a sample of 1490 surgical inpatients. Eur J Anaesthesiol 25, 267–274. Staahl, C., Christrup, L.L., Andersen, S.D., Arendt-Nielsen, L., Drewes, A.M. (2006). A comparative study of oxycodone and morphine in a multi-modal, tissue-differentiated experimental pain model. Pain 123, 28–36. Staahl, C., Drewes, A.M. (2004). Experimental human pain models: A review of standardised methods for preclinical testing of analgesics. Basic Clin Pharmacol Toxicol 95, 97–111. Tracey, I. (2008). Imaging pain. Br J Anaesth 101, 32–39. Tremblay, J., Hamet, P. (2010). Genetics of pain, opioids, and opioid responsiveness. Metabolism 59(Suppl 1), S5–S8. VanDenKerkhof, E.G., Hopman, W.M., Goldstein, D.H., Wilson, R.A., Towheed, T.E., Lam, M., Harrison, M.B., Reitsma, M.L., Johnston, S.L., Medd, J.D., Gilron, I. (2012). Impact of perioperative pain intensity, pain qualities, and opioid use on chronic pain after surgery: A prospective cohort study. Reg Anesth Pain Med 37, 19–27.
Eur J Pain 19 (2015) 1552--1561
1561
