EEG asymmetry detection in patients with severe acquired brain injuries via machine learning methods

Highlights

• EEG-derived symmetry indexes are markers of recovery for patients with acquired brain injury.

• Machine learning algorithms can be crossvalidated to automatically extract symmetry index.

• The solution detected asymmetry with a test accuracy of 85% (sensitivity 92%, specificity 80%)

• It can speed up analysis and improve quality of care in settings lacking skilled staff.

Abstract

Lateral brain symmetry indexes, detected by electroencephalography (EEG), are markers of rehabilitative recovery widely used in patients with severe acquired brain injury (sABI). In this study, Machine Learning algorithms were cross-validated to detect consistent asymmetries, starting from a completely automated features extraction pipeline in the EEG recordings of 54 patients with sABI, classified by two expert neurophysiologists. Raw data were filtered and segmented in two-seconds non-overlapping epochs. Low data quality in frontal electrodes caused up to 40% of epochs rejection, whilst central and posterior electrodes contributed with the greatest number of artefacts-free epochs. Out of more than 3000 extracted features, ∼300 significantly differentiated symmetric and asymmetric EEG recordings, most of them extracted from pairs and lines of electrodes. The best performing solution (nested-cross-validated and optimized Support Vector Machine classifier) detected asymmetry with a test accuracy of 85% (sensitivity 92%, specificity 80%). The application of the proposed approach to our sample size supports the generalizability of our model and its translation to clinical practice. The algorithm, heading to automatic EEG analysis, has the potential to speed up analysis of long recordings and, to improve quality of care in settings lacking skilled staff.

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.