Professional Documents
Culture Documents
a r t i c l e i n f o a b s t r a c t
Article history: The aim of this study was to validate a novel classification for the diagnosis of PNESs. Fifty-five PNES video-EEG
Received 11 November 2015 recordings were retrospectively analyzed by four epileptologists and one psychiatrist in a blind manner and clas-
Revised 19 March 2016 sified into four distinct groups: Hypermotor (H), Akinetic (A), Focal Motor (FM), and with Subjective Symptoms
Accepted 21 March 2016
(SS). Eleven signs and symptoms, which are frequently found in PNESs, were chosen for statistical validation of
Available online xxxx
our classification. An artificial neural network (ANN) analyzed PNES video recordings based on the signs and
Keywords:
symptoms mentioned above. By comparing results produced by the ANN with classifications given by examiners,
Psychogenic seizures we were able to understand whether such classification was objective and generalizable. Through accordance
Video-EEG metrics based on signs and symptoms (range: 0–100%), we found that most of the seizures belonging to class
Machine learning A showed a high degree of accordance (mean ± SD = 73% ± 5%); a similar pattern was found for class SS
Classification of PNESs (80% slightly lower accordance was reported for class H (58% ± 18%)), with a minimum of 30% in some cases.
Low agreement arose from the FM group. Seizures were univocally assigned to a given class in 83.6% of seizures.
The ANN classified PNESs in the same way as visual examination in 86.7%. Agreement between ANN classification
and visual classification reached 83.3% (SD = 17.8%) accordance for class H, 100% (SD = 22%) for class A, 83.3%
(SD = 21.2%) for class SS, and 50% (SD = 19.52%) for class FM. This is the first study in which the validity of a new
PNES classification was established and reached in two different ways. Video-EEG evaluation needs to be
performed by an experienced clinician, but later on, it may be fed into ANN analysis, whose feedback will provide
guidance for differential diagnosis. Our analysis, supported by the ML approach, showed that this model of
classification could be objectively performed by video-EEG examination.
© 2016 Elsevier Inc. All rights reserved.
http://dx.doi.org/10.1016/j.yebeh.2016.03.031
1525-5050/© 2016 Elsevier Inc. All rights reserved.
198 A. Magaudda et al. / Epilepsy & Behavior 60 (2016) 197–201
from witnesses, EEG, video-EEG monitoring, neurophysiology, neuro- comparing results produced by the ANN with classifications given by
humoral monitoring, neuroimaging, neuropsychological testing, hypno- examiners after visual inspection, we were in a position to understand
sis, and conversation analysis [17]. Differential diagnosis cannot only whether such classification was objective and generalizable.
rely on clinical features of PNESs because the majority of signs and
symptoms overlap with ES. A longer duration, a gradual onset, 2.1. Data analysis
waxing-and-waning movements, asynchronous movements, pelvic
thrusting, closed eyes, lateralized head movements, and ictal crying en- 2.1.1. Feature inspection
compass the complex clinical pattern of PNESs, yet none of these is Binary features considered were the following: Minor Limb Tremors,
disease-specific. Closed eyes and lateralized head movements have Vocalization, Eyelid Myoclonia, Rigor-like Movements, Responsiveness,
been reported to be the most specific signs of PNESs. Psychogenic Waxing-and-Waning Patterns, Pedaling, Body Rocking, Pelvic Thrust-
nonepileptic seizures display a polymorphic semiology, mimicking dif- ing, Head Lateralization Movements, and Whole Body Rigidity. First,
ferent types of ES, thus, a classification of PNESs would be useful to in- we correlated those variables by computing the phi coefficient; the sta-
crease diagnostic accuracy. tistically significant level was set to α = 0.05, corrected for multiple
Several PNES clinical classifications have been proposed [18], but comparisons using Sidak correction.
their validation and robustness have not yet been reached. The aim
of this study was to propose a novel simplified clinical classification of 2.1.2. Objectiveness of PNES visual classification
PNESs and verify its validity in terms of reproducibility, either at a We examined the generalizability of PNES classification by looking
subjective level or in an automatic analysis framework, as provided by at the degree of accordance between groupings, as performed by five
a machine learning (ML) approach. expert examiners who separately inspected videos and classified the
entire dataset.
2. Patients and methods
2.2. Machine learning classification
Fifty-five patients with PNESs diagnosed by video-EEG recordings
performed at the Adult Epilepsy Center and Unit of Infantile Neuropsy- Machine learning is a data modeling tool borrowed from scientific
chiatry of the University of Messina were included in the study. Patients computing, which is able to find complex patterns within data and dif-
or their caregivers gave written informed consent. We recorded at least ferentiate them into subclasses. To validate an ML model, datasets are
one PNES for each patient, occurring spontaneously or, in most patients, usually divided into a training set (TS) and a cross-validation set (CV).
induced by suggestion maneuvers (verbal suggestion, intermittent light The TS is used to estimate parameters of the model; the latter is then ap-
stimulation, a patch placed on the forearm, or oral administration of plied to the CV set in order to test how this model is able to correctly
a placebo). We showed PNES videos to a caregiver, to ascertain if the classify new data. Data are randomly assigned to TS and CV several
semiology of recorded PNESs was the same as those usually presented times; different models obtained from these trials are tested to find
by the patient. the one that, on average, produces the best results. For this study, at
Four epileptologists, belonging to two different tertiary care epilepsy each trial, 60% of PNES classifications were assigned to TS and the
centers, and one psychiatrist visually analyzed PNES videos in a blind remaining 40% to CV. Finally, average performances were reported. An
manner and classified recorded PNESs into four distinct groups: important aspect to be taken into account is the numerosity of each sub-
Hypermotor (H), Akinetic (A), Focal Motor (FM), and with Subjective group to be included in the analysis. The sample included in this study is
Symptoms (SS), according to a clinical classification proposed by one not perfectly balanced, i.e., there are some subclasses less represented
(AM) of the five physicians, highly experienced with PNESs. We then in- than others. This situation could present problems when assigning sam-
vestigated the concordance of the classification of different observers. ples to TS and CV: in the worst scenario, an entire subclass could be ei-
We considered as belonging to class H those seizures mainly character- ther assigned only to TS or to CV, thus making the classification task
ized by either tonic, clonic, or dystonic generalized movements. erroneous. We resolved this potential problem by performing a pseudo-
Seizures mainly characterized by the absence of movement were random subdivision, in a way to be sure, at each trial, that 60% of each
assigned to class A. Seizures with focal motor movements were instead subclass was part of TS. Among the ML techniques available, we chose
assigned to FM, whereas SS seizures were those mainly characterized by the ANN because it is considered as having the best performance, due
experiential phenomena reported by the patients. Video-EEG record- to its ability to find nonlinear hidden interactions within data [19]. We
ings showing typical patterns for each of the four PNES types are fed the ANN with data features previously reported and classification
found in the Supplementary Material (Videos 1–4). In Video 1 provided by examiners: for a given PNES, more than one class might
(Hypermotor), the attack started with a moaning, fine tremors in both be elected, thus, the classification used for ANN analysis was based on
the hands, resistance to eye opening, generalized hypertonus, and dif- a majority vote between individual subdivisions. Details of ANN models
fuse jerks. In Video 2 (Akinetic), the patient showed unresponsiveness are the following: one input layer with 11 nodes, one hidden layer with
and absence of movements. In Video 3 (Subjective Symptoms), the pa- 10 nodes, and one output layer consisting of 4 nodes; 1000 rounds of
tient referred to epigastric discomfort and that “I feel as the arms are cross-validation were used to obtain results. Analyses were performed
dry”. Finally, in Video 4 (Focal), the patient showed flapping of the using the MATLAB Software Package (www.mathworks.com).
right arm on the bed surface and wide jerks to the lower limbs. For sta-
tistical validation of our new classification, we choose eleven signs and 3. Results
symptoms more frequently found in PNESs than in ES. It is usually diffi-
cult for a single symptom to be clearly coupled to a specific PNES class; 3.1. Feature inspection
this leads to problems when attempting to find statistically significant
differences between subclasses. Hence, while a single sign may be un- The percentage of each selected feature in the different subclasses is
able to accomplish this task, an analysis that attempts to consider all summarized in Table 1. Among feature correlations, only Pedaling and
signs simultaneously as a unique multivariate type of information may Head Lateralized Movements were significantly positively correlated
instead be able to individuate hidden relations between these symp- (phi = 0.523, p b 1 ∗ 10−6), i.e., they were either both present or both
toms in a way to clearly differentiate PNES classes. An artificial neural absent. In Fig. 1, the degree of accordance between PNESs is shown,
network (ANN) analyzed PNES video recordings based on the signs based on the concomitant presence or absence of a given sign: for a
and symptoms mentioned above in this multivariate framework, giving given couple of seizures, a count of features that are in common was
as output subclasses which were automatically individuated. By performed, and the normalized sum was displayed with a colored
A. Magaudda et al. / Epilepsy & Behavior 60 (2016) 197–201 199
Fig. 1. On the left, symmetric correlation matrix between subjects that participated in the study. Blocks indicate class subdivision according to majority voting. Subjects are ordered
according to the class they have been assigned to: from top to bottom (left to right): Hypermotor, Akinetic, Subjective Symptoms, and Focal Motor. Colors represent a normalized
count of features in common between each pair of subjects. Class subdivision is highlighted by means of black blocks, showing comparisons within each class (main diagonal) and
comparisons between subjects belonging to different classes. On the right, correlation matrix scheme. (For interpretation of the references to color in this figure legend, the reader is
referred to the web version of this article.)
200 A. Magaudda et al. / Epilepsy & Behavior 60 (2016) 197–201
Fig. 2. Fitting of classification performances during training (on the left in blue), CV (in the middle in red) processes, and overall mean results (on the right in green). Fitting is computed to
the posterior probability of a sample to belong to a given class: the closer the colored line is to the black dotted one, the higher is the probability of subjects to belong to the class they were
assigned to. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
symptoms used for ANN automatic analysis, only Pedaling and Head the most common (60%), followed by the focal–clonic type (10%) and
Lateralization Movements were significantly positively related. This the akinetic type (10%) [36]. None of these classifications was validated
demonstrates that the symptoms that we chose (Minor Limb Tremors, by interobserver reliability combined with statistical analysis as in our
Vocalization, Eyelid Myoclonia, Rigor-like Movements, Responsiveness, work.
Waxing-and-Waning Patterns, Pedaling, Body Rocking, Pelvic Thrust- We proposed four classes corresponding to the most frequently
ing, Head Lateralization Movements, and Whole Body Rigidity) are rel- described ones in clinical practice. Their definition has been dictated,
atively independent: this aspect is important from a mathematical point within this context, as well as in previous articles by other groups, by
of view for the quality of input space used for the ANN technique. Fur- the need of a differential diagnosis with respect to ES. In our experience,
thermore, these symptoms are greatly significant in terms of clinical PNES subclasses in clinical practice are similar to different subtypes
classification of PNESs. Recently, ML approaches have been applied to belonging to the domain of proper epileptic seizures. That is, the H
epilepsy and ictal video decoding. A complete automatic procedure group could be placed in correspondence with the so-called convulsive
was performed based on image processing of video recordings with ES; similarly, the A group would correspond to atonic or absence ES.
the goal of training a robotic system to understand arousal of epileptic Then, FM subtypes could be considered in correspondence with epilep-
seizures [28]. Such an approach was not oriented to PNESs and could tic partial motor seizure; eventually, the SS subgroup could be placed in
not be implemented in clinical settings. Several classification schemes correspondence with epileptic partial seizure with experiential phe-
were proposed based on cluster analysis of the clinical signs of the nomena. We did not place a mixed group. Looking at previous literature,
PNESs collected by video-EEG recordings [17] or on detailed seizure ob- Groppel et al. and Hubsch et al. [28,29] did not report a mixed group. A
servation and tabulation [28–30]. Groppel et al. analyzed the clinical se- mixed PNES group (combination of features characterizing the other
miology of seizures in 27 patients with PNESs collected by video-EEG subgroups of the classification) was reported with very different
recording [28]. Cluster analysis of recurrent clinical features allowed percentages: 5.2% in Seneviratne's paper [31] and 32.9% in Dhiman's re-
them to identify three groups of PNESs: “psychogenic motor seizures” port [37]. We have classified PNESs by considering the most prominent
with predominantly hypertonic and clonic movements of the head seizure characteristics. For example, as shown in Table 1, “minor limb
and limbs, pelvic thrusting, and tonic posturing; “psychogenic minor tremors”, “rigor-like movements”, and “whole body rigidity”, which
motor or trembling seizures” with parcellar motor manifestations or are really frequent in class H, can be found, to a lower extent, in class
fine tremors; and “psychogenic atonic seizures” with a sudden loss of A as well. However, we classified as belonging to the A subtype those
postural tone and consciousness. Hubsch et al. proposed a more symp- PNESs in which the main and most evident symptom throughout a sei-
tom-specific cluster analysis for PNES classification that distinguished zure was akinesia.
PNESs into the following: dystonic attacks, paucikinetic attacks with Moreover, it is our advice that the creation of a classification that
preserved responsiveness, pseudosyncope, hyperkinetic prolonged at- would consider too many subclasses, at this stage, could cause confusion
tacks with hyperventilation and aura, and axial dystonic prolonged at- in the absence of a classification accepted by the medical community. It
tacks [29]. Seneviratne et al. reported a proposal of classification of is worth mentioning that categorization does not relate to the treatment
330 PNESs recorded from 61 patients, visually analyzed by just one ex- choice, which depends on the comorbid psychiatric disorder.
pert epileptologist [31]. In this case, PNESs were grouped into six sub- Our analysis, supported by the ML approach, showed that this model
types: rhythmic motor (with minor motor phenomena or rigor-like of PNES classification could be objectively performed by video-EEG
movements), hyperkinetic, complex motor, dialeptic, nonepileptic examination. Different examiners agreed for the most part on the subdi-
auras, and mixed. Furthermore, Selwa et al. proposed a symptom- vision into classes, and only in very few cases was there discordance in
based model of PNES classification, grouping seizures into the follow- at least one of the examiner's assignments. Usually, ML algorithms are
ing: catatonic, thrusting, automatism, tremor, intermittent, and subjec- applied to huge datasets; in our study, only 55 subjects were involved
tive, among which the catatonic subtype showed a better outcome than in the analysis. It might be argued that our optimistic results need the
the thrashing ones [32]. Further simplified models of PNES classifica- model to be applied to a larger dataset and tested again for its accuracy.
tion, based on the predominance of positive or negative motor and/or It is worth noticing that, as we already mentioned, assignment of
sensory manifestations, have been proposed [33–35]. Bhatia and Sapra each PNES to a subclass was decided via a majority vote between exam-
reported that PNESs that mimicked generalized tonic–clonic ES were iners; such decision was then fed to the ML algorithm to provide results.
A. Magaudda et al. / Epilepsy & Behavior 60 (2016) 197–201 201
Because in some cases there was discordance regarding classification of [13] Reuber M, Baker GA, Gill R, Smith DF, Chadwick DW. Failure to recognize psychogenic
nonepileptic seizures may cause death. Neurology 2004;62:834–5.
PNES subtype, a probabilistic labeling could have been used as input to [14] LaFrance Jr WC, Benbadis SR. Avoiding the costs of unrecognized psychological
the ANN. Such choice might permit taking into account uncertainty in nonepileptic seizures. Neurology 2006;66:1620–1.
category assignment. For the present work, we preferred to uniquely as- [15] Reuber M, Elger CE. Psychogenic nonepileptic seizures: review and update. Epilepsy
Behav 2003;4:205–16.
sign each PNES to a different subtype in the attempt to follow clinical [16] Reuber M. Psychogenic nonepileptic seizures: answers and questions. Epilepsy
practice, in which a unique decision has to be taken even when clini- Behav 2008;12:622–35.
cians disagree in diagnosis. Exploring the impact of such probabilistic [17] LaFrance Jr WC, Baker GA, Duncan R, Goldstein LH, Reuber M. Minimum require-
ments for the diagnosis of psychogenic nonepileptic seizures: a staged approach. A
assignment to the ANN may thus be of interest for future work.
report from the International League Against Epilepsy Nonepileptic Seizures Task
Moreover, to the best of our knowledge, this is the first time in which Force. Epilepsia 2013;54:2005–18.
such a sophisticated ML model was used to provide a blind classification [18] Cianci V, Ferlazzo E, Condino, Mauvais HS, Farnarier G, Labate A, et al. Rating scale
for psychogenic nonepileptic seizures: scale development and clinimetric testing.
of PNESs. We know that different ML models may be chosen, such as
Epilepsy Behav 2011;21:128–31.
decision trees. Future studies may be dedicated to explore results and [19] Langley P. Elements of machine learning. Morgan Kaufmann; 1996.
robustness coming from those alternative approaches. [20] Li J, Zhen X, Liu X, Ouyang G. Classifying normal and abnormal status based on video
In conclusion, a semiologically focused classification of PNESs may recordings of epileptic patients. Scientific World Journal 2014;2014:459636.
[21] Leis AA, Ross MA, Summers AK. Psychogenic seizures: ictal characteristics and
address proper diagnosis and standardization across different studies. diagnostic pitfalls. Neurology 1992;42:9.5–9.
Ultimately, etiologic understanding and management of this important [22] Dreifuss F. Classification of epileptic seizures. In: Engel JJ, Pedley TA, editors. Epilepsy: a
public health issue would improve. comprehensive textbook. Philadelphia: Lippincott-Raven; 1997. p. 525–32.
[23] Luders H, Acharya J, Banmgartner C, Benbadis S, Bleasel A, Burgess R, et al. Semiological
Supplementary data to this article can be found online at http://dx. seizure classification. Epilepsia 1998;39:1006–13.
doi.org/10.1016/j.yebeh.2016.03.031. [24] Riggio S. Psychogenic seizures. Emerg Med Clin North Am 1994;12:1001–12.
[25] Gulick TA, Spinks IP, King DW. Pseudoseizures: ictal phenomena. Neurology 1982;
32:24–30.
Conflict of interest [26] Gates JR, Ramani V, Whalen S, Loewenson R. Ictal characteristics of pseudoseizures.
Arch Neurol 1985;42:1183–7.
The authors have nothing to disclose. [27] Chee MWL, Kotagal P, van Ness PC, Gragg L, Murphy D, Luders HO. Lateralizing signs
in intractable partial epilepsy: blinded multiple-observer analysis. Neurology 1993;
43:19–25.
References [28] Groppel G, Kapitany T, Baumgartner C. Cluster analysis of clinical semiology of
psychogenic nonepileptic seizures. Epilepsia 2000;415:610–4.
[1] Lesser RP. Psychogenic seizures. Neurology 1996;46:1499–507. [29] Hubsch C, Baumann C, Hingray C, Gospodaru N, Vignal JP, Vespignani H, et al. Clinical
[2] Sigurdadottir KR, Olafsson E. Incidence of psychogenic seizures in adults: a classification of psychogenic non-epileptic seizures based on video-EEG analysis and
population-based study in Iceland. Epilepsia 1998;39:749–52. automatic clustering. J Neurol Neurosurg Psychiatry 2011;82:955–60.
[3] Sirven JI, Glosser DS. Psychogenic nonepileptic seizures: theoretic and clinical [30] Griffith NM, Szaflarski JP, Schefft BK, Isaradisaikul D, Meckler JM, McNally KA, et al. Re-
considerations. Neuropsychiatry Neuropsychol Behav Neurol 1998;11:225–35. lationship between semiology of psychogenic nonepileptic seizures and Minnesota
[4] Gumnit RJ, Walczak TS. Guidelines for essential services, personnel, and facilities in Multiphasic Personality Inventory profile. Epilepsy Behav 2007;11:105–11.
specialized epilepsy centers in the United States. National Association of Epilepsy [31] Seneviratne U, Reutens D, D'Souza W. Stereotypy of psychogenic nonepileptic
Centers. Epilepsia 2001;42:804–:814. seizures: insights from video-EEG monitoring. Epilepsia 2010;51:1159–68.
[5] Reuber M, Micoulaud-Franchi J-A, Gülich E, Bartolomei F, McGonigal A. What the [32] Selwa LM, Geyer J, Nikakhtar N, Brown MB, Schuh LA, Drury I. Non-epileptic seizure
patient's history tells us about their nonepileptic seizures. Neurophysiol Clin 2014; outcome varies by type of spell and duration of illness. Epilepsia 2000;41:1330–4.
44:375–88. [33] Meierkord H, Will B, Fish D, Shorvon S. The clinical features and prognosis of
[6] Smith PE, Myson V, Gibbon F. A teenager epilepsy clinic: observational study. Eur J pseudoseizures diagnosed using video-EEG telemetry. Neurology 1991;41:1643–6.
Neurol 2002;9:373–6. [34] Reuber M, Pukrop R, Bauer J, Helmstaedter C, Tessendorf N, Elger CE. Outcome in
[7] Pakalnis A, Paolicchi J. Frequency of secondary conversion symptoms in children psychogenic non-epileptic seizures: 1 to 10-year follow-up in 164 patients. Ann
with psychogenic nonepileptic seizures. Epilepsy Behav 2003;4:753–6. Neurol 2003;53:305–11.
[8] Kutluay E, Selwa L, Minecan D, Edwards J, Beydoun A. Nonepileptic paroxysmal [35] van Merode T, de Krom MC, Knottnerus JA. Gender-related differences in non-
events in a pediatric population. Epilepsy Behav 2010;17:272–5. epileptic attacks: a study of patient's cases in the literature. Seizure 1997;6:311–6.
[9] Asadi-Pooya AA, Emami M. Reasons for uncontrolled seizures in children: the impact [36] Bhatia MS, Sapra S. Pseudoseizures in children: a profile of 50 cases. Clin Pediatr
of pseudointractability. Epilepsy Behav 2012;25:341–: 344. (Phila) 2005;44:617–21.
[10] Szabo L, Siegler Z, Zubek L, Liptai Z, Körhegyi I, Bánsági B. A detailed semiologic [37] Dhiman V, Sinha S, Rawat VS, Harish T, Chaturvedi SK, Satishchandra P. Semiological
analysis of childhood psychogenic nonepileptic seizures. Epilepsia 2012;53:565–70. characteristics of adults with psychogenic nonepileptic seizures (PNESs): an attempt
[11] Kotagal P, Costa M, Wyllie E, Wolgamuth B. Paroxysmal nonepileptic events in towards a new classification. Epilepsy Behav 2013;27:427–32.
children and adolescents. Pediatrics 2002;11:46.
[12] Kim SH, Kim H, Lim BC, Chae JH, Kim KJ, Hwang YS, et al. Paroxysmal nonepileptic
events in pediatric patients confirmed by long-term video-EEG monitoring — single
tertiary center review of 143 patients. Epilepsy Behav 2012;24:336–40.