EEG asymmetry detection in patients with severe acquired brain injuries via machine learning methods
Introduction
Severe Acquired Brain Injury (sABI) is often defined as cerebral damage caused by any type of event, such as traumatic, post-anoxic, vascular, or others, that may lead to an alteration of consciousness for at least 24 h and a Glasgow Coma Scale (GCS) varying between 3 and 8 after 24 h [1]. In the last years, improvements in the treatment of patients with sABIs during intensive care have notably increased their survival chances, although they often can remain with a permanent disability. Inevitably, a conspicuous number of patients could also develop a Disorder of Consciousness (DoC) [2]. Following a coma condition, patients can persist in an Unresponsive Wakefulness Syndrome (UWS), evolve in a Minimally Conscious State (MCS), or emerge from the MCS (E-MCS) [3], [4]. Clearly, this differential assessment in patients with sABI is fundamental for the definition of a good rehabilitation plan and an accurate prognosis. These reasons led to the introduction of specific diagnostic tools for assessing consciousness level and the Coma Recovery Scale-Revised (CRS-R) has been identified as the gold standard for the clinical assessment of consciousness [5]. However, given the presence of a relevant misdiagnosis risk when the diagnosis is based only on clinical assessment, the latest international guidelines have endorsed the introduction of instrumental evaluations such as functional neuro-imaging or electroencephalography (EEG), in combination with clinical assessments [4], [6], [7], [8].
Among the various instrumental techniques, EEG has been widely introduced into clinical practice in patients with sABI during the last years [9], [10], [11], [12]. Overall, EEG is a reliable, safe, easy-to-use, and low-cost assessment to be carried out. All these aspects give evidence of the flexibility of this tool, possible to use also in countries with an underdeveloped healthcare system or home-based assessments via wearable devices. Conversely, the wide heterogeneity and complexity of such pathologies complicate the interpretation of the signals and the extraction of the clinical meaning. Furthermore, interpretation of electroencephalographic patterns in patients with sABI is affected by a conspicuous number of confounding factors related to the severe clinical condition of these patients. Specifically, unwanted motion artefacts, electromyographic (EMG) activity, and unintentional eye movements may seriously harm the interpretability and usability of electroencephalographic data. Clearly, these noise sources can also negatively affect the quantitative and automatic elaboration of signals. For this reason, robust signal pre-processing and rejection of epochs with artefacts are necessary steps to improve the algorithm performance.
In the attempt of standardizing guidelines for EEG medical reporting, several studies tried to define a consensus for EEG interpretation. The reference remains the classification according to the American Clinical Neurophysiology Society (ACNS) terminology [6], which created objective rules used to read EEG patterns.
Even if extracted via qualitative means, such Critical Care EEG terminology has been shown to decrease the inter-rater variability and to increase the precision in assessing specific clinical/consciousness conditions. Currently though, quantitative elaboration of EEG data (qEEG) paves the way for a further reduction of inter-rater and intra-rater variability of the neuro-physio-pathological misdiagnosis in patients with a sABI [8], [13].
In this context, we aim to use advanced signal analysis techniques to automate the identification of ACNS features from EEG signals acquired in clinical routine, namely in the form of brain symmetry. Indeed, lateral brain symmetry indexes have been widely used to detect specific pathological conditions [8], [14], [15], [16] derive prognostic markers of rehabilitative recovery in patients with severe neurological [13], [17] or behavioural complications [18]. Symmetry-related indexes include different other basic EEG descriptors (such as frequency and amplitude) in their core definition. In addition, symmetry is also highly related to the topography of brain damage and a proxy of its recovery course [19], [20]. Such behaviour is enhanced with hemorrhagic/vascular conditions since lesions are often more lateralized in hemorrhagic etiologies than with multifocal traumatic patients. Despite their strongly informative content, such quantitative variables result in highly complex to interpret and to relate to a clinical condition. For this reason, Computer-Aided Diagnosis (CAD) has assumed a role in interpreting complex neurophysiological patterns on EEG data [21], [22], [23], [24], [25], [26]. In particular, epileptic and focal activities has been one of the most targeted neurological disorders starting from template matching techniques [27], machine [28], [29] and deep [30] learning models based on the direct signal or on specific transforms (e.g. wavelet [31]). Also, among the aims of CAD systems, detection of depression [25], assessment of focal brain injury lesions [22] and stroke [32], [33] are often targeted. On the other hand, for what concerns automated medical reporting, deep-learning and rule-based methods targeting EEG abnormalities, presence and type of seizure (complex partial, myoclonic, tonic-clonic etc…) resulted in accuracies ranging between 80 and 95 % [34]. Biswal et al. proposed “EEGtoText”, an application capable of automatically identify epileptic discharges, sleep spindles, seizures and focal slowing activity [35]. For what concerns, asymmetric brain patterns a wide variety of symmetry indexes have been proposed both band-specific [14] or generally related to background activity [32], [36], [37], [38]. These indexes define specific brain dynamics, especially concerning sABI patients and their high variability during inpatient, but cannot be easily calculated in daily clinical practice. In particular, assessment of brain symmetry results fundamental for the diagnosis and prognosis in stroke patients [33], [39], [40], [41], [42], [43]. For what concerns the assessment of stroke, it has been shown how even a rapid assessment of brain symmetry via the 3-minutes EEG recording can predict the severity of the stroke [33]. Furthermore, a decrease of brain asymmetry was observed on repeated monthly recordings, showing how the Brain Symmetry Index is higher during the subacute stroke phase and correlates with motor function after 30–60 days after the event [41], [44]. Also, it has been shown by Fanciullacci et al. how affected and unaffected hemispheres differ in subcortical strokes for what concerns absolute delta power and how pairwise derived Brain Symmetry Index is negatively correlated with stroke severity (measured with the National Institutes of Health Stroke Scale, NIHSS) [14]. Overall, EEG estimates of brain symmetry are explicative of brain activity related to lesion location and they could allow precise definition of diagnostic-therapeutic algorithms in brain injury patients.
Overall, EEG estimates of brain symmetry are explicative of brain activity related to lesion location and they could allow precise definition of diagnostic-therapeutic algorithms in brain injury patients. For this reason, instead than computing directly such quantitative indexes of symmetry, we used raw EEG recordings to automatically detect the gold-standard symmetry classification in patients with brain injury. The work starts from the calculation of symmetry-related features with scalable topographies from daily clinical 10–20 system EEG. Then, via an initial feature screening, channels/channel groups carrying most of the significant correlations were retained. Finally, Machine Learning (ML) models were trained and optimized to predict the ACNS symmetry classification with a robust nested cross-validation approach and then embedded with explainability techniques (SHAP). This work, up to our knowledge, introduces for the first time an automated classification of brain symmetry using the ACNS classification, in cohort of patients with a severe brain injury, not strictly including only stroke patients. Furthermore, it demonstrates how, via a scaled grouping of the EEG electrodes, different symmetry levels can be detected by interpretable ML models.
Section snippets
Data and EEG collection
This observational retrospective study was approved by the local ethics committee (Comitato Area Vasta: 17505_oss). The present study followed the principles of the Declaration of Helsinki, and written consent was obtained from the legal guardians of all patients. Patients admitted to the IRU of the IRCCS Fondazione Don Carlo Gnocchi between 1/1/2020 and 1/6/2020 were enrolled in the study (Npatients = 54). A diagnosis of consciousness was made after a clinical evaluation using the CRS-R scale,
Results
Fifty-four patients with sABI (17 females, median age 65.5 years [IQR = 21]) were included in the study, the 58 % of which suffered from a DoC. Specifically, 16 patients were in a UWS/vS state and 16 in an MCS.
The majority of the patients with asymmetric EEG background activity were admitted to the IRU due to a hemorrhagic aetiology. Conversely, among patients with symmetric brain activity, aetiology were equally distributed across TBIs, vascular and hemorrhagic (Table 1).
Discussions
According to the definition of the American Academy of Neurology and the American Clinical Neurophysiology, quantitative EEG is: “the mathematical processing of digitally recorded EEG in order to highlight specific waveform components, transform the EEG into a format or domain that elucidates relevant information, or associate numerical results with the EEG data for subsequent review or comparison” [61]. Due to the complexity of the EEG signals and their processing pipelines, inferring
Conclusions
The developed algorithm, paving the way for automatic EEG medical reporting in patients with severe Acquired Brain Injuries, has the potential to speed up analysis of long recordings and, to improve quality of care in settings lacking skilled staff. The broad spectrum of etiologies, structural brain lesions and consciousness levels of our patients and the adopted nested cross-validation pipeline allow to speculate that results can be generally valid. Such approach uses data from a standard
CRediT authorship contribution statement
Leonardo Corsi: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Writing – original draft, Writing – review & editing, Visualization. Piergiuseppe Liuzzi: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Writing – original draft, Writing – review & editing, Visualization. Sara Ballanti: Writing – original draft, Writing – review & editing. Maenia Scarpino: Data curation. Antonio Maiorelli: Investigation, Data curation.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. The study was supported by the Italian Ministry of Health with the Ricerca Corrente RC2020-2022 programs and 5xMille funds AF2018: “Data Science in Rehabilitation Medicine” and AF2019: “Study and development of biomedical data science and machine learning methods to support the appropriateness and the decision-making process
References (71)
- et al.
The vegetative and minimally conscious states: diagnosis, prognosis and treatment
Neurol. Clin.
(2011) - et al.
Practice guideline update recommendations summary: disorders of consciousness: report of the guideline development, dissemination, and implementation subcommittee of the American academy of neurology; the American congress of rehabilitation medicine; and the national institute on disability, independent living, and rehabilitation research
Arch. Phys. Med. Rehabil.
(2018) - et al.
Prognostic value of post-acute EEG in severe disorders of consciousness, using American Clinical Neurophysiology Society terminology
Neurophysiol. Clin.
(2019) - et al.
Automated diagnosis of epileptic EEG using entropies
Biomed. Signal Process. Control
(2012) - et al.
Automated EEG-based screening of depression using deep convolutional neural network
Comput. Methods Programs Biomed.
(2018) - et al.
Machine-learning-based diagnostics of EEG pathology
NeuroImage
(2020) - et al.
Quantitative EEG in ischemic stroke: correlation with infarct volume and functional status in posterior circulation and lacunar syndromes
Clin. Neurophysiol.
(2011) - et al.
P-EG007. EEG power asymmetry as predictor of recovery in acute ischemic stroke patients
Clin. Neurophysiol.
(2021) - et al.
Sleep and quantitative EEG in neurodegenerative disorders
J. Psychosom. Res.
(2004) - et al.
A novel approach for computer assisted EEG monitoring in the adult ICU
Clin. Neurophysiol.
(2011)