Electroencephalography in delirium assessment: a scoping review

Six studies analyzed EEGs by using qualitative and quantitative methods [47, 53, 62, 63, 70, 73]. Eleven studies analyzed EEGs by only using qualitative methods [42, 43, 50, 51, 54, 59, 64, 65, 67, 69, 72]. Only three of these used a standardized classification system, i.e., the Mayo Clinic classification system [64, 65, 72]. Knauert et al. [51] and Azabou et al. [43] used an encephalopathy classification introduced by Synek et al. [74] in 1990. Azabou et al. [43] also used the Young classification [75]. Three studies used qualitative methods also for cEEG [51, 54, 67]. Sambin et al. [59] and Naeije et al. [54] searched for characteristic patterns of epileptic activity without describing alternative findings. Vacas et al. [67] used qualitative methods to label different sleep phases and quantify the amount of time spent in each phase. Allahyari et al. [42] examined patients with delirium tremens. They attribute slow waves to effects of medication and classified EEGs as either normal or abnormal. Only exemplary cases were presented in detail. Sweden and Mellerio [69] analyzed qualitative aspects of EEGs recorded during drug withdrawal states in patients with signs of delirium but also clinical signs of epilepsy. Therefore, EEG findings focus on typical epileptic discharges. Finally, Kimchi et al. [50] performed a thorough qualitative analysis and description of rEEGs. The six studies that combined qualitative and quantitative analysis used, for the qualitative part, similar methods as the ones mentioned above. Two used the Mayo Clinic classification system [47, 70]. Matsushima et al. [53] only described whether EEGs were classified as normal or showed a degree of slower activity. The other three studies [62, 63, 73] performed a thorough qualitative analysis. Of these, Thomas et al. [62] also quantified rate of reactivity, frequency variation, and delta excess.

Most studies (21/22) performing quantitative analyses of rEEGs or cEEGs used a frequency domain-based method that subdivides complex waveforms in specific frequency components by using Fast Fourier Transformation (FFT). The remaining one used a time domain-based method of waveform analysis that independently measures amplitude and duration of each wave, in order to detect changes that affect only one of these two components [53]. For technical aspects of FFT and waveform recognition, please see supplementary materials.

The most common findings in qualitative EEG analysis of delirious patients were occipital slowing, excess delta or theta, anteriorization, and loss of reactivity [43, 59, 62, 63, 65, 73]. In the studies by Knauert et al. [51] and Azabou et al. [43], most patients were diagnosed with moderate, moderate to severe, or severe encephalopathy.

With regard to choice of electrode derivations, most studies used the whole range of electrodes and compared each electrode to a reference. Others averaged electrodes of anatomical regions to search for differences. Six studies [35, 45, 46, 52, 53, 56] used two- or three-electrode derivations. More specifically, Hunter et al. [46], van der Kooi et al. [35], and Fleischmann et al. [45] examined different derivations among all included electrodes, while Koponen et al. [52], Matsushima et al. [53], and Numan et al. [56] just analyzed the T6-O2 or T5-O1 derivation, the C3-O1 derivation, or Fp2-Pz and T8-Pz derivations, respectively.

Hunter et al. [46] generated a ratio between slow (< 13 Hz) and fast (13-45 Hz) frequencies derived from the electrodes C3, P3 and T7. Van der Kooi et al. [35] evaluated quantitative EEG data of bipolar electrode derivations. They studied patients under two conditions to identify which setting achieves the highest accuracy in delirium detection: A) ICU patients with eyes open (15 derivations, since all frontal, temporal and parietal electrodes were excluded to avoid blinking artefacts), and B) ICU patients with their eyes closed (210 derivations from 21 electrodes: F10, F9, Fp2, Fp1, F8, F4, Fz, F3, F7, T8, C4, Cz, C3, T7, P8, P4, Pz, P3, P7, O2, O1). They showed that 60 s artifact-free EEG recordings in ICU patients with closed eyes could discriminate delirious from non-delirious patients by just using two electrodes. An increase in relative delta power of delirious patients in derivations P8-Pz and Fp2-Fpz showed the highest sensitivity (100%) and specificity (95-96%). Fleischmann et al. [45] did a similar in-depth analysis for each of the 210 derivations × 70 frequencies. They identified F3-C4, F3-P4, and O2-F3 at 2 Hz as best classifiers to distinguish patients with and without delirium. These results were confirmed when tested on an unmatched large sample of controls with normal EEGs and an even larger real-world population. Applied to the latter, F3-C4 and F3-P4 at 2 Hz achieved sensitivities of 100% and specificities of 91 and 93%, respectively. When combined with C3-O1 at 19 Hz, specificity increased to 95%. Thus, all three studies demonstrate high sensitivities and specificities for derivations with increased relative delta band in frontal and parieto-occipital regions.

Matsushima et al. [53] found similar results by positioning the two electrodes only in central and occipital regions. In their study the theta/alpha ratio was increased even prior to clinical delirium manifestations. However, this result must be interpreted with caution due to the small sample size (n = 20). In addition to the above-mentioned studies included in this review, researchers around Gen Shinozaki have applied the novel bispectral EEG to delirium [76‐79]. They showed that using two electrodes only, algorithms based on quantitative EEG can differentiate between delirious and non-delirious individuals as well as estimate prognosis and mortality of delirious patients [76‐78]. Furthermore, the bispectral EEG device shows benefits with regard to small size and simple application. Of note, these articles report only modified and pre-interpreted EEG signals and were thus excluded during the literature search.

Koponen et al. [52] also found significant results for a reduced delta frequency in the P7-O1 or P8-O2 derivations. Since other electrode combinations were not derived, a direct comparison with the above-mentioned results is not possible. Moreover, a number of the elderly patients included in the study by Koponen et al. [52] showed substantive cognitive decline, which may explain a proportion of EEG results [62, 70] affecting specificity for delirium detection. Numan et al. [56] found a significant increase in delta power (frequency 0-4 Hz or 0-6 Hz) of delirious patients by using Fp2-Fz and T8-Fz derivations.

With regard to the lower frequencies, most (17/22) quantitative studies showed an increase in relative and absolute power in spectral analysis in delta and theta (mostly in frontal regions), and a decreased relative and absolute power in alpha (mostly in occipital or parietal regions) in delirious compared to non-delirious patients [35, 45, 46, 48, 49, 52, 55‐58, 61‐63, 68, 70, 71, 73].

With regard to the higher frequencies, Fleischmann et al. [45] also highlighted the importance of a decrease in the relative beta power in detecting delirium, especially in the C3-O1 derivation. Nielsen et al. [73] and Hunter et al. [46] also observed a reduction of beta activity in qualitative analysis of EEGs recorded from delirious ICU patients. In the study by Numan et al. [56], a decrease in relative beta power was one of the best discriminators for delirium detection, as shown by a random forest classifier. Hunter et al. [46] report a substantial reduction of gamma power in five delirious compared to five non-delirious patients. However, there were no differences in gamma power in the much larger study by Fleischmann et al. [71]. Tanabe et al. [61] also observed a decrease in high frequencies among delirious patients.

Moreover, one study reports a decreased centroid frequency (i.e., frequency that divides area of the spectrum in two equal parts) [57], another a decreased peak frequency [52], and two studies a decreased mean frequency [52, 66]. Other parameters described are increased theta/alpha ratio [53], decreased alpha/theta ratio [52, 57], decreased (alpha + beta)/(theta + delta) ratio [52], and a decreased scaled alpha-to-delta ratio, defined as the ratio of EEG power in the alpha band and delta band [49]. Keijzer et al. [49] also looked at the fraction of time not spent in suppression in EEG, which was lower in delirious patients after cardiac arrest compared to non-delirious patients.

With regard to connectivity analyses, the studies by Numan et al. [55], van Dellen et al. [68], and Tanabe et al. [61] found a significantly lower average phase lag index (PLI) for the alpha frequency band in delirious compared to non-delirious patients.

The study by Numan et al. [55] also found loss of posterior-anterior directionality in the alpha band, and loss of integration of the network organization. The latter was shown by the comparison of minimum spanning tree (MST) measures between hypoactive delirium patients and non-delirious patients. Delirious patients showed a decrease in degree, leaf fraction, and maximum betweenness centrality in the alpha band during delirium. Similarly, also Numan et al. [55] found a disturbed posterior-anterior connectivity in the alpha band. Van Dellen et al. [68] also report a decreased path length in the alpha band of delirious patients compared to controls. On the other hand, the clustering coefficient and small-world index did not differ between the groups. With regard to directed connectivity, delirious patients in both studies demonstrated a loss of posterior to anterior orientation in the alpha band [55, 68]. In the study by van Dellen et al. [68], patients also showed a lower delta band directed PLI (dPLI) in anterior regions and a higher dPLI in central regions than non-delirious patients [68]. This may indicate a flow of information within the delta band towards frontal regions. There were no differences in posterior regions or in dPLIs of other bands [68]. Of note, Tanabe et al. [61] found an increased frontal functional connectivity in patients that developed a postoperative delirium. The authors hypothesize a compensatory mechanism for a decreased structural connectivity (most likely due to neurodegenerative processes). Impaired structural connectivity has been confirmed by DTI studies [61, 80].

Fleischmann et al. [71] found global alpha and regional beta band disconnectivity as well as theta band hyperconnectivity in delirious patients. Similarly, also Hunter et al. [46] reported a general disconnectivity in delirious patients. A link between disturbance of consciousness and disconnectivity in the alpha band has already been shown in studies on the effects of ketamine [81] and propofol [82]. Thus, Fleischmann et al. interpret their findings as a sign of disturbed consciousness in delirium. In fact, the abnormalities in connectivity were spread across multiple networks engaged in consciousness, attention, working memory, executive functioning, and salience detection. In summary, functional connectivity seems to be impaired in delirious patients. Differences in affected regions, direction of connectivity, and affected band ranges might be explained by small sample sizes, divergent patient populations, different methods and study designs.

Spectral EEG analysis of 51 ICU patients with hyperactive, hypoactive, and mixed types of delirium did not demonstrate any significant difference in relative alpha, beta, theta, or delta power, alpha/theta ratio, (alpha + beta)/(theta + delta) ratio, or mean frequency values between the different delirium types [52]. This result is limited by the small number of patients in the different categories and the fact that most patients were diagnosed with dementia, which may explain some of the EEG changes. In line with these results, Numan et al. [56] did not find a significant difference in relative delta power between different types of delirium by using three electrodes. Van Dellen et al. [68] also studied delirious patients with and without hallucinations. Presence of hallucinations did not make any difference in alpha band PLI values, path length in graph theoretical analysis, or anterior to posterior dPLI gradient. However, patients with hallucinations showed a significant lower clustering coefficient and small-world index compared to delirious patients without hallucinations.

Tanabe et al. [61] reported a high correlation between an increase of slow wave activity in occipital regions and deliriums severity. The highest correlation was found at electrode Oz. In the study of Knauert et al. [51], reduction or absence of K-complexes during delirium was associated with worse outcome. Moreover, absence of sleep spindles correlated with unfavorable modified Rankin Scale scores.

Matsushima et al. [53] reported a significant slowing in recordings from central and occipital electrodes prior to a clinically overt delirium. This was measured by the theta/alpha ratio derived by a quantitative waveform recognition method of individuals in a serial measurement over many days after surgery. As mentioned above, Tanabe et al. [61] observed an increased alpha power preoperatively in patients who developed delirium postoperatively. Nielsen et al. [73] found continuous delta or theta activity in cEEG, loss of beta activity, and reactivity in evolving delirium. Resolution of delirium was characterized by re-occurrence of beta and reduction in delta activity. In an early study by Jacobson et al. [47], delirious patients that improved in cognitive functioning up to 19 months after the initial testing showed a significant increase in relative alpha band power and brain map changes with reduction of theta and delta in follow-up EEGs. However, the results are limited by a small sample size (n = 34), selection bias, a number of confounders, and absence of reporting of a follow-up delirium score associated with the second EEG.

Through analysis of sleep patterns in EEG, Evans et al. [44] demonstrated that patients with delirium following routine surgery need more time to fall asleep and sleep less during the first night after surgery prior to presenting clinical signs of delirium. In addition, delirium severity was negatively correlated with the amount of sleep during this first night as well as latency to falling asleep. Vacas et al. [67] did not find any difference in polysomnographic variables in cEEG between ICU patients that developed delirium and those that did not. Results are limited by a small sample size (n = 23), missing temporal and occipital electrodes, and duration and time of EEG (only assessed the first day after surgery). Despite the limitations of both studies, loss of physiological sleep structure may be an early indicator of delirium. This is in line with a previous study that used actigraphy [83]. EEG seems to offer a benefit since it provides objective criteria for sleep. Furthermore, the validity of estimates of latency of sleep onset and total duration of sleep based on self-reports is limited [84]. Further research is needed to evaluate the role of EEG, time point of recording, number of electrodes, and type of analysis as an early predictor for the development of delirium.

Using 24 h cEEG, Sambin et al. [59] identified sporadic epileptiform discharges (SEDs) in ten of 50 (20%) delirious patients. Periodic discharges (PDs) were observed in eleven of 50 patients (22%), eight of which had generalized PDs (GPDs), and three had lateralized PDs (LPDs). Moreover, seven of 50 patients (14%) had seizures, six of which had NCSE. Similarly, in the study by Azabou et al. [43], five of 22 (23%) delirious and six of 42 (14%) non-delirious patients had PDs. In addition, seven of 22 (32%) delirious and four of 42 (10%) non-delirious patients had electrographic seizures. Naeije et al. [54] found NCSE in cEEGs of nine of 32 (28%) delirious patients in the emergency department. rEEG detected NCSE only in two (6%) of the 32 patients. Nielsen et al. [73] observed no evidence of NCSE in cEEG recordings of delirious patients. However, they detected lateralized or bilateral PDs in 13 of 66 (20%) delirious, and two of 36 (6%) non-delirious patients.

Although delirium has shown to be associated with ictal or post-ictal conditions [85], the role of interictal activity is less clear. The incidence of SEDs, LPDs, GPDs, or NCSE in delirium cannot be calculated by these studies, due to the small number of patients, selection bias and bias through concomitant treatments with antiepileptic medication. However, cEEG in this setting can improve the diagnostic performance.

Jacobson et al. [70] discriminated patients with dementia, delirium, both conditions, or no encephalopathy. More specifically, normal from encephalopathic records could be differentiated by Mini-Mental State Examination (MMSE) in 85% of cases, by relative power in alpha in 91%, and MMSE plus relative power in alpha in 94% of cases. Delirium and dementia could be differentiated by theta activity (89%), brain map rating (89%), or combinations of theta activity, brain map rating, and/or relative delta power (up to 93%). Similar results were achieved by Koponen et al. [52]. Based on relative delta/alpha2-power density during activation in qEEG, Thomas et al. [62] differentiated patients with dementia and delirium from patients with delirium in 83% of cases. Numan et al. [55] demonstrated a less integrated and less organized functional network of delirious patients compared to healthy controls as shown by the reduced degree, leaf fraction, and maximum betweenness centrality in the alpha band. However, similar results were obtained in patients with Lewy Body Dementia and Alzheimer’s Disease, limiting the specificity [86, 87]. A direct comparison between patients with delirium and patients with dementia using advanced network analysis methods with more electrodes is needed to evaluate the discriminatory potentials of EEG in this context.

In the study by Numan et al. [55], both delirious patients as well as patients recovering from anesthesia (both with altered consciousness) showed an increased relative delta power and reduced relative alpha power in spectral analysis. Although the reduction of relative alpha power was stronger among patients with hypoactive delirium, the discriminatory potential of a spectral analysis between delirious patients and patients during recovery from anesthesia was low. This was also the case when applying functional connectivity measures such as PLI and directionality of connectivity [55]. Neuroleptic drugs were previously said to influence EEG [88, 89]. However, van der Kooi et al. [35] and Koponen et al. [52] did not find differences in relative delta power reduction in delirious patients that were treated with either haloperidol or chlorpromazine-equivalents, compared to delirious patients that were not treated. Other medication that could influence the EEG were often not reported systematically.

Qualitative EEG parameters cannot effectively discriminate patients with delirium and patients with dementia [53, 62]. In contrast, quantitative methods such as computerized waveform analysis demonstrated a significant difference in the ratio of theta over alpha waves [53]. In addition, other quantitative parameters like the increase of the relative delta band or reduction of the relative alpha band showed much higher odds ratios and reached 67% sensitivity and 91% specificity in one study [62]. In two studies by Trzepacz et al. [64, 65], a qualitative EEG analysis of delirious patients after liver transplantation demonstrated sensitivities of 83 and 75%, and specificities of 78 and 88%, respectively, for reduced dominant posterior rhythm. Both studies are limited by small sample sizes (n = 12 and 18 delirious patients) and selection bias.

In two studies, an increase of relative theta power in quantitative analysis was identified as the most sensitive characteristic of delirious patients compared to non-delirious patients [62, 63]. Numan et al. [55] performed a random forest analysis including spectral analysis, functional and directed connectivity and network topology. 77% sensitivity and 95% specificity were reached. Relative alpha, beta, and delta powers were the best discriminators. Elaborate quantitative analysis of > 200 bipolar derivations in all frequency bands did not confirm the expected sensitivity of the relative power in the theta band [35]. One reason for this could be the different margins of the theta frequency. Thomas et al. [62, 63] sub-classified the theta frequency into a lower frequency part (3-5 Hz) and an upper frequency part (5-7 Hz). The high sensitivity applied only to the lower part, which is nearer to the delta frequency that showed the highest sensitivity in the study by van der Kooi et al. [35]. Moreover, Numan et al. [56] found a better discriminatory potential between delirious and non-delirious patients in a 1-6 Hz frequency range than 1-4 Hz. Van der Kooi et al. [35] achieved around 100% sensitivity and 95% specificity for certain bi-electrode derivations, as mentioned above. Fleischmann et al. [45] found similar sensitivity and specificity in a larger not pre-specified cohort. Still, there are limitations to both studies [35, 45] since the sensitivity of the confirmation test (i.e., CAM-ICU) based on which patients were divided into delirium and non-delirium groups has been shown to be only 47% in the routine clinical setting [31]. Finally, EEG was performed only once. Thus, longitudinal studies on non-pre-selected groups of patients with quantitative cEEG are needed to further investigate sensitivity and specificity of EEG in delirium detection.