Pain in patients should be controlled for physiological reasons as well as for moral, humanitarian and ethical reasons. Therefore, obtaining information about the responsiveness level in patients under minimally invasive procedures, such as endoscopies, is one of the most important requirements for the anesthesiologist in order to guarantee a comfortable state for the patient during the medical procedure. Sedation-analgesia is a continuum that comprises four sedation levels from minimal sedation (anxiolysis) to general anesthesia [1
]. The first level corresponds to a drug-induced state during which ventilatory and cardiovascular functions are unaffected and patients can respond normally to verbal commands, although cognitive function and coordination may be impaired. Moderate sedation/analgesia is the next level that relates to a drug-induced depression of consciousness during which ventilation is adequate and cardiovascular function is usually maintained, avoiding any interventions to maintain an open airway. Patients can respond purposefully to verbal commands, either alone or accompanied by light tactile stimulation. The following level is deep sedation/analgesia, a drug-induced depression of consciousness during which cardiovascular function is usually maintained but the ability to independently maintain ventilatory function may be impaired, requiring the assistance in maintaining a patent airway. In this level, patients cannot be easily aroused but respond purposefully following repeated or painful stimulation. General Anesthesia corresponds to the last level, a drug-induced loss of consciousness during which patients are not arousable, even by painful stimulation. This level is characterized by a depressed spontaneous ventilation, requiring assistance in maintaining a clear airway and positive pressure ventilation in the patients, and also the cardiovascular function might be depressed as well as the cardiovascular control.
Adverse drug responses can be detected and treated in a timely manner by monitoring the level of consciousness, thus avoiding many of the complications associated with sedation and analgesia [1
]. Several methods have been developed for the noninvasive assessment of the level of consciousness during general anesthesia, which include pharmacodynamic assessment of the anesthetic and analgesic agents, study of hemodynamic evolution during anesthesia and evaluation of the modification of the electroencephalographic (EEG) activity [2
]. Since the main action of anesthetic agents occurs in the brain, the analysis of the EEG signals is one of the most applied techniques to monitor the sedation level. A central topic would be to assess hypnotic effect and pain/nociception-related activation from EEG. Most of the commercial monitors based on EEG have been developed for measuring the hypnotic effect with methods based on bispectrum (BIS, Aspect Medical System, Newton, MA, USA) [7
], entropy (M-entropy module, GE Healthcare, Helsinki, Finland) [9
], auditory evoked potentials (AEP, Monitor/2, Danmeter, Odense, Denmark) [10
], or power spectral analysis on different frequency bands (qCON, Quantium Medical, Barcelona, Spain) [12
Although different methods have been proposed over the last decade to monitoring nociception, this is a topic that has not been completely solved. Most of these methods include the analysis of signals from the brain activity, such as EEG or auditory-evoked potentials (AEP) [10
], or signals from the cardiovascular activity, such as heart rate variability (HRV) [13
], and also consider skin conductance (SC) [14
] and combinations of these signals [15
]. HRV and SC signals are influenced by the sympathetic activity, being the systems sensitive to other disturbances in addition to pain/nociception like changes in blood pressure or heart rate due to sympathomimetic drug delivery, perioperative bleeding or patient’s baseline condition (hypertension and arrhythmias of diverse etiology). There is not consensus in the relationship between different levels of analgesia and changes in EEG features, and therefore methods based on EEG signals tend to be unspecific. The AEP signals are weakly coupled to the levels of analgesia and its amplitude is very small, being difficult to record in clinical practice without significant noise levels [16
]. Recently, studies based on time-frequency representation [17
] and auto-mutual information function (AMIF) [18
] were applied in order to determine associated changes between EEG spectrum and EEG complexity with the prediction of the levels of unconsciousness as measured via the Ramsay Sedation Scale (RSS). In addition, EEG single-scale and multi-scale entropy measures have been effective in tracking changes of drug concentration during general anesthesia [19
]. Similarly, a recent study [21
] was proposed to measure the information coupling of EEG signals in sevoflurane anesthesia by means of a modification in the AMIF, which included permutation entropy, obtaining a measure that can clearly distinguish the awake and anesthesia states. However the discrimination between sedation-analgesia levels with no nociceptive response or sluggish response to painful stimulation still remains an open problem.
The aim of the study is to evaluate complexity of EEG signals during sedation-analgesia using a multiscale approach in order to improve the prediction of pain responses. We exploited the refined multiscale entropy (RMSE) [22
], a refined version of the multiscale entropy (MSE) originally proposed in [23
]. RMSE was selected, instead of a more traditional entropy-based approach, because it assesses complexity of times series. This feature is extremely helpful because it is well-known that EEG dynamics is regulated by several mechanisms operating along different frequency bands. Furthermore, AMIF measure is also considered to give information about the complexity in EEG signals based on the information coupling [24
]. To evaluate the prediction of response to painful stimulus, a multivariate analysis is performed on RMSE variables, AMIF measures, power spectral measures and other parameters such as drug concentration and heart rate, in order to maximize specificity and sensitivity.
3.1. CeProp, CeRemi, BIS, Time and Spectral HR Indexes
shows the mean and standard deviation (mean ± std) of the predicted concentrations of propofol (CeProp
) and remifentanil (CeRemi
), BIS parameter, mean heart rate (mHR
), variability of the heart rate (sdHR
) and spectral power in bands δ, θ, α, and β. Each index was obtained from EEG windows of length of 1 min between 30 s and 90 s before the response annotation of RSS. The values of Pk
calculated with a univariate linear discriminant function are also indicated in this table for Trials 1 and 2.
It can be noted from Table 2
that the mean value of indexes BIS
was higher (p
-value < 0.05) in the responsive groups (2 ≤ RSS ≤ 5 and RSS = 5) than in the unresponsive group (RSS = 6) in both trials. Indexes as CeRemi
had a contrary behavior, showing lower mean value (p
-value < 0.05) in the responsive groups (2 ≤ RSS ≤ 5 and RSS = 5) than in the unresponsive group (RSS = 6). The index CePropo
showed only statistically significant differences in Trial 1, being the mean value lower in the responsive group (2 ≤ RSS ≤ 5) compared with the unresponsive group (RSS = 6). In Trial 1, the Pδ
showed higher values in the responsive group (2 ≤ RSS ≤ 5) than responsive group (RSS = 6), while it was lower in the responsive group (RSS = 5) than the unresponsive group (RSS = 6) for Trial 2.
In Trial 1, the BIS was the index with the best Pk (0.799), also presenting the highest Sen (75.7%). The best Spe (82.2%) was obtained with the Pβ index, although the sdHR index showed also high Spe (80.3%) but it had the worst Sen (37.2%). In Trial 2, the spectral power Pβ was the index with the best Pk (0.663), although CeRemi showed also an important Pk value (0.642). Again, the sdHR index had the best Spe value (78.2%).
3.2. Multiscale Entropy Analysis of EEG: RMSE and AMIF
Mean ± standard error values of the SampEn
computed along the time scale factor ts
) are shown in Figure 2
. These values were obtained from the EEG segments in responsive states (2 ≤ RSS ≤ 5 and RSS = 5 for Trials 1 and 2, respectively) and unresponsive states (RSS = 6 for both Trials). All the RMSE curves exhibited an initial fast increase at short and medium time scale from scale ts
= 1 to ts
= 4, followed by a slow decrease at long time scales, being the entropy values higher at long time scales (ts
= 4 to ts
= 20) than at short time scales (ts
= 1), in both responsive and unresponsive groups. The comparison between responsive and unresponsive states showed a larger difference for Trial 1 than for Trial 2. In general, entropy values were lower in unresponsive state compared with responsive state at short time scales ts
, whereas a contrary behavior was observed at long time scales, where entropy values were higher in the unresponsive group.
RMSE indexes with p
-value < 0.001 and Pk
> 0.60 in at least one Trial were included in Table 3
, which shows the mean and standard deviation (mean ± std), Pk
values for both trials, Trials 1 and 2. Only one of the indexes obtained from the slope of the RMSE course at long time scales fulfill the conditions of Table 3
, and this was RMSE
This slope showed higher absolute values in the responsive group than in the unresponsive group, indicating that RMSE in the responsive group decrease faster than in the unresponsive group as function of the long time scale. In Trial 1, the highest Pk
(0.754) was obtained at long time scales (ts
= 17), while in Trial 2 it was obtained at medium time scales (Pk
= 0.625 at ts
Indexes computed from the AMIF analysis with p
-value < 0.001 and Pk
> 0.60 are also included in Table 3
. In both Trials, it is observed that the values of the indexes FD
(Req = 0.5
and max(Req = 2
were higher in the responsive group compared with unresponsive group, where max
(Req = 2
shows a contrary behavior. For these indexes, the best Pk
(0.734) in Trial 1 was obtained with max
(Req = 2
, while FD
(Req = 0.5
gave the best results (Pk
= 0.642) in Trial 2.
The Pearson correlation analysis between indexes provided that the spectral index Pβ was highly correlated with RMSE1, RMSE2 and FD(Req = 0.5)TB, with the following correlations (p-value < 0.0005): ρ(Pβ, RMSE1) = 0.797; ρ(Pβ, RMSE2) = 0.869; ρ(Pβ, FD(Req = 0.5)TB) = 0.959. The non-linear index FD(Req = 0.5)TB was also correlated (p-value < 0.0005) with RMSE1 and RMSE2: ρ(FD(Req = 0.5)TB, RMSE1) = 0.812; ρ(FD(Req = 0.5)TB, RMSE2) = 0.842.
3.3. Multivariate Statistical Analysis
The results that were obtained using a multivariate discriminant analysis are presented in Table 4
. Only functions with Pk
> 0.60, Sen
> 60% and Spe
> 60%, simultaneously in at least one trial, are included in this table: f1
); and f5
). The correlation between the indexes of the linear functions was weak, being the highest 0.381 (p
-value < 0.0005) which states between RMSEα
, and max
The function f4, which includes RMSE and AMIF indexes, had the best performances (Pk = 0.802) in discriminating responsive from unresponsive group when 2 ≤ RSS ≤ 5 vs. RSS = 6 were considered in Trial 1, but it showed a relative low Pk (0.683) in relation with other functions when RSS = 5 vs. RSS = 6 were compared in Trial 2. On the other hand, the multivariable function f2, which combines RMSE indexes with CeRemi, yields the highest Pk (0.722) in Trial 2.
3.4. Validation in GAG Reflex during Endoscopy Tube Insertion
To validate the robustness of the proposed univariate indexes and functions fi
for Trials 1 and 2, the same indexes and functions were also considered to study GAG reflex during endoscopy tube insertion in the same database patients. Table 5
shows the mean ± std, Pk
(%) and Spe
(%) of the univariate indexes. In order to simplify the information, this table only contains the RMSE and AMIF indexes that were included in the multivariable functions that are shown in Table 4
. It can be observed that all statistically analyzed indexes showed significant differences between groups GAG = 1 vs.
GAG = 0 with the exception of CeProp
Comparing with the analysis in Trials 1 and 2, the validation of the univariate indexes showed a similar behavior in the study of the GAG reflex. Indeed, the mean value of indexes BIS, mHR, sdHR, Pβ and Pδ was higher (p-value < 0.001) in the responsive groups (GAG = 1) than in the unresponsive group (GAG = 0). Indexes such as CeRemi, Pα and Pθ had a contrary behavior, showing lower mean value (p-value < 0.001) in the responsive groups than in the unresponsive group. RMSE1 (SampEn at ts = 1) and RMSEα (slope of RMSE course between 5 ≤ ts ≤ 8) showed higher absolute values in the responsive group than in the unresponsive group. The index max(Req = 2)δ was higher in the responsive group than in the unresponsive group. The best Pk (0.766) and Sen (74.3%) values were reached with the index Pα, while the index sdHR showed the highest Spe value (76.6%).
The results of the Pearson correlation analysis between indexes indicated that the spectral index Pβ was highly correlated with RMSE1, RMSE2 and max(Req = 2), with the following correlations (p-value < 0.0005): ρ(Pβ, RMSE1) = 0.782; ρ(Pβ, RMSE2) = 0.872; ρ(Pβ, max(Req = 2)) = 0.948. The non-linear index FD(Req = 0.5)TB was also correlated (p-value < 0.0005) with RMSE1 and RMSE2: ρ(FD(Req = 0.5)TB, RMSE1) = 0.861; ρ(FD(Req = 0.5)TB, RMSE2) = 0.842.
Finally, the multivariate functions with the best performance in the GAG reflex study were: f3 = f(RMSE1, max(Re2)δ , mHR) and f4 = f(RMSE1, RMSEα, max(Re2)δ). The Pk, Sen and Spe values that were obtained with the function f3 were 0.723, 64.4% and 70.6%, respectively, and with function f4 were 0.827, 68.3% and 76.9%, respectively.
The prediction of response to noxious stimulation is an issue that has not been completely solved, particularly in sedation-analgesia procedures where the patients may be under a sedation state in which the patient cannot be aroused but respond purposefully following repeated or painful stimulation, as the tube insertion during endoscopies, a minimally invasive procedure. Therefore, physiological conditions in these sedation levels are very different from the conditions detectable under general anesthesia where the responses to noxious stimulation are absent. This study demonstrates that a multivariate function of complexity measures as RMSE and AMIF can predict the responsiveness to noxious stimulation, which can be used by anesthesiologist in order to guarantee a comfortable state for the patient during the medical procedure.
Specifically, the function f4
) had the best performances (Pk
> 0.8) in discriminating responsive from unresponsive groups when considering 2 ≤ RSS ≤ 5 vs.
RSS = 6 and GAG = 1 vs.
GAG = 0, while the function f2
) yielded the highest Pk
(0.722) when comparing RSS = 5 vs.
RSS = 6, being these values higher than the values computed with the BIS
index. It is well known that BIS
is able to describe hypnotic effect as it was confirmed by results of Trial 1 (Table 3
). However, as it was seen in Trial 2 where (2 ≤ RSS ≤ 5) and (RSS = 6) were compared or comparing (GAG = 1) vs.
(GAG = 0), the functions based on RMSE and AMIF indexes showed a better capability than BIS to describe the analgesic effect and then to predict the response to noxious stimulation.
Analyzing the results in Table 2
and Table 5
, the statistically significant differences of BIS
values between responsive and unresponsive groups indicate that higher sedation levels can be associated with low probability of nociception. Since in this study the patients were not under general anesthesia but only under sedation, the expected value of the BIS
fluctuates between 100 and 60. Low BIS
values are associated with higher sedation levels, which are ideally reached by an adequate combination of hypnotic and analgesic concentrations, generating less probability of nociception. Thereby, Brocas et al.
] found BIS
values significantly higher during a control period than an alfentanil period when evaluated the effect of an intravenous bolus of alfentanil on the variations in BIS
level. An increase in BIS
values was also observed after a control period tracheal suction. However, BIS
might have the same value for different concentrations of drugs and it is possible that in case of low doses of analgesia, a response to noxious stimuli might be observed even at low BIS
Several studies [36
] relate nociception with the parameters derived from the analysis of heart rate variability (HRV). It has been demonstrated that the tone of the autonomic nervous system is strongly influenced by anesthetic drugs. Therefore, the traditional parameters of HRV in time domain and frequency domain [39
] are influenced by changes in the depth of sedation [40
] and can act as indicators of inadequate analgesia [35
]. The studies [41
] showed that anesthesia induction decreased the heart rate and that during nociception, the HRV not change when the patient receives adequate analgesia. In this study, the lower values of meanHR
in unresponsive state (RSS = 6) confirm that high sedation level can be associated with low heart rate. Furthermore, the lower values of sdHR
in unresponsive state denote a lower variability of the heart rate when the analgesia is adequate. An increase of meanHR
from presence to absence of GAG reflex is most likely to be due to a stimulation of central noradrenergic neurons that could realize a kind of cortical awakens, which reflects adrenergic hyperactivation.
Concerning about EEG bands, changes in δ, θ and α activity have been associated with changes in the sedation levels [43
]. Indeed, an increase of the hypnotic and analgesic doses in patients moves EEG activity from a low-amplitude and high-frequency signal to a high-amplitude and low-frequency signal. Concretely, there is a reduction of the β activity and an increment of the α and δ activities when the levels of propofol anesthesia are incremented [46
]. In the present study, this behavior was corroborated by the statistical differences observed in the spectral power in α,
θ and β frequency bands (see Table 2
and Table 5
) between the responsive (2 ≤ RSS ≤ 5, RSS = 5 and GAG = 1) and unresponsive groups (RSS = 6, GAG = 0). The spectral power in δ band showed different behavior in 2 ≤ RSS ≤ 5 vs.
RSS = 6 and GAG = 1 vs.
GAG = 0 that between RSS = 5 vs.
RSS = 6. Specifically, the increase of the δ activities with high level of sedation was only observed in the comparison between RSS = 5 vs.
RSS = 6, while the δ power decreased when comparing 2 ≤ RSS ≤ 5 vs.
RSS = 6 and GAG = 1 vs.
GAG = 0. This behavior might be explained by assuming that the strong response to nail bed compression (RSS < 5) and the gag reflex after tube insertion (GAG = 1) are associated with lower sedation levels than the sluggish response to nail bed compression (RSS = 5), hence they might contain also ocular activity. In this way, slow eye movements can affect the spectral power in δ band by causing its increase also in low sedation levels RSS < 5 and GAG = 1. It is also important to remember that patients in this study were under sedation-analgesia procedure and not under general anesthesia, and therefore it was not observed a significant increment in δ power at high sedation levels.
The results of RMSE (Table 3
and Table 5
) indicate that the complexity of the EEG signal, in responsive and unresponsive groups, was higher in the low frequency bands (ts
> 5), corresponding to δ, θ and α bands, than in high frequency bands (ts
< 3). These changes were also observed in the index RMSEα
which is the slope of RMSE course corresponding to time scale factors ranging from ts
= 5 to ts
= 10, corresponding approximately to α band at a sampling frequency of 128 Hz. Besides that, complexity of the EEG signal was higher in responsive than unresponsive group at short time scales (ts
< 3). This behavior can be interpreted by considering that short time scales contain frequency bands where scalp and facial muscle artifact are presented, suggesting that this muscle activity, which is more important in patients with low sedation level, is the responsible for the higher complexity in the responsive state at those time scales. In this sense, it is possible to associate an increased activity in the facial muscles with a greater probability of pain.
At long time scales, in which muscle activity has been eliminated, the entropy values indicate that the EEG signal contains more regular patterns in the responsive state than in the unresponsive state. This regularity increases as the time scale is larger for the responsive state but remains almost constant for the unresponsive state. Therefore, the results demonstrated that patients in unresponsive group show a more complex EEG activity in low frequency bands than patients in responsive group. One explication of this result is related with the fact that EEG activity becomes slower as the sedation level increases, and thus, it is expected that patients in unresponsive group present a slower EEG signal than responsive group, increasing the signal complexity in low frequency bands. A similar situation was reported in a study during sevoflurane anesthesia where MSE curves showed an ascending or flat trend for deep anesthesia and a descending trend for waking state [20
Finally, regarding the AMIF indexes (Table 3
and Table 5
), it can be denoted that EEG behavior was more complex in total and in α frequency band for the responsive group compared with the unresponsive group, while the EEG of responsive group was more regular in δ band than unresponsive group. This behavior was in agreement with the results of RMSE at long time scale, which showed lower values, and then more regular EEG, in the responsive group compared to the unresponsive group.