EEG autoregressive modeling analysis: A diagnostic tool for patients with epilepsy without epileptiform discharges

Highlights

• Distinguishing between epileptic seizures and nonepileptic events is not always easy.

• EEG autoregressive model prediction errors were significantly higher in patients with epilepsy than controls.

• The method can assist diagnosis of epilepsy without epileptiform discharges.

Abstract

Objective

Numerous types of nonepileptic paroxysmal events, such as syncopes and psychogenic nonepileptic seizures, may imitate epileptic seizures and lead to diagnostic difficulty. Such misdiagnoses may lead to inappropriate treatment in patients that can considerably affect their lives. Electroencephalogram (EEG) is a commonly used tool in assisting diagnosis of epilepsy. Although the appearance of epileptiform discharges (EDs) in EEG recordings is specific for epilepsy diagnosis, only 25%–56% of patients with epilepsy show EDs in their first EEG examination.

Methods

In this study, we developed an autoregressive (AR) model prediction error–based EEG classification method to distinguish EEG signals between controls and patients with epilepsy without EDs. Twenty-three patients with generalized epilepsy without EDs in their EEG recordings and 23 age-matched controls were enrolled. Their EEG recordings were classified using AR model prediction error–based EEG features.

Results

Among different classification methods, XGBoost achieved the highest performance in terms of accuracy and true positive rate. The results showed that the accuracy, area under the curve, true positive rate, and true negative rate were 85.17%, 87.54%, 89.98%, and 81.81%, respectively.

Conclusions

Our proposed method can help neurologists in the early diagnosis of epilepsy in patients without EDs and might help in differentiating between nonepileptic paroxysmal events and epilepsy.

Significance

EEG AR model prediction errors could be used as an alternative diagnostic marker of epilepsy.

Introduction

Epilepsy is a common chronic neurological disorder affecting more than 50 million people worldwide (Carney et al., 2011). Its symptoms are transient and are caused by hypersynchronous activity in neuronal networks. Epilepsy is characterized by an enduring alteration in the brain that increases the predisposition toward seizures (Fisher et al., 2005). Distinguishing between epileptic seizures and nonepileptic events is not always straightforward. For example, syncopes and psychogenic nonepileptic seizures may imitate seizures and lead to diagnostic difficulty (Edfors and Erdal, 2008, Duncan, 2010). One study reported that the frequency of misdiagnosis of epilepsy ranges from 2% to 71%, with syncope and psychogenic nonepileptic seizures being the commonest imitators. Such misdiagnoses lead to mismanagement with antiepileptic drugs (AEDs), which can considerably affect patients’ lives (Xu et al., 2016).

Electroencephalogram (EEG) is a tool in assisting diagnosis of epilepsy. Although the appearance of epileptiform discharges (EDs) in EEG recordings is specific for epilepsy diagnosis, only 25%–56% of patients with epilepsy show EDs in their first EEG examination (Smith, 2005). To increase the positive EDs rate, a repeat EEG, sleep study, and hyperventilation stimulation are used for epilepsy diagnosis (Flink et al., 2002, Halasz et al., 2002). However, a substantial amount of patients with epilepsy still never show EDs in their EEG recordings.

Conventionally, epileptologists interpret EEG data through visual inspection. However, the interpretation of EEG signals through visual assessment is time consuming, particularly with the increased use of long-term or continuous video EEG recordings, in which hours or days of EEG data must be reviewed manually (Krumholz et al., 2007). In addition, the subjectivity of visual inspection results in varying clinical interpretations based on the EEG reader’s level of expertise (Acharya et al., 2018). The low positive rate of routine EEG studies poses another problem. A patient with epilepsy may undergo a routine EEG, and the results may be completely normal. This may be because the brains of patients with epilepsy generally do not continually emit EDs.

Recently, computer-aided systems and machine learning techniques have emerged as analytical methods in EEG (Giger, 2018). An autoregressive (AR) model learns from a series of timed steps and takes measurements from previous actions as inputs for a regression model to predict the value of the next time step (Jeon and Rabe-Hesketh, 2016). An AR model prediction error is defined as the difference between the predicted and exact values of the signal. It had been reported that in patients with epilepsy, the presence of spikes on the input signal increases the number of AR model prediction errors (Dandapat and Ray, 1997). The total amount of AR model prediction errors for each epoch is calculated and recorded to evaluate the density of spikes on the signal. The total number of AR model prediction errors for an EEG epoch increases dramatically during seizures. A comparison of AR model prediction errors with a threshold value is required to determine ictal state onset (Dandapat and Ray, 1997). According to aforementioned reports, the larger the total prediction error is, the weaker is the EEG stationarity. We hypothesized that AR prediction errors were greater in patients with epileptic seizures than nonepileptic seizures. In this study, we developed an AR model prediction error–based EEG classification method to distinguish between controls and patients with epilepsy without EDs.