Ictal autonomic changes as a tool for seizure detection: a systematic review

Heart rate was monitored using single or multiple lead electrocardiography (ECG) in 14 of 18 studies [8, 22‐26, 28, 32, 34‐37, 42, 43]. Alternative methods included photoplethysmography (PPG) in a wearable sensor (n = 2) [2, 30] and an implanted heart rate sensor (AspireSR) (n = 2) [31, 33].

Heart rate measurement was done using various methods of R-peak detection, including those proposed by Pan and Tompkins [30, 41], Kohler [28], Yeh and Wang [22‐24], or unspecified methods [8, 25, 26, 31‐34, 42]. Some studies applied noise filtering techniques to diminish false R-peak detection, including high- and low-pass noise filters [8, 22‐24, 26, 30] or a specific algorithm (baseline estimation and denoising with sparsity) [42].

One case study prospectively assessed a HR algorithm using a vagal nerve stimulation (VNS) device with a fixed HR sensitivity threshold [33]. Alarms were generated when the HR augmentation exceeded 50% of the baseline HR. Eleven out of 12 seizures were detected (sensitivity 92%), together with 128 false alarms (FAR 1.88/h; 68 h recordings). A second prospective validation study of the same VNS device compared different HR thresholds (≥ 20%, ≥ 40%, and ≥ 60% increases from baseline) in 16 adults with refractory epilepsy [31]. Lower thresholds resulted in higher sensitivity and higher FAR than higher thresholds (e.g., sensitivity 59.3% and FAR 7.2/h for threshold ≥ 20% vs. sensitivity 18.8% and FAR 0.5/h for thresholds ≥  60%).

Similar effects of varying the thresholds (for both the relative HR increase and the duration of HR increase) were reported in two studies on retrospectively validated HR algorithms [32, 34]. A follow-up using the same dataset examined different factors that may influence the probability of seizure detection [44]. The best regression model was created with variables including age, gender, etiology, seizure class, and years with epilepsy.

All of the HRV-focused studies performed retrospective validations [8, 22‐26, 28, 30, 41, 42]. Different HRV features were selected and specific feature thresholds were classified as ‘ictal’ or ‘interictal.’ Nine out of ten HRV studies applied linear analysis [8, 22‐25, 28, 30, 41, 42] using time domain [22‐25, 28, 30, 41, 42] and frequency domain [8, 25, 28, 41, 42] features. Time domain analysis focuses on the instantaneous HR; the interval between two normal QRS complexes, abbreviated to ‘NN.’ Different time domain features, such as the mean NN interval or the distribution of NN have been used for seizure detection. Four studies extracted and classified these time domain features using a support vector machine (SVM) classifier and validated the same HRV algorithm in different populations [22‐24, 30]. The first retrospective study of 17 people with temporal lobe epilepsy found a mean sensitivity of 83.2% with a FAR of 2.01/h [22]. The second study extracted ECG or PPG data from three different heart rate sensors worn by 11 adults with temporal lobe epilepsy [30]. The best performance was obtained using a wearable ECG device, with a sensitivity of 64% and a FAR of 2.35/h. A third study tested the algorithm in 28 children and showed a higher overall sensitivity (81.3%) and a lower FAR (0.75/h) [23]. Performance, particularly FAR, improved when applying a patient-specific heuristic classifier. The latter was confirmed in the fourth study of data from 19 people with temporal lobe epilepsy from a pre-existing epilepsy database [24]. The authors also proposed an adaptive seizure detection algorithm, and showed that similar results were obtained with simulated ‘real-time’ user feedback.

Frequency domain analysis is used to extract the frequency components of the HR signal, each with its own physiological footprint: low frequency (LF 0.04–0.15 Hz), high frequency (HF 0.15–0.40 Hz), very low frequency (VLF 0.0001–0.04 Hz), and very high frequency (VHF 0.4–0.5 Hz). Different frequencies were identified by power spectral density analysis of HRV in four studies [8, 25, 28, 41], and two studies sped up this process by applying an efficiency algorithm [fast Fourier transform (FFT)] [8, 28]. The LF/HF ratio , reflecting the balance of sympathetic and parasympathetic function, was examined in two studies [25, 41]. One of these studies tested a seizure detection algorithm combining both time and frequency domain features on 11 focal seizures upon awakening [25]. Ten of the 11 seizures were detected prior to seizure onset (sensitivity 91%, DL − 494 ± 262 s). Another study of seven adults with focal epilepsy that used time–frequency analysis of HRV based on a combination of the matching-pursuit and Wigner–Ville distribution algorithms reported a sensitivity of 96.4% with high FAR (5.4/h) [42]. Combining ECG and EEG algorithms yielded better performance (sensitivity 100%, FAR 1.6/h).

To assess the dynamic properties of ictal HR changes, nonlinear analysis can be applied, such as a Lorenz (or Poincaré) plot. This method plots the current R–R interval against the next R–R value. Standard deviations in the transverse (SD1) and longitudinal (SD2) directions of these plots can be calculated, and higher ratios of SD2/SD1 reflect increased sympathetic tone. These ratios can be used in seizure detection algorithms, since an increase in sympathetic tone is often seen during the preictal and early ictal phases. One small retrospective study proposed the modified cardio sympathetic index (mCSI) as a new measure in seizure detection that reflects the sympathetic tone [26]. A seizure detection algorithm based on changes in mCSI yielded a sensitivity of 88% in five people with temporal lobe epilepsy (FAR not reported). A larger follow-up study of adults with focal epilepsy compared frequency domain analysis with Lorenz plot analysis [8]. mCSI appeared more sensitive, but FARs were not reported.

The two remaining studies of HRV combined linear and nonlinear analysis [28, 41]. The first retrospective study of seven people with focal epilepsy reported an overall sensitivity of 88.3% with a specificity of 86.2% after selecting an optimal performance threshold for each patient [41]. The second study combined time–frequency and Lorenz plot analysis with a second nonlinear analysis of ‘sample entropy’ [28]. This parameter quantifies the regularity and complexity of a time series, and entropy decreases can be seen during the ictal phase. Applying all of these methods together to ECG data from twelve temporal lobe epilepsy patients resulted in overall sensitivity of 94.1% with a FAR of 0.49/h.

Another retrospective study reported two different seizure detection algorithms based on changes in QRS morphology (algorithm 1) and cardiorespiratory interactions (algorithm 2) [35]. The first algorithm captured five consecutive QRS complexes, aligned them with respect to the R peak, and assembled them into one QRS matrix. Principal component analysis was used to select different features from this QRS matrix. This process was repeated for every heart beat, which resulted in a sensitivity of 89.5–100% for detecting focal onset seizures and 86% for generalized onset seizures. The second algorithm was based on the well-known modulatory effects of respiration on HRV. These cardiorespiratory changes were quantified using phase-rectified signal averaging—a methodology used to detect quasi-periodicities in nonstationary signals such as the resampled RR interval time series—and were used for seizure detection. Slightly better performance was achieved by the second algorithm, which yielded a sensitivity of 100% for focal onset seizures and 90% for generalized onset seizures. In this study, 10.4–90% of the generated alarms were false, and this percentage was lower for the second algorithm.

All multimodal autonomic algorithms were retrospectively validated. A combination of three biosignals, measured by two different devices, was used for seizure detection in a study of ten subjects with focal epilepsy [2]. An algorithm based on a specific seizure pattern of increased HR, decreased SpO2, and increased EDA was able to detect all seizures in six out of ten patients with a low FAR of 0.015/h. Specific thresholds of HR, QTC, and SpO2 were combined in an algorithm tested on a larger study population of 45 people with refractory epilepsy [37]. Only half of the collected data was used for analysis, and a sensitivity of 81–94% was found for focal to bilateral tonic–clonic seizures, while focal seizures without bilateral spreading showed worse performance, with a sensitivity of 25–36%. Overall FAR ranged from 0.4–2.4/h.

Three other retrospective validation studies combined EDA and accelerometry (ACC), measured with one device [28‐40]. Different classifiers were used to select features of EDA and ACC. The first study tested two machine learning algorithms, the k-nearest neighbor (kNN) and random forest classifiers. The kNN classifier achieved the best results with 11 features, and was most sensitive for nonmotor seizures (sensitivity 97.1%, FAR not reported). The random forest classifier selected 26 features and showed its best performance with motor seizures (sensitivity 90.5%, FAR not reported). A second study used a SVM classifier to extract 19 features (16 ACC and 3 EDA) [40]. Fourteen out of 16 focal onset seizures with bilateral spreading were detected (sensitivity 88%) and FAR was 0.04/h. The same feature set was used in the third study and compared to a larger (40 ACC and 6 EDA) and a reduced (22 ACC and 3 EDA) feature set [39]. Retrospectively tested on 24 children and 45 adults with focal epilepsy, the reduced set showed the best performance (sensitivity 94.6%, FAR 0.20/day).

A multicenter study combined HR and ACC measures in 95 people with nocturnal major motor seizures [36]. Data from only 23 patients could be used to retrospectively validate three different algorithms based on changes in HR, ACC, and ‘HR or ACC.’ Clinically urgent seizures were detected well (sensitivity 71–87%), but FAR was relatively high (2.3–6.3/night), with wide variation between subjects.

According to the QUADAS-2 criteria, the overall quality of the included studies was medium–high (Table 2). Seventeen out of 21 studies were at risk of bias, mainly due to an undefined patient selection process and fitting of the algorithm [2, 8, 22‐26, 30, 32, 34, 37‐43]. There was concern regarding the applicability of the selected patients in three studies, because the populations consisted of children only and/or were not well described [23, 25, 33]. Concerns about the applicability of the index test (i.e., the tested algorithm) arose in nine studies, mainly because the algorithm was fitted to one dataset [2, 8, 23, 25, 28, 30, 32, 36, 37].

Based on the standards for the clinical validation of SDDs proposed by Beniczky and Ryvlin [21], most studies were classified as phase 1 proof-of-principle studies, whereas three were classified as phase 0 initial studies [34, 41, 42], and only one as a phase 2 study on a dedicated SDD [31] (Table 3). Seven other studies also tested a dedicated device but included small population sizes or did not address the safety of the device and were therefore classified as phase 1 [2, 30, 33, 36, 38‐40]. Ten studies trained and tested their algorithm on the same dataset [2, 8, 22, 26, 32, 34, 37, 40‐42], and only four used a predefined algorithm or cutoff values [30, 31, 33, 36]. Eighteen studies used video-EEG as reference standard; the remaining three used EEG or ECoG without video recordings [34, 41, 42].