- Email: [email protected]
New spectral thresholds improve the utility of the electroencephalogram for the diagnosis of hepatic encephalopathy
Accepted Manuscript New Spectral Thresholds Improve the Utility of the Electroencephalogram for the Diagnosis of Hepatic Encephalopathy Clive D. Jackson, Mikkel Gram, Edwin Halliday, Søren Schou Olesen, Thomas Holm Sandberg, Asbjørn Mohr Drewes, Marsha Y Morgan PII: DOI: Reference:
S1388-2457(16)30018-9 http://dx.doi.org/10.1016/j.clinph.2016.03.027 CLINPH 2007799
To appear in:
Clinical Neurophysiology
Accepted Date:
14 March 2016
Please cite this article as: Jackson, C.D., Gram, M., Halliday, E., Olesen, S.S., Sandberg, T.H., Drewes, A.M., Morgan, M.Y., New Spectral Thresholds Improve the Utility of the Electroencephalogram for the Diagnosis of Hepatic Encephalopathy, Clinical Neurophysiology (2016), doi: http://dx.doi.org/10.1016/j.clinph.2016.03.027
This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
1
New Spectral Thresholds Improve the Utility of the Electroencephalogram for the Diagnosis of Hepatic Encephalopathy Clive D Jacksona,1, Mikkel Gramb,2, Edwin Hallidaya,3, Søren Schou Olesenb,4, Thomas Holm Sandbergb,5, Asbjørn Mohr Drewesb,c,6, Marsha Y Morgand,7,*
a
Department of Neurophysiology, Royal Free London NHS Foundation Trust, Pond Street, London, NW3 2QG, UK
b
Mech-Sense, Department of Gastroenterology and Hepatology, Aalborg University Hospital, Denmark
c
Department of Clinical Medicine, Aalborg University, Aalborg, Denmark
d
UCL Institute for Liver and Digestive Health, Department of Medicine, Royal Free Campus, University College London, Hampstead, London NW32PF, UK
1
[email protected] [email protected] 3 [email protected] 4 [email protected] 5 [email protected] 6 [email protected] 2
*
Corresponding author:
Dr Marsha Y Morgan, UCL Institute of Liver and Digestive Health, Division of Medicine, Royal Free Campus, University College London, Rowland Hill Street, London, NW3 2PF, UK Tel.: +44 207 433 2873 E-mail: [email protected]
2
Highlights
• New spectral EEG thresholds for the diagnosis of any degree of hepatic encephalopathy have been identified and validated using a machine learning technique. • The performance characteristics of these new thresholds are better balanced than the thresholds currently employed and hence their adoption would enhance diagnostic utility. • Implementation of these new thresholds would not require any changes in data recording or collection.
3
Abstract Objective: The utility of electroencephalogram (EEG) for the diagnosis of hepatic encephalopathy, using conventional spectral thresholds, is open to question. The aim of this study was to optimise this diagnostic performance by defining new spectral thresholds. Methods: EEGs were recorded in 69 healthy controls and 113 patients with cirrhosis whose neuropsychiatric status was classified using clinical and psychometric criteria. New EEG spectral thresholds were calculated, on the parietal P3-P4 lead derivation using an extended multivariable receiver operating characteristic curve analysis. Thresholds were validated in a separate cohort of 68 healthy controls and 113 patients with cirrhosis.
The diagnostic
performance of the newly derived spectral thresholds was further validated using a machine learning technique Results: The diagnostic performance of the new thresholds (sensitivity 75.0%; specificity 77.4%) was better balanced than conventional thresholds (58.3%; 93.2%) and comparable to the performance of a machine learning technique (72.9%; 76.8%). The diagnostic utility of the new thresholds was confirmed in the validation cohort. Conclusions: Adoption of the new spectral thresholds would significantly improve the utility of the EEG for the diagnosis of hepatic encephalopathy. Significance: These new spectral EEG thresholds optimise the performance of the EEG for the diagnosis of hepatic encephalopathy and can be adopted without the need to alter data recording or the initial processing of traces.
Keywords: Diagnostic thresholds; EEG; hepatic encephalopathy; psychometry; spectral analysis; support vector machine learning.
4
Abbreviations PHES: Psychometric Hepatic Encephalopathy Score; ROC: Receiver Operating Characteristic; MV ROC: Multivariable ROC; SVM: Support Vector Machine; SEDACA: Short Epoch, Dominant Activity, Cluster Analysis
5
1. Introduction Hepatic encephalopathy is one of the major complications of cirrhosis.
In unselected
populations approximately 30 to 45% of patients will exhibit clinically apparent neuropsychiatric abnormalities, encompassing a wide spectrum of mental and motor disorders, while a further 22 to 75%, although apparently neuropsychiatrically unimpaired on clinical examination, will show significant abnormalities in both psychometric and neurophysiological performance (Dhiman et al., 2010; Ferenci et al., 2002; Vilstrup et al., 2014). The presence of hepatic encephalopathy, whether minimal or overt, has a considerable impact on the execution of complex tasks, such as driving (Bajaj et al., 2009; Schomerus et al., 1981); health-related quality of life (Groeneweg et al., 1998; Montagnese et al., 2009); patient safety (Roman et al., 2011); neurocognitive function post liver transplantation (Sotil et al., 2009); and, ultimately, survival (D’Amico et al., 2006; Stewart et al., 2007; Wong et al., 2015). Electroencephalography (EEG) provides information useful for detecting, assessing and monitoring this complication of cirrhosis (Montagnese et al., 2004).
The main
electrophysiological characteristic of hepatic encephalopathy is slowing of the mean frequency from the alpha range towards the theta and delta ranges (Niedermeyer, 1998a). The efficacy of the EEG for the diagnosis of hepatic encephalopathy is critically dependent on the type of analysis performed, as even those incorporating semi-quantitative classifications of the background frequency (Conn et al., 1977; Laidlaw, 1959), are subject to inter- and intraoperator variability. Thus, the reported sensitivity of visual analysis of the EEG, for the diagnosis of overt hepatic encephalopathy, ranges from 57 to 100% (Rehnstrom et al., 1977; Weissenborn et al., 1990), while the specificity ranges from 41 to 88% (Parson-Smith et al., 1957; Weissenborn et al., 1990). Spectral analysis provides an automated estimate of the dominant EEG frequency (Van der Rijt et al., 1984) and, as such, is less inter- and intraoperator dependent (Amodio et al., 1996). Nevertheless, although frequency thresholds for the
6
diagnosis of overt hepatic encephalopathy have been identified (Amodio et al., 1999; Van der Rijt et al., 1984) considerable variation is still observed in the reported diagnostic performance of spectral analysis, in this context, with sensitivities ranging from 43 to 100% (Van der Rijt et al., 1984; Weissenborn et al., 1990) and specificity from 64 to 81% (Weissenborn et al., 1990). To date, no thresholds have been identified, using conventional spectral analysis on the parietal P3-P4 lead derivation, to characterize patients with minimal hepatic encephalopathy. Short Epoch, Dominant Activity, Cluster Analysis (SEDACA) is a technique for spatiotemporal decomposition of the EEG (Jackson and Sherratt, 2004; Montagnese et al., 2007). SEDACA-derived spectral estimates allow differentiation of patients with minimal hepatic encephalopathy from a reference population, whereas no such differentiation is possible using conventional spectral analysis (Montagnese et al., 2007). However, the SEDACA-derived spectral thresholds have not been independently validated, to date, and the technique is used primarily in a research setting.
Machine learning techniques explore the construction and study of algorithms to learn from and make predictions on data (Ortiz-Rosario and Adeli, 2013). Such algorithms operate by building a model from input examples to make data-driven predictions or decisions, without employing any theoretical a priori assumptions. There has been very little exploitation of these techniques to study EEG performance in patients with cirrhosis (Amodio et al., 2006).
The aims of this study were (i) to derive new EEG spectral thresholds for the diagnosis of any degree hepatic encephalopathy on both the P3-P4 lead derivation and on SEDACA components; (ii) to validate the diagnostic performance of the newly derived thresholds in an independent cohort; and (iii) to validate the performance of the newly derived thresholds with a machine learning technique.
7
2. Subjects and methods 2.1. The study populations 2.1.1. The patient cohort The patient cohort comprised of 226 patients (149 men: 77 women; mean [range] age 54.8 [2680] yr) with biopsy-proven cirrhosis recruited from the Royal Free Hospital, London between 2008 and 2012. The aetiology of their liver injury was determined using clinical, laboratory, radiological and histological variables, whilst its severity was assessed using Pugh modification of the Child’s grading system (Pugh et al., 1973). All were clinically stable at the time of the study. Patients were excluded if they were <25 or >80 years of age; if they had suffered an episode of major hepatic decompensation within seven days of assessment; had hyponatraemia or renal failure; significant cardiac or respiratory failure; insulin-dependent diabetes mellitus; cerebrovascular disease; epilepsy; a history of significant head injury or other conditions likely to affect cerebral function. Patients were also excluded if they had misused alcohol or drugs in the three months preceding assessment; if their manual dexterity was impaired; if they could not speak English; or were taking psychoactive medications. 2.1.2. The reference population The reference population of 137 healthy volunteers (73 men:64 women: mean age 39 [17-75] yr) was recruited from amongst family, friends and staff working at the Royal Free Hospital, London and individuals who had experienced an isolated episode of fainting/dizziness but in whom clinical examination, the EEG, and cerebral imaging were completely normal. None had a history of liver disease, drank alcohol in excess of 20g daily or took prescription or overthe-counter medicines. 2.1.3. Creation of the identification and validation cohorts The patients were ranked by their raw score on the Psychometric Hepatic Encephalopathy Syndrome (PHES) test battery of psychometric tests (Weissenborn et al., 2001) and
8
alternatively assigned to either an identification or a validation cohort. The 137 reference individuals were ordered by age before alternative assignment to one of the two cohorts. The terms identification and validation cohorts were then applied to the patients and reference populations within each subgroup. 2.2. Assessment and classification of neuropsychiatric status Patients were clinically assessed by two hepatologists, working independently, and their mental state classified, using the West Haven Criteria (Conn et al., 1977) as either clinically unimpaired or as showing features of overt hepatic encephalopathy.
Psychometric
performance was assessed using the PHES battery (Weissenborn et al., 2001), which comprises five paper and pencil tests viz: digit symbol, number connection A and B, serial dotting and line tracing. The PHES data were adjusted and scored using UK normative data (Marks et al., 2008); composite scores of less than two standard deviations below mean reference values were considered abnormal. Neuropsychiatric status was classified, on the day of the study, in the patient population as: (i) unimpaired: no clinical evidence of hepatic encephalopathy and no psychometric abnormalities; (ii) minimal hepatic encephalopathy: no clinical abnormalities but impaired psychometric performance; (iii) overt hepatic encephalopathy: clinically evident, characteristic neuropsychiatric disturbances together with impaired psychometric performance. 2.3. Electroencephalography Electroencephalograms (EEGs) were recorded on one of two digital EEG systems, WalterGraphtek PL-Winsor (Walter-Graphtek GmbH, Emmendingen, Gerrmany) or MicroMed SystemPlus EVOLUTION (MicroMed. S.p.A., Mogliano Veneto (TV), Italy), using 23 silversilver chloride electrodes placed according to the International 10-20 System. The impedance of the electrodes was kept below 5 KΩ. Recordings were undertaken for 10 minutes, in a state
9
of eyes-closed, relaxed, wakefulness with a sample rate of 256 Hz and an on-line band pass filter of 0.05 to 70 Hz. 2.3.1. EEG spectral analysis on the P3-P4 derivation A 60 to 100 seconds section of eye-closed, artefact-free recording was identified on the P3-P4 derivation from each EEG.
The recordings were divided into two second epochs and
multiplied with a cosine window with a one second overlap. The Fast Fourier Transform was applied to obtain the spectral power in each 0.5 Hz bin which was averaged across all epochs to give an average power spectrum for each EEG. Estimates were also obtained for five further spectral parameters viz: Mean dominant frequency (MDF): weighted mean of frequencies in the 10 to 26.5Hz range Relative delta power percentage activity in the 1.0 to 3.5 Hz range; Relative theta power: percentage activity in the 4.0 to 8.0 Hz range; Relative alpha power: percentage activity in the 8.5 to 13.0 Hz range; Relative beta power: percentage activity in the 13.5 to 26.5 Hz range. Spectral analysis was applied to the P3-P4 lead derivations and the EEGs graded by applying the thresholds proposed by Amodio et al., (1999) on the P3-P4 lead. 2.3.2. EEG spectral analysis on the SEDACA components SEDACA (Short Epoch Dominant Activity Clustering Algorithm) is a method for decomposing, or separating, the EEG recording into its separate activities, i.e. alpha rhythms or muscle or eye artefact components.(Jackson and Sherratt, 2004) Prior to SEDACA analysis of the 60 to 100 seconds of EEG recording selected for the P3-P4 spectral analysis, an extra 10 seconds of EEG recording containing eye-movement artefact was added to enhance the separation of the eye-movement components from the other activities in the recording. After decomposition of the EEG by SEDACA, a single component was selected to represent each
10
record; the extra 10 seconds removed so further analysis was over the same 60 – 100 second time periods as used in the analysis of the P3-P4 lead derivation. Spectral thresholds comparable to those used in standard P3-P4 spectral analysis were applied, as previously proposed by Montagnese et al., (2006). 2.3.3. Machine learning analysis The EEGs were exported, as ASCII files, to Mech-Sense, Department of Gastroenterology and Hepatology, Aalborg University Hospital, Denmark for analysis using Support Vector Machine (SVM) Learning which was performed on the spectral data using the libSVM toolbox (Version 3.20) (Chang and Lin, 2011) in Matlab 2012a (The Matworks, Inc., Natick, MA, USA). A linear kernel function was used to avoid over-fitting of the data (Gong et al., 2011). The cost parameter C of the SVM was set to 1 (Auffarth et al., 2010). 2.4. Derivation of new spectral EEG thresholds New spectral thresholds, for the diagnosis of any degree of hepatic encephalopathy, were derived for MDF, relative theta and relative delta power, from the P3-P4 lead derivation and SEDACA components, in the identification cohort based on classical and modified Receiver Operating Characteristic (ROC) curve analyses. Subjects were grouped as follows: Group I - the reference population and neuropsychiatrically unimpaired patients. Group II –patients and minimal or overt hepatic encephalopathy. 2.4.1. Standard ROC curve analysis The threshold for each of the spectral variables was defined by the point on the ROC curve closest to the top left hand corner of the graph. 2.4.2. Modified ROC curve analysis Sets of potential thresholds were tested, each set comprising of a threshold on MDF, relative delta power and relative theta power. The optimal set of thresholds was defined by the point
11
on this ROC-like curve closest to the top left hand corner of the graph of sensitivity against 1specificity. These are referred to as the multivariable ROC (MV ROC) thresholds. 2.5. Application of EEG spectral thresholds The conventional and newly derived spectral thresholds for the diagnosis of any degree of hepatic encephalopathy were individually applied to the P3-P4 lead derivation and the SEDACA components. An EEG was considered abnormal if the MDF were below threshold or the relative theta power or relative delta power was above threshold. 2.6. Cross-validation of the spectral thresholds A Support Machine Vector (SVM) model was used to cross validate the new spectral thresholds. Differential weighting was applied to Groups 1 and 2 in the identification cohort to account for differences in sample size. Group 1, comprising of the reference population and the neuropsychiatrically unimpaired patients, was weighted by the number of subjects in Group 2, while Group 2, comprising the patients with minimal and overt hepatic encephalopathy, as the smaller of the two groups was weighted 1.4 times the number of subjects in Group 1. One SVM model was trained, based on each group, using the MDF, relative delta power and relative theta power, and subsequently each model was tested using data from both groups. 2.7. Performance and validation of the new spectral EEG thresholds The diagnostic performance of the new spectral EEG thresholds derived in the identification cohort, using ROC and MV ROC analysis, were applied to the validation population. Performance characteristics for the diagnosis of any degree of hepatic encephalopathy are reported as sensitivity, specificity, positive predictive value and negative predictive values for MDF, relative delta power and relative theta power and for a composite assessment in which an EEG is considered abnormal if the MDF or the relative delta power or the relative theta power is abnormal.
12
2.8. Statistical analysis The distribution of variables was tested using the Shapiro-Wilk’s test. Differences between non-normally distributed variables were examined using the Kruskal-Wallis test and subsequent between-groups comparisons performed using the Mann-Whitney U test with Bonferroni corrections. Differences between normally distributed variables were examined by one way ANOVA and subsequent between group comparisons performed using the Tukey test. The statistical analyses were undertaken using the software packages STATA version 11.2 (StataCorp LP, College Station, Texas) and R version 3.0.2 (CRAN.R-project.org/doc/FAQ/RFAQ.html). 2.9. Ethics The study was conducted according to the Declaration of Helsinki (Hong Kong Amendment) and Good Clinical Practice (European guidelines). The protocol was approved by the Royal Free Hampstead NHS Trust Ethics Committee and the equivalent body in Hannover. All participating subjects provided written, informed consent.
13
3. Results 3.1 The study populations The aetiology of the cirrhosis in the patient cohort was alcohol in 157 (69.5%) of the 226; alcohol and hepatitis C in 21 (9.3%); cryptogenic in 12 (5.3%); fatty liver disease in 11 (4.9%); hepatitis C/hepatitis B in eight (3.5%); primary biliary cirrhosis in seven (3.1%); chronic active hepatitis in five (2.2%); haemochromatosis in two (0.9%); and ‘other’ in three (1.3%). Functionally: 135 (59.7%) patients were classified as Child’s Grade A; 51 (22.6%) as Child’s Grade B; and 40 (17.7%) as Child’s Grade C. On the day of study: 127 (56.2 %) patients were classified as neuropsychiatrically unimpaired; 21 (9.3%) as having minimal and 78 (34.5%) as having overt hepatic encephalopathy. The patients with overt hepatic encephalopathy had, as expected, significantly less functional hepatic reserve than the patients in the other two groups, evidenced by a higher mean Pugh’s score (Table 1). The reference subjects were significantly younger than the patients with cirrhosis (Table 1). There were no significant differences in demographic and measurement variables between the identification and validation cohorts (Supplementary Table S1). 3.2. EEG analysis EEG data were available for spectral analysis, on the P3-P4 lead derivation, from all patients and reference individuals; EEGs were available for spectral analysis on SEDACA components from 218 (96.5%) and 129 (94.2%) patients and reference individuals, respectively. 3.3. EEG spectral variables in patients, by neuropsychiatric status The mean (± 1SD) relative theta power on both the P3-P4 lead derivation and the SEDACA components was significantly increased and the relative alpha power significantly decreased in
14
the patients compared to the reference population; the mean MDF on the SEDACA component was also significantly lower (Table 2). Changes were observed in the EEG spectral variables in the patients with cirrhosis in relation to their neuropsychiatric status (Figure 1). The mean MDF in the patients with overt hepatic encephalopathy was significantly reduced, relative to the reference population and the other patient subgroups, on both the P3-P4 lead derivation and the SEDACA components, primarily reflecting increases in the mean relative theta power and reductions in the mean relative alpha power (Table 2). Significant slowing of the MDF was also observed in the patients with minimal hepatic encephalopathy compared to both the neuropsychiatrically unimpaired patients and the reference population but only on the SEDACA components (Table 2). The mean theta power showed a progressive reduction across patient subgroups, by neuropsychiatric status but did not distinguish the unimpaired patients from those with minimal hepatic encephalopathy (Table 2). The pattern of spectral EEG changes observed in the identification and validation cohorts were comparable and mirrored the changes observed in the population overall (Supplementary Tables S2a and b). 3.4. Definition of new spectral EEG thresholds The new spectral thresholds differed from the conventional thresholds currently in use (Table 3). The ROC-derived thresholds for MDF were higher, on both the P3-P4 lead derivation and the SEDACA components, while the thresholds for delta and theta activity were considerably lower. The MV ROC-derived thresholds for MDF and delta activity, on P3-P4 lead derivation, were broadly comparable to those currently in use but the threshold for theta activity was again lower.
In contrast, the MV ROC-derived thresholds for MDF and theta activity on the
15
SEDACA components were lower than conventional thresholds while the threshold for delta activity was higher (Table 3). 3.5. Diagnostic performance of conventional and new spectral EEG thresholds 3.5.1. Identification cohort The performance of the conventional spectral thresholds on the P3-P4 lead derivation for the diagnosis of any degree of hepatic was characterised by high specificity and low sensitivity (Table 4). The performance of the conventional spectral variables on the SEDACA component was generally better with a notable increase in the sensitivity of the composite assessment which, nevertheless, remained only moderate. The diagnostic performance of the new ROC-derived thresholds varied considerably by spectral variable on both the P3-P4 lead derivation and the SEDACA components; the performance of the threshold for relative theta power was balanced with high sensitivity and high specificity while the composite assessments was characterised by high sensitivity and relatively poor specificity. The thresholds derived for the individual spectral indices using the MV ROC analysis provided comparable performance data, for MDF and relative delta power, on both the P3-P4 lead derivation and the SEDACA components, to the conventional thresholds, whilst the performance characteristics for relative theta power were the same as those provided by ROC analysis. The composite assessment showed balanced performance characteristics. The spectral variable with the best performance characteristics on both the P3-P4 lead derivation and the SEDACA components, irrespective of the applied threshold, was relative theta power. 3.5.2. Validation cohort
16
The performance of the conventional, ROC-derived and MV ROC-derived spectral thresholds in the validation cohort (Table 5) closely mirrored performance in the identification cohort (Table 4). The performance characteristics of the SVM modelling were similar to those of the MV ROC analysis in the validation cohort (Tables 4 and 5).
17
4. Discussion New spectral EEG thresholds for the diagnosis of any degree of hepatic encephalopathy have been identified and validated in two independent populations and by reference to performance variables obtained using SVM. The performance characteristics of these new thresholds are superior and better balanced than the thresholds currently employed; their adoption would significantly improve the utility of the EEG for the diagnosis of hepatic encephalopathy, in clinical practice. The EEG has been used to diagnose hepatic encephalopathy since the 1950s (Foley et al., 1950). In 1957, Parsons-Smith et al. provided a detailed description and grading of the EEG findings in patients with cirrhosis which correlated fairly well with the degree of clinical neuropsychiatric disturbance. They also observed EEG changes in patients with cirrhosis who had no clinical evidence of hepatic encephalopathy which, subsequently, was recognised as one of the defining features of minimal hepatic encephalopathy. The raw EEG signal is complex comprising an admixture of different activities. Thus, visual analytical techniques, even those incorporating semi-quantitative classifications of the background frequency (Conn et al., 1977; Laidlaw, 1959), are subject to inter- and intraoperator variability(Amodio et al. 1996).
Nevertheless, visual inspection, by a trained
neurophysiologist, is still one of the most frequently utilized methods of assessment of the EEG in this patient population. Under these circumstances an EEG is considered abnormal, and compatible with a diagnosis of a metabolic or drug-induced encephalopathy, if there is slowing of the rhythmicity with a dominant frequency below 8 Hz, although other changes, such as desynchronization and anteriorisation of the dominant activity may also be seen (Parson-Smith et al., 1957). The decision as to whether these changes represent hepatic encephalopathy is a clinical one based on exclusion of drug-induced and other causes of metabolic encephalopathies such as renal failure, hyponatraemia, carbon dioxide retention and
18
chronic hypoglycaemia (Bauer and Bauer, 1998; Niedermeyer, 1998b).
In patients with
cirrhosis, in whom other causes for the EEG changes have been excluded, the presence of compatible clinical symptoms and signs will support a diagnosis of overt hepatic encephalopathy or when absent, will identify the presence of minimal hepatic encephalopathy. Spectral analysis was first used to study the EEG in patients with cirrhosis in the late 1980s/ early 1990s. It is less inter- and intra-operator dependent than visual assessment (Amodio et al., 1999), provides an automated estimate of the dominant EEG frequency and allows the relative contributions of the different rhythms to the overall background frequency to be estimated (Van der Rijt et al., 1984). The recommended spectral thresholds for MDF, and the relative powers of delta and theta on the P3-P4 lead derivation were derived essentially to characterize patients with overt hepatic encephalopathy which explains why they are set so exactingly (Van der Rijt et al., 1984). Thus, while an EEG would be considered abnormal, on visual inspection, if the dominant frequency were below 8 Hz, the conventional spectral threshold is ≤6.8 Hz. To date, no thresholds have been identified, using conventional spectral analysis, to characterize patients with minimal hepatic encephalopathy. In consequence, the spectral thresholds used to detect more substantial degrees of neuropsychiatric impairment are applied. Spatio-temporal analysis of the EEG using SEDACA provides better diagnostic information than conventional spectral analysis (Jackson and Sherratt, 2004; Montagnese et al., 2007). Thus, SEDACA-derived spectral estimates correlate better with neuropsychiatric status and allow differentiation of patients with minimal hepatic encephalopathy from a reference population (Montagnese et al., 2007). However, this is primarily a research technique and is not generally available. The overall utility of the EEG for the diagnosis of hepatic encephalopathy is difficult to gauge in the absence of a diagnostic gold standard and is very dependent on the analytical technique
19
applied.
However, in patients with overt hepatic encephalopathy the reported diagnostic
sensitivity ranges from 43 to 100% while the specificity ranges from 41 to 88%, (Parson-Smith et al., 1957; Rehnstrom et al., 1977; Van der Rijt et al., 1984; Weissenborn et al., 1990). In addition, EEG abnormalities consistent with a diagnosis of hepatic encephalopathy are observed in from 8 to 40% of neuropsychiatrically unimpaired patients with cirrhosis, (Montagnese et al., 2004). The primary aim of the present study was to optimise the performance of the EEG by determining spectral thresholds for the diagnosis of any degree of hepatic encephalopathy. This was determined to be the most pragmatic approach as it is unlikely that a robust threshold could be identified which would differentiate minimal from overt hepatic encephalopathy. Thus, if the EEG were abnormal, classification of the patients as having minimal or overt hepatic encephalopathy would be made by the clinician on the basis of the history and clinical examination. The results of this study show that the method used to determine EEG spectral thresholds has a large effect on measured performance. Use of the conventional thresholds was associated with high specificity for the diagnosis of any degree of hepatic encephalopathy but poor to moderate sensitivity. The ROC-derived threshold for relative theta power had good diagnostic utility but the composite assessment, based on the performance of the three spectral variables, was characterised by high sensitivity and low specificity. Finally, use of the MV ROC-derived thresholds was associated with superior and better balanced diagnostic performance for both the individual and composite assessments on both the P3-P4 lead derivation and the SEDACA component. There was greater concordance in the diagnostic performance of the P3-P4 lead and SEDACA spectral estimates when utilizing the ROC- or MV ROC-derived thresholds. The validity of MV ROC-derived thresholds for the diagnosis of any degree of hepatic encephalopathy was confirmed in the validation cohort and further supported by the SVM
20
classification. This indicates that using the MV ROC-derived threshold to evaluate EEG spectral variables provides information comparable to that produced using more advanced machine learning methods. Undertaking an EEG does not require patient co-operation and the recordings are not subject to learning effects—both problems which beset many other assessment tools used to diagnose hepatic encephalopathy.(Guerit et al., 2009)
This was recognised in the recent Practice
Guideline on Hepatic Encephalopathy in Chronic Liver Disease published jointly by the American and European Associations for the Study of the Liver.(Vilstrup et al., 2014) However, the Practice Guideline also pointed out that the EEG findings in patients with cirrhosis are non-specific and states that ‘possibly the reliability of the EEG analysis [can] increase with quantitative analysis’. The results of this study may help to address this unmet need.
21
5. Conclusion Use of the new MV ROC-derived spectral thresholds considerably improves the performance of the EEG for the diagnosis of any degree of hepatic encephalopathy. A threshold for relative theta power of 22.7% on the P3-P4 lead derivation will identify patients with any degree of hepatic encephalopathy with a sensitivity of 74.5% and a specificity of 79.4%. Patients with cirrhosis with increased relative theta activity should then be clinically reviewed and undergo psychometric testing to confirm and classify their degree of neuropsychiatric disturbance. The relative theta power can also be used with utility to monitor patients’ progress over time and in response to treatment. These thresholds can be adopted easily without the need to alter data recording or initial processing. Conflicts of interest None of the authors have potential conflicts of interest to be disclosed. Financial support None.
22
References Amodio P, Marchetti P, Del Piccolo F, de Tourtchaninoff M, Varghese P, Zuliani C et al. Spectral versus visual EEG analysis in mild hepatic encephalopathy. Clin Neurophysiol 1999;110:1334-44. Amodio P, Quero JC, Del PF, Gatta A, Schalm SW. Diagnostic tools for the detection of subclinical
hepatic
encephalopathy:
comparison
of
standard
and
computerized
psychometric tests with spectral-EEG. Metab Brain Dis 1996;11:315-27. Amodio P, Pellegrini A, Ubiali E, Mathy I, Del Piccolo F, Orsato R, et al. The EEG assessment of low-grade hepatic encephalopathy: comparison of an artificial neural network-expert system (ANNES) based evaluation with visual EEG readings and EEG spectral analysis. Clin Neurophysiol 2006;117:2243–51. Auffarth B, López M, Cerquides J. Comparison of redundancy and relevance measures for feature selection in tissue classification of CT images. In: Perner P editor. Advances in data mining: application and theoretical aspects. Heidelberg: Springer-Verlag Berlin; 2010. p. 248–62 Bajaj JS, Saeian K, Schubert CM, Hafeezullah M, Franco J, Varma RR, et al. Minimal hepatic encephalopathy is associated with motor vehicle crashes: the reality beyond the driving test. Hepatology 2009;50:1175-83. Bauer G, Bauer R. EEG, drug effects and central nervous system poisoning. In Niedermeyer E, Lopes da Silva, editors. Electroencephalography, basic principles, clinical applications, and related fields. 4th ed. Baltimore: Wilkins & Williams; 1998. p. 671-91. Chang, C-C, Lin C-J. LIBSVM: A library for support vector machines. ACM Trans Intell Syst Technol 2011;2:1–27.
23
Conn HO, Leevy CM, Vlahcevic ZR, Rodgers JB, Maddrey WC, Seeff, L et al. Comparison of lactulose and neomycin in the treatment of chronic portal-systemic encephalopathy. A double blind controlled trial. Gastroenterology 1977;72:573-83. D'Amico G, Garcia-Tsao G, Pagliaro L. Natural history and prognostic indicators of survival in cirrhosis: a systematic review of 118 studies. J Hepatol 2006;44:217-31. Dhiman RK, Saraswat VA, Sharma BK, Sarin SK, Chawla YK, Butterworth R et al. Minimal hepatic encephalopathy: consensus statement of a working party of the Indian National Association for Study of the Liver. J Gastroenterol Hepatol 2010;25:1029-41. Ferenci P, Lockwood A, Mullen K, Tarter, R, Weissenborn K, Blei AT.
Hepatic
encephalopathy--definition, nomenclature, diagnosis, and quantification: final report of the working party at the 11th World Congresses of Gastroenterology, Vienna, 1998. Hepatology 2002;35:716-21. Foley JM, Watson CW, Adams RD. Significance of the electroencephalographic changes in hepatic coma. Trans Am Neurol Assoc 1950;51:161–5. Gong Q, Wu Q, Scarpazza C, Lui S, Jia Z, Marquand A. Prognostic prediction of therapeutic response in depression using high-field MR imaging. Neuroimage 2011;55:1497-503. Groeneweg M, Quero JC, De BI, Hartmann IJ, Essink-bot ML, Hop WC, Schalm SW. Subclinical hepatic encephalopathy impairs daily functioning. Hepatology1998;28:45-9. Guerit JM, Amantini A, Fischer C, Kaplan PW, Mecarelli O, Schnitzler A et al. Neurophysiological investigations of hepatic encephalopathy: ISHEN practice guidelines. Liver Int 2009;29:789-96. Jackson C, Sherratt M.
A novel spatio-temporal decomposition of the EEG: derivation,
validation and clinical application. Clin Neurophysiol 2004;115:227-37.
24
Laidlaw J. The application in general medical conditions of a visual method of assessing and representing generalized electroencephalographic abnormalities.
J Neurol Neurosurg
Psychiatry 1959;22:69-76. Marks ME, Jackson CD, Montagnese S, Jenkins CW, Head IM, Morris RW, et al. Derivation of a normative UK database for the psychometric hepatic encephalopathy score (PHES): confounding effect of ethnicity and test scoring. J Hepatol 2008;48(Suppl 2):S119. Montagnese S, Amodio P, Morgan MY. Methods for diagnosing hepatic encephalopathy in patients with cirrhosis: a multidimensional approach. Metab Brain Dis 2004;19:281-312. Montagnese S, Middleton B, Skene DJ, Morgan MY. Nighttime sleep disturbance does not correlate with neuropsychiatric impairment in patients with cirrhosis.
Liver Int
2009;29:1372-82. Montagnese S, Jackson C, Morgan MY.
Spatio-temporal decomposition of the
electroencephalogram in patients with cirrhosis. J Hepatol 2007;46:447-58. Niedermeyer E. The normal EEG of the waking adult. In Niedermeyer E, Lopes da Silva, editors. Electroencephalography, basic principles, clinical applications, and related fields. 4th ed. Baltimore: Wilkins & Williams; 1998a. p. 149-73. Niedermeyer E. Metabolic central nervous system disorders. In Niedermeyer E, Lopes da Silva, editors. Electroencephalography, basic principles, clinical applications, and related fields. 4th ed. Baltimore: Wilkins & Williams; 1998b. p. 416-31. Ortiz-Rosario A, Adeli H. Brain-computer interface technologies: from signal to action. Rev Neurosci 2013;24:537-52. Parsons-Smith BG, Summerskill WH, Dawson AM, Sherlock S. The electroencephalograph in liver disease. Lancet 1957;273:867-71.
25
Pugh RN, Murray-Lyon IM, Dawson JL, Pietroni MC, Williams R.
Transection of the
oesophagus for bleeding oesophageal varices. Br J Surg 1973;60:646-9. Rehnstrom S, Simert G, Hansson JA, Johnson G, Vang J. Chronic hepatic encephalopathy. A psychometrical study. Scand J Gastroenterol 1977;12:305-11. Román E, Córdoba J, Torrens M, Torras X, Villanueva C, Vargas V et al. Minimal hepatic encephalopathy is associated with falls. Am J Gastroenterol 2011;106:476-82. Schomerus H, Hamster W, Blunck H, Reinhard U, Mayer K, Dolle W. Latent portasystemic encephalopathy. I. Nature of cerebral functional defects and their effect on fitness to drive. Dig Dis Sci 1981;26:622-30. Sotil EU, Gottstein J, Ayala E, Randolph C, Blei AT. Impact of preoperative overt hepatic encephalopathy on neurocognitive function after liver transplantation.
Liver Transpl
2009;15:184-92. Stewart CA, Malinchoc M, Kim WR, Kamath PS. Hepatic encephalopathy as a predictor of survival in patients with end-stage liver disease. Liver Transpl 2007;13:1366-71. Van der Rijt CC, Schalm SW, De Groot GH, De VM. Objective measurement of hepatic encephalopathy by means of automated EEG analysis.
Electroencephalogr Clin
Neurophysiol 1984;57:423-6. Vilstrup H, Amodio P, Bajaj J, Cordoba J, Ferenci P, Mullen KD, Weissenborn K, Wong P. Hepatic encephalopathy in chronic liver disease: 2014 Practice Guideline by the American Association for the Study of Liver Diseases and the European Association for the Study of the Liver. Hepatology 2014;60:715-35.
26
Weissenborn K, Scholz M, Hinrichs H, Wiltfang J, Schmidt FW, Kunkel H. Neurophysiological assessment of early hepatic encephalopathy. Electroencephalogr Clin Neurophysiol 1990;75:289-95. Weissenborn K, Ennen JC, Schomerus H, Ruckert N, Hecker H.
Neuropsychological
characterization of hepatic encephalopathy. J Hepatol 2001;34:768-73. Wong RJ, Aguilar M, Gish RG, Cheung R, Ahmed A. The impact of pretransplant hepatic encephalopathy on survival following liver transplantation. Liver Transpl 2015;21:873-80.
27
Figure Legend Figure 1. Representative EEG recordings with corresponding spectral variables obtained from the P3-P4 derivation in a healthy control subject and three patients with cirrhosis by degree of neuropsychiatric impairment. The positional and numerical designation of the 21 electrode is marked to the left of their corresponding traces.
Figure 1
A
B
C
Healthy Control Unimpaired MDF (Hz) Delta (%) Theta (%) Alpha (%) Beta (%)
11.0 6.0 7.1 67.1 19.8
10.4 6.0 10.1 65.7 18.2
D
Patients with cirrhosis Minimal HE 7.4 23.0 39.3 29.0 8.7
Overt HE 4.3 56.0 34.5 6.6 3.0
Tables 1-5
1
Table 1. Demographic and assessment variables, in the patient and the reference populations, by neuropsychiatric status. Study population
Age
Gender
Pugh's score
NCT-A
NCT-B
SD
LTT-t
LTT-e
(5-15)
DS* (n correct)
(n)
(yr)
(% men)
Reference population (137)
39.2 (18.0-75.0)
53
Patient with cirrhosis (226)
54.8 (26.0-80.0)
Unimpaired (127)
54.1^^^ (32.0-80.0)
Minimal HE (21)
56.1^^^ (42.0-71.0)
Overt HE (78)
55.5^^^ (26.0-78.0)
PHES
(s)
(s)
(s)
(s)
(n errors)
66
6.8 (5.0-13.0)
34.5 (0.0-80.0)
56.1 (15.7-480.0)
151.4 (21.9-480.0)
70.6 (29.0-332.5)
102.4 (25.0-430.8)
64.6 (0.0-389.0)
-1.5 (-6.7:1.6)
65
5.7 (5.0-12.0)
42.2 (21.0-80.0)
37.1 (15.7-86.6)
98.7 (21.9-400.0)
54.6 (29.0-111.0)
82.2 (25.0-212.9)
48.9 (2.0-197.0)
-0.6 (-2.0:1.6)
67
5.7 (5.0-10.0)
25.0*** (0.0-40.0)
67.4*** (36.3-148.0)
213.2*** (73.0-480.0)
77.3*** (48.1-138.0)
117.4*** (67.0-230.5)
71.8 (5.0-204.0)
-2.6*** (-3.8:-2.0)
68
9.0***### (6.0-13.0)
24.4*** (3.0-45.0)
83.9*** (24.0-480.0)
220.7*** (54.8-480.0)
94.8*** (35.0-332.5)
131.4*** (39.0-430.8)
88.2*** (0.0-389.0)
-2.8*** (-6.7:-0.4)
Data are crude – not logarithmically transformed – and are expressed as mean (range) or absolute number (%); *normally distributed variable Abbreviations: DS: Digit Symbol test; NCT-A/B: Number Connection Tests A/B; SD: Serial Dotting test; LTT-t/e: Line Tracing Test time/errors; PHES: Psychometric Hepatic Encephalopathy Score; HE: hepatic encephalopathy Differences between the reference population and the various patient subgroups; ^ p<0.05; ^^ p<0.01; ^^^ p<0.001 Differences between the unimpaired patients and the minimal/overt HE; * p < 0.05; ** p<0.01; *** p<0.001 Differences between the patients with minimal and overt HE; # p < 0.05; ## p<0.01; ### p<0.001
2
Table 2. EEG spectral indices in the patient and reference populations, by neuropsychiatric status. Study population (n)
MDF (Hz)
Delta (%)
Theta (%)
Alpha (%)
Beta (%)
P3-P4 spectral indices Reference population (137)
9.9 ± 1.5 (7.0-15.5)
17.3 ± 9.2 (4.4-49.8)
16.2 ± 7.4 (4.9-53.9)
45.6 ± 16.7 (10.8-78.9)
20.9 ± 11.8 (5.3-68.1)
Patients with cirrhosis (226)
9.4 ± 2.0 (2.4-13.5)
16.7 ± 12.1 (2.6-89.1)
25.5 ± 15.6^^^ (5.5-81.3)
37.4 ± 17.5^^^ (1.2-79.1)
20.4 ± 12.0 (1.1-56.5)
Unimpaired (127)
10.2 ± 1.4 (6.5-13.2)
14.0 ± 7.5^ (3.1-40.6)
18.9 ± 11.6 (6.3-62.3)
43.4 ± 16.1 (10.6-79.1)
23.8 ± 11.9 (5.9-56.1)
Minimal HE (21)
9.9 ± 1.5 (7.4-13.5)
14.9 ± 7.2 (2.6-33.8)
21.6 ± 11.1 (5.5-48.4)
40.8 ± 16.2 (17.7-77.1)
22.7 ± 12.2 (8.4-56.5)
Overt HE (78)
8.0 ± 2.2^^^***## (2.4-12.2)
21.6 ± 16.9* (4.0-89.1)
37.2 ± 15.4^^^***### (8.2-81.3)
26.8 ± 15.3^^^***## (1.2-70.9)
14.3 ± 9.8^^^***## (1.1-42.0)
SEDACA spectral indices Reference population (129)
9.5 ± 1.2 (6.4-14.3)
13.4 ± 8.7 (2.3-42.9)
13.7 ± 10.5 (2.0-79.5)
60.6 ± 19.9 (8.6-90.3)
12.3 ± 10.5 (2.3-63.0)
Patients with cirrhosis (218)
8.7 ± 1.8^^^ (3.3-14.4)
17.0 ± 14.0 (2.2-75.2)
28.7 ± 20.1^^^ (2.2-79.7)
40.9 ± 24.4^^^ (4.9-88.2)
13.4 ± 10.4 (2.3-58.1)
Unimpaired (122)
9.5 ± 1.4 (5.9-14.4)
13.8 ± 10.2 (2.2-57.9)
19.8 ± 15.4^^ (2.2-68.5)
50.6 ± 22.6^^ (9.4-88.2)
15.7 ± 11.7^ (2.4-58.1)
Minimal HE (21)
8.6 ± 1.5^^* (4.4-11.8)
18.8 ± 16.6 (3.6-75.2)
26.6 ± 15.6^^^ (7.8-61.6)
41.2 ± 22.5^^ (5.1-81.8)
13.4 ± 10.4 (4.4-43.2)
Overt HE (75)
7.4 ± 1.7^^^***# (3.3-11.8)
21.7 ± 17.1^^** (5.2-75.1)
43.9 ± 19.3^^^***## (8.4-79.7)
24.8 ± 19.1^^^***## (4.9-74.5)
9.6 ± 6.4*** (2.3-38.6)
Data are crude – not logarithmically transformed – and are expressed as means ± 1 SD (range) Abbreviations: MDF: Mean Dominant Frequency; Delta %: relative delta power; Theta %: relative theta power; Alpha %: relative alpha power; Beta %: relative beta power; HE: hepatic encephalopathy Difference between the reference population and the patient population and subgroups: ^ p < 0.05; ^^ p < 0.01; ^^^ p < 0.001 Differences between unimpaired patients and those with minimal/overt HE: * p < 0.05; ** p < 0.01; *** p < 0.001 Differences between the patient with minimal and overt HE: # p < 0.05; ## p < 0.01; ### p < 0.001
3
Table 3. Conventional and new thresholds on the spectral variables derived in the identification cohort for the diagnosis of any degree of hepatic encephalopathy. P3-P4 lead derivation MDF (Hz)
Delta (%)
Theta (%)
SEDACA components MDF (Hz)
Delta (%)
Theta (%)
43.5
40.0
13.6
21.9
62.5
21.9
Conventional thresholds1,2 6.8
49.0
35.0
6.8
ROC-derived thresholds 9.2
16.2
22.7
8.8
MV ROC-derived thresholds 7.0
50.0
22.7
5.7
MDF: Mean Dominant Frequency; Theta%: relative theta power; Delta%: relative delta power; ROC: receiver operating characteristics; MV ROC: multivariable ROC 1 Conventional thresholds on P3-P4 (Amodio et al. 1999) 2 Conventional thresholds on selected SEDACA components (Montagnese et al. 2007)
4
Table 4. Performance characteristics of spectral variables in the identification cohort using both conventional and newly derived thresholds. P3-P4 lead derivation Spectral variable
Sensitivity (%)
Specificity (%)
PPV (%)
SEDACA components NPV (%)
Sensitivity (%)
Specificity (%)
PPV (%)
NPV (%)
Conventional thresholds MDF (Hz)
19.6
100.0
100.0
76.2
20.8
98.4
83.3
76.5
Delta (%)
5.9
99.2
75.0
73.0
10.4
99.2
83.3
74.4
Theta (%)
41.2
95.4
77.8
80.6
56.2
95.2
81.8
85.1
Composite*
47.1
94.7
77.4
82.1
60.4
93.7
78.4
86.1
ROC derived thresholds MDF (Hz)
60.8
72.5
46.3
82.6
75.0
72.2
50.7
88.3
Delta (%)
51.0
61.1
33.8
76.2
58.3
65.9
39.4
80.6
Theta (%)
74.5
79.4
58.5
88.9
79.2
78.6
58.5
90.8
Composite*
80.4
47.3
37.3
86.1
87.5
50.0
40.0
91.3
MV ROC derived thresholds MDF (Hz)
21.6
100.0
100.0
76.6
16.7
100.0
100.0
75.9
Delta (%)
5.9
100.0
100.0
73.2
4.2
100.0
100.0
73.3
Theta (%)
74.5
79.4
58.5
88.9
79.2
78.6
58.5
90.8
Composite*
76.5
79.4
59.1
89.7
81.2
78.6
59.1
91.7
Abbreviations: PPV: Positive Predictive Value; NPV: Negative Predictive Value; MDF: Mean Dominant Frequency; Theta%: relative theta power; Delta%: relative delta power; ROC: receiver operating characteristics; MV ROC: multivariable ROC Composite *: an EEG is considered abnormal if the MDF or relative delta power or relative theta power is abnormal
5
Table 5. Performance of the composite assessment* based on thresholds derived from/trained on the identification cohort, for the diagnosis of any degree of hepatic encephalopathy, in the validation cohort. P3-P4 lead derivation
SEDACA components
Sensitivity
Specificity
PPV
NPV
Sensitivity
Specificity
PPV
NPV
(%)
(%)
(%)
(%)
(%)
(%)
(%)
(%)
91.2
73.2
86.4
46.4
38.0
89.2
81.2
72.0
52.7
90.9
77.1
76.8
56.1
89.7
Conventional thresholds 58.3
93.2
75.7
86.1
62.5
ROC derived thresholds 83.3
51.9
38.5
89.6
85.4
MV ROC derived thresholds 75.0
77.4
54.5
89.6 SVM
72.9
76.8
56.1
88.5
Composite *: an EEG is considered abnormal if the MDF or relative delta power or relative theta power is abnormal Abbreviations: PPV: Positive Predictive Value; NPV: Negative Predictive Value; ROC: receiver operating characteristics; MV-ROC: multivariable ROC; SVM: Support Vector Machine
S1388-2457(16)30018-9 http://dx.doi.org/10.1016/j.clinph.2016.03.027 CLINPH 2007799
To appear in:
Clinical Neurophysiology
Accepted Date:
14 March 2016
Please cite this article as: Jackson, C.D., Gram, M., Halliday, E., Olesen, S.S., Sandberg, T.H., Drewes, A.M., Morgan, M.Y., New Spectral Thresholds Improve the Utility of the Electroencephalogram for the Diagnosis of Hepatic Encephalopathy, Clinical Neurophysiology (2016), doi: http://dx.doi.org/10.1016/j.clinph.2016.03.027
This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
1
New Spectral Thresholds Improve the Utility of the Electroencephalogram for the Diagnosis of Hepatic Encephalopathy Clive D Jacksona,1, Mikkel Gramb,2, Edwin Hallidaya,3, Søren Schou Olesenb,4, Thomas Holm Sandbergb,5, Asbjørn Mohr Drewesb,c,6, Marsha Y Morgand,7,*
a
Department of Neurophysiology, Royal Free London NHS Foundation Trust, Pond Street, London, NW3 2QG, UK
b
Mech-Sense, Department of Gastroenterology and Hepatology, Aalborg University Hospital, Denmark
c
Department of Clinical Medicine, Aalborg University, Aalborg, Denmark
d
UCL Institute for Liver and Digestive Health, Department of Medicine, Royal Free Campus, University College London, Hampstead, London NW32PF, UK
1
[email protected] [email protected] 3 [email protected] 4 [email protected] 5 [email protected] 6 [email protected] 2
*
Corresponding author:
Dr Marsha Y Morgan, UCL Institute of Liver and Digestive Health, Division of Medicine, Royal Free Campus, University College London, Rowland Hill Street, London, NW3 2PF, UK Tel.: +44 207 433 2873 E-mail: [email protected]
2
Highlights
• New spectral EEG thresholds for the diagnosis of any degree of hepatic encephalopathy have been identified and validated using a machine learning technique. • The performance characteristics of these new thresholds are better balanced than the thresholds currently employed and hence their adoption would enhance diagnostic utility. • Implementation of these new thresholds would not require any changes in data recording or collection.
3
Abstract Objective: The utility of electroencephalogram (EEG) for the diagnosis of hepatic encephalopathy, using conventional spectral thresholds, is open to question. The aim of this study was to optimise this diagnostic performance by defining new spectral thresholds. Methods: EEGs were recorded in 69 healthy controls and 113 patients with cirrhosis whose neuropsychiatric status was classified using clinical and psychometric criteria. New EEG spectral thresholds were calculated, on the parietal P3-P4 lead derivation using an extended multivariable receiver operating characteristic curve analysis. Thresholds were validated in a separate cohort of 68 healthy controls and 113 patients with cirrhosis.
The diagnostic
performance of the newly derived spectral thresholds was further validated using a machine learning technique Results: The diagnostic performance of the new thresholds (sensitivity 75.0%; specificity 77.4%) was better balanced than conventional thresholds (58.3%; 93.2%) and comparable to the performance of a machine learning technique (72.9%; 76.8%). The diagnostic utility of the new thresholds was confirmed in the validation cohort. Conclusions: Adoption of the new spectral thresholds would significantly improve the utility of the EEG for the diagnosis of hepatic encephalopathy. Significance: These new spectral EEG thresholds optimise the performance of the EEG for the diagnosis of hepatic encephalopathy and can be adopted without the need to alter data recording or the initial processing of traces.
Keywords: Diagnostic thresholds; EEG; hepatic encephalopathy; psychometry; spectral analysis; support vector machine learning.
4
Abbreviations PHES: Psychometric Hepatic Encephalopathy Score; ROC: Receiver Operating Characteristic; MV ROC: Multivariable ROC; SVM: Support Vector Machine; SEDACA: Short Epoch, Dominant Activity, Cluster Analysis
5
1. Introduction Hepatic encephalopathy is one of the major complications of cirrhosis.
In unselected
populations approximately 30 to 45% of patients will exhibit clinically apparent neuropsychiatric abnormalities, encompassing a wide spectrum of mental and motor disorders, while a further 22 to 75%, although apparently neuropsychiatrically unimpaired on clinical examination, will show significant abnormalities in both psychometric and neurophysiological performance (Dhiman et al., 2010; Ferenci et al., 2002; Vilstrup et al., 2014). The presence of hepatic encephalopathy, whether minimal or overt, has a considerable impact on the execution of complex tasks, such as driving (Bajaj et al., 2009; Schomerus et al., 1981); health-related quality of life (Groeneweg et al., 1998; Montagnese et al., 2009); patient safety (Roman et al., 2011); neurocognitive function post liver transplantation (Sotil et al., 2009); and, ultimately, survival (D’Amico et al., 2006; Stewart et al., 2007; Wong et al., 2015). Electroencephalography (EEG) provides information useful for detecting, assessing and monitoring this complication of cirrhosis (Montagnese et al., 2004).
The main
electrophysiological characteristic of hepatic encephalopathy is slowing of the mean frequency from the alpha range towards the theta and delta ranges (Niedermeyer, 1998a). The efficacy of the EEG for the diagnosis of hepatic encephalopathy is critically dependent on the type of analysis performed, as even those incorporating semi-quantitative classifications of the background frequency (Conn et al., 1977; Laidlaw, 1959), are subject to inter- and intraoperator variability. Thus, the reported sensitivity of visual analysis of the EEG, for the diagnosis of overt hepatic encephalopathy, ranges from 57 to 100% (Rehnstrom et al., 1977; Weissenborn et al., 1990), while the specificity ranges from 41 to 88% (Parson-Smith et al., 1957; Weissenborn et al., 1990). Spectral analysis provides an automated estimate of the dominant EEG frequency (Van der Rijt et al., 1984) and, as such, is less inter- and intraoperator dependent (Amodio et al., 1996). Nevertheless, although frequency thresholds for the
6
diagnosis of overt hepatic encephalopathy have been identified (Amodio et al., 1999; Van der Rijt et al., 1984) considerable variation is still observed in the reported diagnostic performance of spectral analysis, in this context, with sensitivities ranging from 43 to 100% (Van der Rijt et al., 1984; Weissenborn et al., 1990) and specificity from 64 to 81% (Weissenborn et al., 1990). To date, no thresholds have been identified, using conventional spectral analysis on the parietal P3-P4 lead derivation, to characterize patients with minimal hepatic encephalopathy. Short Epoch, Dominant Activity, Cluster Analysis (SEDACA) is a technique for spatiotemporal decomposition of the EEG (Jackson and Sherratt, 2004; Montagnese et al., 2007). SEDACA-derived spectral estimates allow differentiation of patients with minimal hepatic encephalopathy from a reference population, whereas no such differentiation is possible using conventional spectral analysis (Montagnese et al., 2007). However, the SEDACA-derived spectral thresholds have not been independently validated, to date, and the technique is used primarily in a research setting.
Machine learning techniques explore the construction and study of algorithms to learn from and make predictions on data (Ortiz-Rosario and Adeli, 2013). Such algorithms operate by building a model from input examples to make data-driven predictions or decisions, without employing any theoretical a priori assumptions. There has been very little exploitation of these techniques to study EEG performance in patients with cirrhosis (Amodio et al., 2006).
The aims of this study were (i) to derive new EEG spectral thresholds for the diagnosis of any degree hepatic encephalopathy on both the P3-P4 lead derivation and on SEDACA components; (ii) to validate the diagnostic performance of the newly derived thresholds in an independent cohort; and (iii) to validate the performance of the newly derived thresholds with a machine learning technique.
7
2. Subjects and methods 2.1. The study populations 2.1.1. The patient cohort The patient cohort comprised of 226 patients (149 men: 77 women; mean [range] age 54.8 [2680] yr) with biopsy-proven cirrhosis recruited from the Royal Free Hospital, London between 2008 and 2012. The aetiology of their liver injury was determined using clinical, laboratory, radiological and histological variables, whilst its severity was assessed using Pugh modification of the Child’s grading system (Pugh et al., 1973). All were clinically stable at the time of the study. Patients were excluded if they were <25 or >80 years of age; if they had suffered an episode of major hepatic decompensation within seven days of assessment; had hyponatraemia or renal failure; significant cardiac or respiratory failure; insulin-dependent diabetes mellitus; cerebrovascular disease; epilepsy; a history of significant head injury or other conditions likely to affect cerebral function. Patients were also excluded if they had misused alcohol or drugs in the three months preceding assessment; if their manual dexterity was impaired; if they could not speak English; or were taking psychoactive medications. 2.1.2. The reference population The reference population of 137 healthy volunteers (73 men:64 women: mean age 39 [17-75] yr) was recruited from amongst family, friends and staff working at the Royal Free Hospital, London and individuals who had experienced an isolated episode of fainting/dizziness but in whom clinical examination, the EEG, and cerebral imaging were completely normal. None had a history of liver disease, drank alcohol in excess of 20g daily or took prescription or overthe-counter medicines. 2.1.3. Creation of the identification and validation cohorts The patients were ranked by their raw score on the Psychometric Hepatic Encephalopathy Syndrome (PHES) test battery of psychometric tests (Weissenborn et al., 2001) and
8
alternatively assigned to either an identification or a validation cohort. The 137 reference individuals were ordered by age before alternative assignment to one of the two cohorts. The terms identification and validation cohorts were then applied to the patients and reference populations within each subgroup. 2.2. Assessment and classification of neuropsychiatric status Patients were clinically assessed by two hepatologists, working independently, and their mental state classified, using the West Haven Criteria (Conn et al., 1977) as either clinically unimpaired or as showing features of overt hepatic encephalopathy.
Psychometric
performance was assessed using the PHES battery (Weissenborn et al., 2001), which comprises five paper and pencil tests viz: digit symbol, number connection A and B, serial dotting and line tracing. The PHES data were adjusted and scored using UK normative data (Marks et al., 2008); composite scores of less than two standard deviations below mean reference values were considered abnormal. Neuropsychiatric status was classified, on the day of the study, in the patient population as: (i) unimpaired: no clinical evidence of hepatic encephalopathy and no psychometric abnormalities; (ii) minimal hepatic encephalopathy: no clinical abnormalities but impaired psychometric performance; (iii) overt hepatic encephalopathy: clinically evident, characteristic neuropsychiatric disturbances together with impaired psychometric performance. 2.3. Electroencephalography Electroencephalograms (EEGs) were recorded on one of two digital EEG systems, WalterGraphtek PL-Winsor (Walter-Graphtek GmbH, Emmendingen, Gerrmany) or MicroMed SystemPlus EVOLUTION (MicroMed. S.p.A., Mogliano Veneto (TV), Italy), using 23 silversilver chloride electrodes placed according to the International 10-20 System. The impedance of the electrodes was kept below 5 KΩ. Recordings were undertaken for 10 minutes, in a state
9
of eyes-closed, relaxed, wakefulness with a sample rate of 256 Hz and an on-line band pass filter of 0.05 to 70 Hz. 2.3.1. EEG spectral analysis on the P3-P4 derivation A 60 to 100 seconds section of eye-closed, artefact-free recording was identified on the P3-P4 derivation from each EEG.
The recordings were divided into two second epochs and
multiplied with a cosine window with a one second overlap. The Fast Fourier Transform was applied to obtain the spectral power in each 0.5 Hz bin which was averaged across all epochs to give an average power spectrum for each EEG. Estimates were also obtained for five further spectral parameters viz: Mean dominant frequency (MDF): weighted mean of frequencies in the 10 to 26.5Hz range Relative delta power percentage activity in the 1.0 to 3.5 Hz range; Relative theta power: percentage activity in the 4.0 to 8.0 Hz range; Relative alpha power: percentage activity in the 8.5 to 13.0 Hz range; Relative beta power: percentage activity in the 13.5 to 26.5 Hz range. Spectral analysis was applied to the P3-P4 lead derivations and the EEGs graded by applying the thresholds proposed by Amodio et al., (1999) on the P3-P4 lead. 2.3.2. EEG spectral analysis on the SEDACA components SEDACA (Short Epoch Dominant Activity Clustering Algorithm) is a method for decomposing, or separating, the EEG recording into its separate activities, i.e. alpha rhythms or muscle or eye artefact components.(Jackson and Sherratt, 2004) Prior to SEDACA analysis of the 60 to 100 seconds of EEG recording selected for the P3-P4 spectral analysis, an extra 10 seconds of EEG recording containing eye-movement artefact was added to enhance the separation of the eye-movement components from the other activities in the recording. After decomposition of the EEG by SEDACA, a single component was selected to represent each
10
record; the extra 10 seconds removed so further analysis was over the same 60 – 100 second time periods as used in the analysis of the P3-P4 lead derivation. Spectral thresholds comparable to those used in standard P3-P4 spectral analysis were applied, as previously proposed by Montagnese et al., (2006). 2.3.3. Machine learning analysis The EEGs were exported, as ASCII files, to Mech-Sense, Department of Gastroenterology and Hepatology, Aalborg University Hospital, Denmark for analysis using Support Vector Machine (SVM) Learning which was performed on the spectral data using the libSVM toolbox (Version 3.20) (Chang and Lin, 2011) in Matlab 2012a (The Matworks, Inc., Natick, MA, USA). A linear kernel function was used to avoid over-fitting of the data (Gong et al., 2011). The cost parameter C of the SVM was set to 1 (Auffarth et al., 2010). 2.4. Derivation of new spectral EEG thresholds New spectral thresholds, for the diagnosis of any degree of hepatic encephalopathy, were derived for MDF, relative theta and relative delta power, from the P3-P4 lead derivation and SEDACA components, in the identification cohort based on classical and modified Receiver Operating Characteristic (ROC) curve analyses. Subjects were grouped as follows: Group I - the reference population and neuropsychiatrically unimpaired patients. Group II –patients and minimal or overt hepatic encephalopathy. 2.4.1. Standard ROC curve analysis The threshold for each of the spectral variables was defined by the point on the ROC curve closest to the top left hand corner of the graph. 2.4.2. Modified ROC curve analysis Sets of potential thresholds were tested, each set comprising of a threshold on MDF, relative delta power and relative theta power. The optimal set of thresholds was defined by the point
11
on this ROC-like curve closest to the top left hand corner of the graph of sensitivity against 1specificity. These are referred to as the multivariable ROC (MV ROC) thresholds. 2.5. Application of EEG spectral thresholds The conventional and newly derived spectral thresholds for the diagnosis of any degree of hepatic encephalopathy were individually applied to the P3-P4 lead derivation and the SEDACA components. An EEG was considered abnormal if the MDF were below threshold or the relative theta power or relative delta power was above threshold. 2.6. Cross-validation of the spectral thresholds A Support Machine Vector (SVM) model was used to cross validate the new spectral thresholds. Differential weighting was applied to Groups 1 and 2 in the identification cohort to account for differences in sample size. Group 1, comprising of the reference population and the neuropsychiatrically unimpaired patients, was weighted by the number of subjects in Group 2, while Group 2, comprising the patients with minimal and overt hepatic encephalopathy, as the smaller of the two groups was weighted 1.4 times the number of subjects in Group 1. One SVM model was trained, based on each group, using the MDF, relative delta power and relative theta power, and subsequently each model was tested using data from both groups. 2.7. Performance and validation of the new spectral EEG thresholds The diagnostic performance of the new spectral EEG thresholds derived in the identification cohort, using ROC and MV ROC analysis, were applied to the validation population. Performance characteristics for the diagnosis of any degree of hepatic encephalopathy are reported as sensitivity, specificity, positive predictive value and negative predictive values for MDF, relative delta power and relative theta power and for a composite assessment in which an EEG is considered abnormal if the MDF or the relative delta power or the relative theta power is abnormal.
12
2.8. Statistical analysis The distribution of variables was tested using the Shapiro-Wilk’s test. Differences between non-normally distributed variables were examined using the Kruskal-Wallis test and subsequent between-groups comparisons performed using the Mann-Whitney U test with Bonferroni corrections. Differences between normally distributed variables were examined by one way ANOVA and subsequent between group comparisons performed using the Tukey test. The statistical analyses were undertaken using the software packages STATA version 11.2 (StataCorp LP, College Station, Texas) and R version 3.0.2 (CRAN.R-project.org/doc/FAQ/RFAQ.html). 2.9. Ethics The study was conducted according to the Declaration of Helsinki (Hong Kong Amendment) and Good Clinical Practice (European guidelines). The protocol was approved by the Royal Free Hampstead NHS Trust Ethics Committee and the equivalent body in Hannover. All participating subjects provided written, informed consent.
13
3. Results 3.1 The study populations The aetiology of the cirrhosis in the patient cohort was alcohol in 157 (69.5%) of the 226; alcohol and hepatitis C in 21 (9.3%); cryptogenic in 12 (5.3%); fatty liver disease in 11 (4.9%); hepatitis C/hepatitis B in eight (3.5%); primary biliary cirrhosis in seven (3.1%); chronic active hepatitis in five (2.2%); haemochromatosis in two (0.9%); and ‘other’ in three (1.3%). Functionally: 135 (59.7%) patients were classified as Child’s Grade A; 51 (22.6%) as Child’s Grade B; and 40 (17.7%) as Child’s Grade C. On the day of study: 127 (56.2 %) patients were classified as neuropsychiatrically unimpaired; 21 (9.3%) as having minimal and 78 (34.5%) as having overt hepatic encephalopathy. The patients with overt hepatic encephalopathy had, as expected, significantly less functional hepatic reserve than the patients in the other two groups, evidenced by a higher mean Pugh’s score (Table 1). The reference subjects were significantly younger than the patients with cirrhosis (Table 1). There were no significant differences in demographic and measurement variables between the identification and validation cohorts (Supplementary Table S1). 3.2. EEG analysis EEG data were available for spectral analysis, on the P3-P4 lead derivation, from all patients and reference individuals; EEGs were available for spectral analysis on SEDACA components from 218 (96.5%) and 129 (94.2%) patients and reference individuals, respectively. 3.3. EEG spectral variables in patients, by neuropsychiatric status The mean (± 1SD) relative theta power on both the P3-P4 lead derivation and the SEDACA components was significantly increased and the relative alpha power significantly decreased in
14
the patients compared to the reference population; the mean MDF on the SEDACA component was also significantly lower (Table 2). Changes were observed in the EEG spectral variables in the patients with cirrhosis in relation to their neuropsychiatric status (Figure 1). The mean MDF in the patients with overt hepatic encephalopathy was significantly reduced, relative to the reference population and the other patient subgroups, on both the P3-P4 lead derivation and the SEDACA components, primarily reflecting increases in the mean relative theta power and reductions in the mean relative alpha power (Table 2). Significant slowing of the MDF was also observed in the patients with minimal hepatic encephalopathy compared to both the neuropsychiatrically unimpaired patients and the reference population but only on the SEDACA components (Table 2). The mean theta power showed a progressive reduction across patient subgroups, by neuropsychiatric status but did not distinguish the unimpaired patients from those with minimal hepatic encephalopathy (Table 2). The pattern of spectral EEG changes observed in the identification and validation cohorts were comparable and mirrored the changes observed in the population overall (Supplementary Tables S2a and b). 3.4. Definition of new spectral EEG thresholds The new spectral thresholds differed from the conventional thresholds currently in use (Table 3). The ROC-derived thresholds for MDF were higher, on both the P3-P4 lead derivation and the SEDACA components, while the thresholds for delta and theta activity were considerably lower. The MV ROC-derived thresholds for MDF and delta activity, on P3-P4 lead derivation, were broadly comparable to those currently in use but the threshold for theta activity was again lower.
In contrast, the MV ROC-derived thresholds for MDF and theta activity on the
15
SEDACA components were lower than conventional thresholds while the threshold for delta activity was higher (Table 3). 3.5. Diagnostic performance of conventional and new spectral EEG thresholds 3.5.1. Identification cohort The performance of the conventional spectral thresholds on the P3-P4 lead derivation for the diagnosis of any degree of hepatic was characterised by high specificity and low sensitivity (Table 4). The performance of the conventional spectral variables on the SEDACA component was generally better with a notable increase in the sensitivity of the composite assessment which, nevertheless, remained only moderate. The diagnostic performance of the new ROC-derived thresholds varied considerably by spectral variable on both the P3-P4 lead derivation and the SEDACA components; the performance of the threshold for relative theta power was balanced with high sensitivity and high specificity while the composite assessments was characterised by high sensitivity and relatively poor specificity. The thresholds derived for the individual spectral indices using the MV ROC analysis provided comparable performance data, for MDF and relative delta power, on both the P3-P4 lead derivation and the SEDACA components, to the conventional thresholds, whilst the performance characteristics for relative theta power were the same as those provided by ROC analysis. The composite assessment showed balanced performance characteristics. The spectral variable with the best performance characteristics on both the P3-P4 lead derivation and the SEDACA components, irrespective of the applied threshold, was relative theta power. 3.5.2. Validation cohort
16
The performance of the conventional, ROC-derived and MV ROC-derived spectral thresholds in the validation cohort (Table 5) closely mirrored performance in the identification cohort (Table 4). The performance characteristics of the SVM modelling were similar to those of the MV ROC analysis in the validation cohort (Tables 4 and 5).
17
4. Discussion New spectral EEG thresholds for the diagnosis of any degree of hepatic encephalopathy have been identified and validated in two independent populations and by reference to performance variables obtained using SVM. The performance characteristics of these new thresholds are superior and better balanced than the thresholds currently employed; their adoption would significantly improve the utility of the EEG for the diagnosis of hepatic encephalopathy, in clinical practice. The EEG has been used to diagnose hepatic encephalopathy since the 1950s (Foley et al., 1950). In 1957, Parsons-Smith et al. provided a detailed description and grading of the EEG findings in patients with cirrhosis which correlated fairly well with the degree of clinical neuropsychiatric disturbance. They also observed EEG changes in patients with cirrhosis who had no clinical evidence of hepatic encephalopathy which, subsequently, was recognised as one of the defining features of minimal hepatic encephalopathy. The raw EEG signal is complex comprising an admixture of different activities. Thus, visual analytical techniques, even those incorporating semi-quantitative classifications of the background frequency (Conn et al., 1977; Laidlaw, 1959), are subject to inter- and intraoperator variability(Amodio et al. 1996).
Nevertheless, visual inspection, by a trained
neurophysiologist, is still one of the most frequently utilized methods of assessment of the EEG in this patient population. Under these circumstances an EEG is considered abnormal, and compatible with a diagnosis of a metabolic or drug-induced encephalopathy, if there is slowing of the rhythmicity with a dominant frequency below 8 Hz, although other changes, such as desynchronization and anteriorisation of the dominant activity may also be seen (Parson-Smith et al., 1957). The decision as to whether these changes represent hepatic encephalopathy is a clinical one based on exclusion of drug-induced and other causes of metabolic encephalopathies such as renal failure, hyponatraemia, carbon dioxide retention and
18
chronic hypoglycaemia (Bauer and Bauer, 1998; Niedermeyer, 1998b).
In patients with
cirrhosis, in whom other causes for the EEG changes have been excluded, the presence of compatible clinical symptoms and signs will support a diagnosis of overt hepatic encephalopathy or when absent, will identify the presence of minimal hepatic encephalopathy. Spectral analysis was first used to study the EEG in patients with cirrhosis in the late 1980s/ early 1990s. It is less inter- and intra-operator dependent than visual assessment (Amodio et al., 1999), provides an automated estimate of the dominant EEG frequency and allows the relative contributions of the different rhythms to the overall background frequency to be estimated (Van der Rijt et al., 1984). The recommended spectral thresholds for MDF, and the relative powers of delta and theta on the P3-P4 lead derivation were derived essentially to characterize patients with overt hepatic encephalopathy which explains why they are set so exactingly (Van der Rijt et al., 1984). Thus, while an EEG would be considered abnormal, on visual inspection, if the dominant frequency were below 8 Hz, the conventional spectral threshold is ≤6.8 Hz. To date, no thresholds have been identified, using conventional spectral analysis, to characterize patients with minimal hepatic encephalopathy. In consequence, the spectral thresholds used to detect more substantial degrees of neuropsychiatric impairment are applied. Spatio-temporal analysis of the EEG using SEDACA provides better diagnostic information than conventional spectral analysis (Jackson and Sherratt, 2004; Montagnese et al., 2007). Thus, SEDACA-derived spectral estimates correlate better with neuropsychiatric status and allow differentiation of patients with minimal hepatic encephalopathy from a reference population (Montagnese et al., 2007). However, this is primarily a research technique and is not generally available. The overall utility of the EEG for the diagnosis of hepatic encephalopathy is difficult to gauge in the absence of a diagnostic gold standard and is very dependent on the analytical technique
19
applied.
However, in patients with overt hepatic encephalopathy the reported diagnostic
sensitivity ranges from 43 to 100% while the specificity ranges from 41 to 88%, (Parson-Smith et al., 1957; Rehnstrom et al., 1977; Van der Rijt et al., 1984; Weissenborn et al., 1990). In addition, EEG abnormalities consistent with a diagnosis of hepatic encephalopathy are observed in from 8 to 40% of neuropsychiatrically unimpaired patients with cirrhosis, (Montagnese et al., 2004). The primary aim of the present study was to optimise the performance of the EEG by determining spectral thresholds for the diagnosis of any degree of hepatic encephalopathy. This was determined to be the most pragmatic approach as it is unlikely that a robust threshold could be identified which would differentiate minimal from overt hepatic encephalopathy. Thus, if the EEG were abnormal, classification of the patients as having minimal or overt hepatic encephalopathy would be made by the clinician on the basis of the history and clinical examination. The results of this study show that the method used to determine EEG spectral thresholds has a large effect on measured performance. Use of the conventional thresholds was associated with high specificity for the diagnosis of any degree of hepatic encephalopathy but poor to moderate sensitivity. The ROC-derived threshold for relative theta power had good diagnostic utility but the composite assessment, based on the performance of the three spectral variables, was characterised by high sensitivity and low specificity. Finally, use of the MV ROC-derived thresholds was associated with superior and better balanced diagnostic performance for both the individual and composite assessments on both the P3-P4 lead derivation and the SEDACA component. There was greater concordance in the diagnostic performance of the P3-P4 lead and SEDACA spectral estimates when utilizing the ROC- or MV ROC-derived thresholds. The validity of MV ROC-derived thresholds for the diagnosis of any degree of hepatic encephalopathy was confirmed in the validation cohort and further supported by the SVM
20
classification. This indicates that using the MV ROC-derived threshold to evaluate EEG spectral variables provides information comparable to that produced using more advanced machine learning methods. Undertaking an EEG does not require patient co-operation and the recordings are not subject to learning effects—both problems which beset many other assessment tools used to diagnose hepatic encephalopathy.(Guerit et al., 2009)
This was recognised in the recent Practice
Guideline on Hepatic Encephalopathy in Chronic Liver Disease published jointly by the American and European Associations for the Study of the Liver.(Vilstrup et al., 2014) However, the Practice Guideline also pointed out that the EEG findings in patients with cirrhosis are non-specific and states that ‘possibly the reliability of the EEG analysis [can] increase with quantitative analysis’. The results of this study may help to address this unmet need.
21
5. Conclusion Use of the new MV ROC-derived spectral thresholds considerably improves the performance of the EEG for the diagnosis of any degree of hepatic encephalopathy. A threshold for relative theta power of 22.7% on the P3-P4 lead derivation will identify patients with any degree of hepatic encephalopathy with a sensitivity of 74.5% and a specificity of 79.4%. Patients with cirrhosis with increased relative theta activity should then be clinically reviewed and undergo psychometric testing to confirm and classify their degree of neuropsychiatric disturbance. The relative theta power can also be used with utility to monitor patients’ progress over time and in response to treatment. These thresholds can be adopted easily without the need to alter data recording or initial processing. Conflicts of interest None of the authors have potential conflicts of interest to be disclosed. Financial support None.
22
References Amodio P, Marchetti P, Del Piccolo F, de Tourtchaninoff M, Varghese P, Zuliani C et al. Spectral versus visual EEG analysis in mild hepatic encephalopathy. Clin Neurophysiol 1999;110:1334-44. Amodio P, Quero JC, Del PF, Gatta A, Schalm SW. Diagnostic tools for the detection of subclinical
hepatic
encephalopathy:
comparison
of
standard
and
computerized
psychometric tests with spectral-EEG. Metab Brain Dis 1996;11:315-27. Amodio P, Pellegrini A, Ubiali E, Mathy I, Del Piccolo F, Orsato R, et al. The EEG assessment of low-grade hepatic encephalopathy: comparison of an artificial neural network-expert system (ANNES) based evaluation with visual EEG readings and EEG spectral analysis. Clin Neurophysiol 2006;117:2243–51. Auffarth B, López M, Cerquides J. Comparison of redundancy and relevance measures for feature selection in tissue classification of CT images. In: Perner P editor. Advances in data mining: application and theoretical aspects. Heidelberg: Springer-Verlag Berlin; 2010. p. 248–62 Bajaj JS, Saeian K, Schubert CM, Hafeezullah M, Franco J, Varma RR, et al. Minimal hepatic encephalopathy is associated with motor vehicle crashes: the reality beyond the driving test. Hepatology 2009;50:1175-83. Bauer G, Bauer R. EEG, drug effects and central nervous system poisoning. In Niedermeyer E, Lopes da Silva, editors. Electroencephalography, basic principles, clinical applications, and related fields. 4th ed. Baltimore: Wilkins & Williams; 1998. p. 671-91. Chang, C-C, Lin C-J. LIBSVM: A library for support vector machines. ACM Trans Intell Syst Technol 2011;2:1–27.
23
Conn HO, Leevy CM, Vlahcevic ZR, Rodgers JB, Maddrey WC, Seeff, L et al. Comparison of lactulose and neomycin in the treatment of chronic portal-systemic encephalopathy. A double blind controlled trial. Gastroenterology 1977;72:573-83. D'Amico G, Garcia-Tsao G, Pagliaro L. Natural history and prognostic indicators of survival in cirrhosis: a systematic review of 118 studies. J Hepatol 2006;44:217-31. Dhiman RK, Saraswat VA, Sharma BK, Sarin SK, Chawla YK, Butterworth R et al. Minimal hepatic encephalopathy: consensus statement of a working party of the Indian National Association for Study of the Liver. J Gastroenterol Hepatol 2010;25:1029-41. Ferenci P, Lockwood A, Mullen K, Tarter, R, Weissenborn K, Blei AT.
Hepatic
encephalopathy--definition, nomenclature, diagnosis, and quantification: final report of the working party at the 11th World Congresses of Gastroenterology, Vienna, 1998. Hepatology 2002;35:716-21. Foley JM, Watson CW, Adams RD. Significance of the electroencephalographic changes in hepatic coma. Trans Am Neurol Assoc 1950;51:161–5. Gong Q, Wu Q, Scarpazza C, Lui S, Jia Z, Marquand A. Prognostic prediction of therapeutic response in depression using high-field MR imaging. Neuroimage 2011;55:1497-503. Groeneweg M, Quero JC, De BI, Hartmann IJ, Essink-bot ML, Hop WC, Schalm SW. Subclinical hepatic encephalopathy impairs daily functioning. Hepatology1998;28:45-9. Guerit JM, Amantini A, Fischer C, Kaplan PW, Mecarelli O, Schnitzler A et al. Neurophysiological investigations of hepatic encephalopathy: ISHEN practice guidelines. Liver Int 2009;29:789-96. Jackson C, Sherratt M.
A novel spatio-temporal decomposition of the EEG: derivation,
validation and clinical application. Clin Neurophysiol 2004;115:227-37.
24
Laidlaw J. The application in general medical conditions of a visual method of assessing and representing generalized electroencephalographic abnormalities.
J Neurol Neurosurg
Psychiatry 1959;22:69-76. Marks ME, Jackson CD, Montagnese S, Jenkins CW, Head IM, Morris RW, et al. Derivation of a normative UK database for the psychometric hepatic encephalopathy score (PHES): confounding effect of ethnicity and test scoring. J Hepatol 2008;48(Suppl 2):S119. Montagnese S, Amodio P, Morgan MY. Methods for diagnosing hepatic encephalopathy in patients with cirrhosis: a multidimensional approach. Metab Brain Dis 2004;19:281-312. Montagnese S, Middleton B, Skene DJ, Morgan MY. Nighttime sleep disturbance does not correlate with neuropsychiatric impairment in patients with cirrhosis.
Liver Int
2009;29:1372-82. Montagnese S, Jackson C, Morgan MY.
Spatio-temporal decomposition of the
electroencephalogram in patients with cirrhosis. J Hepatol 2007;46:447-58. Niedermeyer E. The normal EEG of the waking adult. In Niedermeyer E, Lopes da Silva, editors. Electroencephalography, basic principles, clinical applications, and related fields. 4th ed. Baltimore: Wilkins & Williams; 1998a. p. 149-73. Niedermeyer E. Metabolic central nervous system disorders. In Niedermeyer E, Lopes da Silva, editors. Electroencephalography, basic principles, clinical applications, and related fields. 4th ed. Baltimore: Wilkins & Williams; 1998b. p. 416-31. Ortiz-Rosario A, Adeli H. Brain-computer interface technologies: from signal to action. Rev Neurosci 2013;24:537-52. Parsons-Smith BG, Summerskill WH, Dawson AM, Sherlock S. The electroencephalograph in liver disease. Lancet 1957;273:867-71.
25
Pugh RN, Murray-Lyon IM, Dawson JL, Pietroni MC, Williams R.
Transection of the
oesophagus for bleeding oesophageal varices. Br J Surg 1973;60:646-9. Rehnstrom S, Simert G, Hansson JA, Johnson G, Vang J. Chronic hepatic encephalopathy. A psychometrical study. Scand J Gastroenterol 1977;12:305-11. Román E, Córdoba J, Torrens M, Torras X, Villanueva C, Vargas V et al. Minimal hepatic encephalopathy is associated with falls. Am J Gastroenterol 2011;106:476-82. Schomerus H, Hamster W, Blunck H, Reinhard U, Mayer K, Dolle W. Latent portasystemic encephalopathy. I. Nature of cerebral functional defects and their effect on fitness to drive. Dig Dis Sci 1981;26:622-30. Sotil EU, Gottstein J, Ayala E, Randolph C, Blei AT. Impact of preoperative overt hepatic encephalopathy on neurocognitive function after liver transplantation.
Liver Transpl
2009;15:184-92. Stewart CA, Malinchoc M, Kim WR, Kamath PS. Hepatic encephalopathy as a predictor of survival in patients with end-stage liver disease. Liver Transpl 2007;13:1366-71. Van der Rijt CC, Schalm SW, De Groot GH, De VM. Objective measurement of hepatic encephalopathy by means of automated EEG analysis.
Electroencephalogr Clin
Neurophysiol 1984;57:423-6. Vilstrup H, Amodio P, Bajaj J, Cordoba J, Ferenci P, Mullen KD, Weissenborn K, Wong P. Hepatic encephalopathy in chronic liver disease: 2014 Practice Guideline by the American Association for the Study of Liver Diseases and the European Association for the Study of the Liver. Hepatology 2014;60:715-35.
26
Weissenborn K, Scholz M, Hinrichs H, Wiltfang J, Schmidt FW, Kunkel H. Neurophysiological assessment of early hepatic encephalopathy. Electroencephalogr Clin Neurophysiol 1990;75:289-95. Weissenborn K, Ennen JC, Schomerus H, Ruckert N, Hecker H.
Neuropsychological
characterization of hepatic encephalopathy. J Hepatol 2001;34:768-73. Wong RJ, Aguilar M, Gish RG, Cheung R, Ahmed A. The impact of pretransplant hepatic encephalopathy on survival following liver transplantation. Liver Transpl 2015;21:873-80.
27
Figure Legend Figure 1. Representative EEG recordings with corresponding spectral variables obtained from the P3-P4 derivation in a healthy control subject and three patients with cirrhosis by degree of neuropsychiatric impairment. The positional and numerical designation of the 21 electrode is marked to the left of their corresponding traces.
Figure 1
A
B
C
Healthy Control Unimpaired MDF (Hz) Delta (%) Theta (%) Alpha (%) Beta (%)
11.0 6.0 7.1 67.1 19.8
10.4 6.0 10.1 65.7 18.2
D
Patients with cirrhosis Minimal HE 7.4 23.0 39.3 29.0 8.7
Overt HE 4.3 56.0 34.5 6.6 3.0
Tables 1-5
1
Table 1. Demographic and assessment variables, in the patient and the reference populations, by neuropsychiatric status. Study population
Age
Gender
Pugh's score
NCT-A
NCT-B
SD
LTT-t
LTT-e
(5-15)
DS* (n correct)
(n)
(yr)
(% men)
Reference population (137)
39.2 (18.0-75.0)
53
Patient with cirrhosis (226)
54.8 (26.0-80.0)
Unimpaired (127)
54.1^^^ (32.0-80.0)
Minimal HE (21)
56.1^^^ (42.0-71.0)
Overt HE (78)
55.5^^^ (26.0-78.0)
PHES
(s)
(s)
(s)
(s)
(n errors)
66
6.8 (5.0-13.0)
34.5 (0.0-80.0)
56.1 (15.7-480.0)
151.4 (21.9-480.0)
70.6 (29.0-332.5)
102.4 (25.0-430.8)
64.6 (0.0-389.0)
-1.5 (-6.7:1.6)
65
5.7 (5.0-12.0)
42.2 (21.0-80.0)
37.1 (15.7-86.6)
98.7 (21.9-400.0)
54.6 (29.0-111.0)
82.2 (25.0-212.9)
48.9 (2.0-197.0)
-0.6 (-2.0:1.6)
67
5.7 (5.0-10.0)
25.0*** (0.0-40.0)
67.4*** (36.3-148.0)
213.2*** (73.0-480.0)
77.3*** (48.1-138.0)
117.4*** (67.0-230.5)
71.8 (5.0-204.0)
-2.6*** (-3.8:-2.0)
68
9.0***### (6.0-13.0)
24.4*** (3.0-45.0)
83.9*** (24.0-480.0)
220.7*** (54.8-480.0)
94.8*** (35.0-332.5)
131.4*** (39.0-430.8)
88.2*** (0.0-389.0)
-2.8*** (-6.7:-0.4)
Data are crude – not logarithmically transformed – and are expressed as mean (range) or absolute number (%); *normally distributed variable Abbreviations: DS: Digit Symbol test; NCT-A/B: Number Connection Tests A/B; SD: Serial Dotting test; LTT-t/e: Line Tracing Test time/errors; PHES: Psychometric Hepatic Encephalopathy Score; HE: hepatic encephalopathy Differences between the reference population and the various patient subgroups; ^ p<0.05; ^^ p<0.01; ^^^ p<0.001 Differences between the unimpaired patients and the minimal/overt HE; * p < 0.05; ** p<0.01; *** p<0.001 Differences between the patients with minimal and overt HE; # p < 0.05; ## p<0.01; ### p<0.001
2
Table 2. EEG spectral indices in the patient and reference populations, by neuropsychiatric status. Study population (n)
MDF (Hz)
Delta (%)
Theta (%)
Alpha (%)
Beta (%)
P3-P4 spectral indices Reference population (137)
9.9 ± 1.5 (7.0-15.5)
17.3 ± 9.2 (4.4-49.8)
16.2 ± 7.4 (4.9-53.9)
45.6 ± 16.7 (10.8-78.9)
20.9 ± 11.8 (5.3-68.1)
Patients with cirrhosis (226)
9.4 ± 2.0 (2.4-13.5)
16.7 ± 12.1 (2.6-89.1)
25.5 ± 15.6^^^ (5.5-81.3)
37.4 ± 17.5^^^ (1.2-79.1)
20.4 ± 12.0 (1.1-56.5)
Unimpaired (127)
10.2 ± 1.4 (6.5-13.2)
14.0 ± 7.5^ (3.1-40.6)
18.9 ± 11.6 (6.3-62.3)
43.4 ± 16.1 (10.6-79.1)
23.8 ± 11.9 (5.9-56.1)
Minimal HE (21)
9.9 ± 1.5 (7.4-13.5)
14.9 ± 7.2 (2.6-33.8)
21.6 ± 11.1 (5.5-48.4)
40.8 ± 16.2 (17.7-77.1)
22.7 ± 12.2 (8.4-56.5)
Overt HE (78)
8.0 ± 2.2^^^***## (2.4-12.2)
21.6 ± 16.9* (4.0-89.1)
37.2 ± 15.4^^^***### (8.2-81.3)
26.8 ± 15.3^^^***## (1.2-70.9)
14.3 ± 9.8^^^***## (1.1-42.0)
SEDACA spectral indices Reference population (129)
9.5 ± 1.2 (6.4-14.3)
13.4 ± 8.7 (2.3-42.9)
13.7 ± 10.5 (2.0-79.5)
60.6 ± 19.9 (8.6-90.3)
12.3 ± 10.5 (2.3-63.0)
Patients with cirrhosis (218)
8.7 ± 1.8^^^ (3.3-14.4)
17.0 ± 14.0 (2.2-75.2)
28.7 ± 20.1^^^ (2.2-79.7)
40.9 ± 24.4^^^ (4.9-88.2)
13.4 ± 10.4 (2.3-58.1)
Unimpaired (122)
9.5 ± 1.4 (5.9-14.4)
13.8 ± 10.2 (2.2-57.9)
19.8 ± 15.4^^ (2.2-68.5)
50.6 ± 22.6^^ (9.4-88.2)
15.7 ± 11.7^ (2.4-58.1)
Minimal HE (21)
8.6 ± 1.5^^* (4.4-11.8)
18.8 ± 16.6 (3.6-75.2)
26.6 ± 15.6^^^ (7.8-61.6)
41.2 ± 22.5^^ (5.1-81.8)
13.4 ± 10.4 (4.4-43.2)
Overt HE (75)
7.4 ± 1.7^^^***# (3.3-11.8)
21.7 ± 17.1^^** (5.2-75.1)
43.9 ± 19.3^^^***## (8.4-79.7)
24.8 ± 19.1^^^***## (4.9-74.5)
9.6 ± 6.4*** (2.3-38.6)
Data are crude – not logarithmically transformed – and are expressed as means ± 1 SD (range) Abbreviations: MDF: Mean Dominant Frequency; Delta %: relative delta power; Theta %: relative theta power; Alpha %: relative alpha power; Beta %: relative beta power; HE: hepatic encephalopathy Difference between the reference population and the patient population and subgroups: ^ p < 0.05; ^^ p < 0.01; ^^^ p < 0.001 Differences between unimpaired patients and those with minimal/overt HE: * p < 0.05; ** p < 0.01; *** p < 0.001 Differences between the patient with minimal and overt HE: # p < 0.05; ## p < 0.01; ### p < 0.001
3
Table 3. Conventional and new thresholds on the spectral variables derived in the identification cohort for the diagnosis of any degree of hepatic encephalopathy. P3-P4 lead derivation MDF (Hz)
Delta (%)
Theta (%)
SEDACA components MDF (Hz)
Delta (%)
Theta (%)
43.5
40.0
13.6
21.9
62.5
21.9
Conventional thresholds1,2 6.8
49.0
35.0
6.8
ROC-derived thresholds 9.2
16.2
22.7
8.8
MV ROC-derived thresholds 7.0
50.0
22.7
5.7
MDF: Mean Dominant Frequency; Theta%: relative theta power; Delta%: relative delta power; ROC: receiver operating characteristics; MV ROC: multivariable ROC 1 Conventional thresholds on P3-P4 (Amodio et al. 1999) 2 Conventional thresholds on selected SEDACA components (Montagnese et al. 2007)
4
Table 4. Performance characteristics of spectral variables in the identification cohort using both conventional and newly derived thresholds. P3-P4 lead derivation Spectral variable
Sensitivity (%)
Specificity (%)
PPV (%)
SEDACA components NPV (%)
Sensitivity (%)
Specificity (%)
PPV (%)
NPV (%)
Conventional thresholds MDF (Hz)
19.6
100.0
100.0
76.2
20.8
98.4
83.3
76.5
Delta (%)
5.9
99.2
75.0
73.0
10.4
99.2
83.3
74.4
Theta (%)
41.2
95.4
77.8
80.6
56.2
95.2
81.8
85.1
Composite*
47.1
94.7
77.4
82.1
60.4
93.7
78.4
86.1
ROC derived thresholds MDF (Hz)
60.8
72.5
46.3
82.6
75.0
72.2
50.7
88.3
Delta (%)
51.0
61.1
33.8
76.2
58.3
65.9
39.4
80.6
Theta (%)
74.5
79.4
58.5
88.9
79.2
78.6
58.5
90.8
Composite*
80.4
47.3
37.3
86.1
87.5
50.0
40.0
91.3
MV ROC derived thresholds MDF (Hz)
21.6
100.0
100.0
76.6
16.7
100.0
100.0
75.9
Delta (%)
5.9
100.0
100.0
73.2
4.2
100.0
100.0
73.3
Theta (%)
74.5
79.4
58.5
88.9
79.2
78.6
58.5
90.8
Composite*
76.5
79.4
59.1
89.7
81.2
78.6
59.1
91.7
Abbreviations: PPV: Positive Predictive Value; NPV: Negative Predictive Value; MDF: Mean Dominant Frequency; Theta%: relative theta power; Delta%: relative delta power; ROC: receiver operating characteristics; MV ROC: multivariable ROC Composite *: an EEG is considered abnormal if the MDF or relative delta power or relative theta power is abnormal
5
Table 5. Performance of the composite assessment* based on thresholds derived from/trained on the identification cohort, for the diagnosis of any degree of hepatic encephalopathy, in the validation cohort. P3-P4 lead derivation
SEDACA components
Sensitivity
Specificity
PPV
NPV
Sensitivity
Specificity
PPV
NPV
(%)
(%)
(%)
(%)
(%)
(%)
(%)
(%)
91.2
73.2
86.4
46.4
38.0
89.2
81.2
72.0
52.7
90.9
77.1
76.8
56.1
89.7
Conventional thresholds 58.3
93.2
75.7
86.1
62.5
ROC derived thresholds 83.3
51.9
38.5
89.6
85.4
MV ROC derived thresholds 75.0
77.4
54.5
89.6 SVM
72.9
76.8
56.1
88.5
Composite *: an EEG is considered abnormal if the MDF or relative delta power or relative theta power is abnormal Abbreviations: PPV: Positive Predictive Value; NPV: Negative Predictive Value; ROC: receiver operating characteristics; MV-ROC: multivariable ROC; SVM: Support Vector Machine