1. Introduction
Epilepsy is a chronic neurological disorder characterized by recurrent seizures that affects approximately 1% of the world’s population [
1]. Antiepileptic drugs are ineffective in controlling seizures in up to 30% of patients to whom it is administered, who can develop intractable epilepsy with the administration of more than three antiepileptic drugs [
2].
Convulsions or loss of consciousness associated with uncontrolled seizures may cause serious injuries to the patients themselves and the people around them. Early detection of epileptic seizures prior to symptoms allows for early intervention with fast-acting medication, vagus nerve stimulation, or evasive action against accidents. A questionnaire answered by 141 epilepsy patients showed that more than 90% of the patients desired seizure prediction [
3].
Although epileptic seizure prediction using electroencephalogram (EEG) data has achieved reliable sensitivity and specificity [
4,
5,
6,
7], intolerance to motion artifacts with scalp EEG and craniotomy techniques required for invasive EEG would be a major barrier to dissemination. In addition, because the real-time analysis of the EEG channels acquired at a sampling rate higher than 500 Hz requires huge machine power for calculation, the current application software achieving standalone seizure prediction can hardly be implemented on a handheld device, such as a smartphone.
As epileptic seizures affect the autonomic nervous function, peri-ictal changes are seen in the heart rate and electrocardiogram (ECG) data [
8,
9,
10]. The parasympathetic and sympathetic outputs to the heart are modulated by the central autonomic network, which includes the insular cortex, orbitofrontal cortex, cingulate, amygdala, hypothalamus, and periaqueductal gray matter. Excessive neural activation associated with epileptic seizures distributed in these brain regions affects the functioning of the central autonomic network, which is reflected by heart rate variability. Hence, the practicality of prediction care can be improved by monitoring the cardiac autonomic changes in the early ictal phase [
11,
12,
13] and preictal phase [
14] by heart rate variability (HRV) analysis [
15,
16] with a portable Holter monitor. A systematic review of HRV and epileptic seizures was performed, which showed that the HRV and heart rate change before seizure onset [
17]. Feature extraction from HRV indices preceding seizure onset using the machine learning technique should contribute toward realizing automatic seizure prediction, whereas the Lorenz-plot-based method predicted seizures in only three of the seventeen patients [
18]. Although time-frequency analysis [
19], nonlinear analysis of HRV using fuzzy clustering [
20], and the machine learning method based on support vector machine [
21] demonstrated greater prediction sensitivity than the random predictor in retrospective analysis of preictal ECG collected from video-EEG recording, the feasibility of real-time analysis and prospective clinical applications has not been examined. Evaluation of the real-time capability through ECG data acquisition, calculation of HRV indices, and calculation of analytic features operating on a wearable system is crucial for demonstrating the feasibility of the prediction system, because seizure warnings would help avoid seizure-related accidents, improve patient quality of life, and provide novel treatment strategies, particularly for intractable epilepsy.
To improve the positive predictive value in the seizure prediction, our group reported a novel analytical method using machine learning. The method employs multivariate statistical process control (MSPC) [
22] for anomaly detection to identify minority samples that have different characteristics to the remaining samples.
In this study, the feasibility and reliability of seizure prediction as a diagnostic tool that is practical for daily use by patients is evaluated with a prototype in hospitals during long-term video-EEG monitoring, which is categorized as the phase 2 validation framework [
23].
Section 2 describes the bespoke wearable system for HRV measurement and analysis for seizure prediction. The experimental methods for evaluating the seizure prediction capability are explained in
Section 3 and the results are shown in
Section 4.
Section 5 discusses the results from instrumentation engineering and biomedical significance perspectives.
Section 6 concludes the study.
3. Experimental Methods
3.1. MSPC Model Construction
The interictal HRV data for constructing the MSPC model were collected from refractory epilepsy patients who were admitted to three departments in Japan: the Department of Neurosurgery, Medical Hospital, Tokyo Medical and Dental University (TMDU); the Department of Psychiatry, National Center Hospital of Neurology and Psychiatry (NCNP); and the Department of Epileptology, Tohoku University Hospital (TUH). Patients underwent clinical video-EEG monitoring for presurgical evaluation or seizure assessment. Data from fourteen patients collected by Fujiwara et al. (2016) were used with the approval of the Medical Research Ethics Committees from all participating departments. Two or more experts from the Japan Epilepsy Society labeled seizures and epileptic interictal spikes from video-EEG data of the awake state to allow creation of a dataset containing only normal data. The HRV indices were calculated from the cropped ECG and the learning procedure of the MSPC model was performed by MATLAB according to (Novak et al., 1999). The MSPC does not model abnormality (i.e., seizures), and collecting clean ECG data during a seizure is difficult because of the contamination of seizure motion artifacts and the low frequency of seizures.
3.2. Measurement Setup and Protocols
The prospective study was conducted on epilepsy patients admitted to TMDU, NCNP, and the Department of Psychiatry of Nara Medical Center (NMC), Japan, for clinical video-EEG monitoring. The experiments were approved by the Medical Research Ethics Committees of TMDU, NCNP, and NMC, and written informed consent was obtained from each participant.
The subjects rested in a supine or sitting position in a bed inside a monitoring room and their prescribed antiepileptic drugs were reduced from the usual dosage, according to the standard procedure of video-EEG monitoring. The video, ECG, and EEG data were recorded simultaneously using a long-term video-EEG monitoring system (Neurofax EEG-1200, NIHON KOHDEN, Japan) with sampling frequencies of 500 or 1000 Hz for both ECG and EEG.
The fabricated wearable RRI telemeter, which has disposable ECG electrodes and a bipolar CS5 lead configuration, was used simultaneously with the video-EEG monitoring system. Data were collected with an Android smartphone (Blade S Lite g02, ZTE Japan K. K., Japan) fixed to the bedside and tethered for continuous power. The internal clock of the smartphone was manually synchronized with the clock of the video-EEG monitoring system, because the network security regulations do not permit use of the network time protocol. The recording with the wearable system was continued as long as possible during the video-EEG monitoring.
The CLs were calculated during post-processing after video-EEG monitoring, because the interictal periods must be determined from the video-EEG data diagnosed by clinical specialists. Therefore, only steps 6–12 in Algorithm 1 were executed on the PC after the experiments.
The system was evaluated with video-EEG monitoring, which is considered the gold standard for commercial seizure detection devices [
31,
32,
33]. However, video-EEG monitoring restricts body motion and autonomic change from normal daily life, which may lead to the underestimation of the number of false positives. Therefore, in addition to the recommended validation methods [
34], the system was evaluated on the healthy subjects (controlled to be consistent in gender and age) and focused on false positives from the uncontrolled situation of their daily life. The healthy subjects without a history of cardiac and neuronal diseases were asked to wear the system during their daily life, except while sleeping. The same criteria for data exclusion with the patient data was used.
5. Discussion
EEG is the most popular biomarker for seizure prediction, but HRV is rarely used [
41], although the effects of epileptic seizures on the cardiac autonomic function have been reported in the early ictal phase [
11,
12,
13] and the preictal phase [
14]. Previous approaches of seizure prediction achieved comparable performance with HRV and time-frequency analysis [
19], fuzzy clustering [
20], and a support vector machine classifier [
21]. However, only retrospective analysis of video-EEG data was performed, and real-time analyses are not reported. In this study, a wearable system was used with video-EEG monitoring. The HRV indices were calculated in real time on a smartphone that contained a precalibrated MSPC model. Only the control limit, which is a discrimination threshold of seizure prediction, was calculated after measurement, since the interictal periods should be defined by specialists through a diagnosis of the video-EEG data. In clinical situations, the proposed system would be used after a definitive diagnosis has been made with video-EEG monitoring, with the control limits being calculated prior to use. In addition to prediction, the smartphone app sends data to a server, so remote monitoring in telemedicine may enable rapid fine-tuning of CLs and
τ via the network. Thus, the proposed system demonstrated adequate real-time HRV monitoring and seizure prediction.
The small size of the proposed system, consisting of a telemeter and a smartphone, allows patients to hold the system. The battery life depends on the smartphone model and the telemeter’s operating duration of more than 48 h after a full charge. Considering that a fully charged Android smartphone depletes its 3010 mAh battery in 14.12 ± 0.68 h (mean ± SD,
n = 6) under continuous operation, the proposed wearable system has sufficient battery life for daily activity monitoring from 09:00 to 18:00. Although the fabricated telemeters failed to measure the RRIs of 3.46 ± 1.91% (mean ± SD) in the interictal episodes, the failure rates are sufficiently small to be used for HRV analysis, because an RRI failure that is less than 5% does not cause a significant error in the HRV indices when they are compensated properly [
16,
42]. Meanwhile, the effect of the RRI outliers can be minimized through RRI outlier compensation (
Section 2.2). The intersubject variance in the RRI failure rates suggests that it was affected by the stability of the electrodes and patients, because the threshold-based R-wave detection method adopted in the proposed telemeter is relatively weak against the baseline drift of ECG. Since the present study showed that the original telemeter exhibited sufficient performance only in situations where patient activity was limited and the false positive rates in healthy controls were relatively higher than that in patients, device lead-off and motion artifacts in daily life activity should be discussed in the upcoming phase 3 and 4 studies. Garment-type ECG electrodes could be a solution for improving usability and reducing motion artifacts originating from multiplied wire leads [
43].
Although the numbers of seizures and patients used for evaluation are relatively small in comparison to the retrospective studies for seizure prediction [
44], the number of patients used for the learning dataset is comparable because the MSPC model was constructed from fourteen subjects in our previous study [
22]. In general, the number of events per variable is recommended to be as least ten [
45] when a statistical model is built. The proposed MSPC model has eight input HRV features that were calculated from an RRI between two heartbeats. The model was constructed with 18.9 h of interictal data that included more than ten thousand heartbeats [
22]. Thus, the sample size for model construction was appropriate for this study. Meanwhile, statistical significance was shown for the prediction sensitivity, because fourteen seizures were obtained from seven patients. A study with a larger number of subjects with controlled syndromes and focus locations may help in understanding the variance in prediction latency.
In Fujiwara et al. [
22], epileptic seizure prediction with a sensitivity of 91% and a false positive rate of 0.7 times/h was performed using an offline retrospective analysis of the ECG extracted from long-term video-EEG monitoring. In this current study, the measurement and analysis functions were implemented in the wearable system and operated in real time. The developed system realized a sensitivity of 85.7% and a false positive rate of 0.62 times/h, where the MSPC model was constructed with previously unseen data (i.e., different patients) and the
Q statistics were used independently for seizure prediction. The proposed system’s sensitivity is comparable to previous systems (78–100%) introduced in the review papers, which summarized 23 seizure prediction methods [
40,
46]; however, most of them did not consider wearability or real-time analysis capability, so are not considered suitable for practical use. Conversely, the retrospective studies of EEG-based seizure prediction achieved a false positive rate within the range of 0.06–0.39 times/h. The clinical studies of the seizure detection (not prediction) system [
33,
47] were comparable within the range of 0.04–0.16 times/h. In recent retrospective studies on seizure prediction using HRV analysis, Pavei et al. [
48] presented an epileptic seizure prediction algorithm adopting an SVM classifier for HRV signals that forecasted seizures with a sensitivity of 94.1% and a false positive rate of 0.49 in patients with epilepsy and 0.19 in healthy subjects. Billeci et al. [
21] proposed an HRV-based patient-specific seizure prediction method using recurrence quantification analysis, with an average sensitivity of 89.6 and a false positive rate of 0.41 per hour. The interview-based patient survey [
3] showed that patients require high sensitivity from seizure prediction devices, whereas they regard specificity as secondary. However, the majority of patients required false alarms to be less than the correct predictions. Therefore, the present method requires further development to reduce the false positive rate.
There was no significant reduction in the false positive rate from our previous study. The results support our previous study’s hypothesis that the false positives may be due to body motion affecting the autonomic nervous system and HRV. As the patients rested in a sitting or supine position for most of the time during the video-EEG monitoring, the probability of body motion occurrence in the present study is considered similar to the previous study. The calculation of the t-test showed that the difference in the false positive rate of Q and T2 statistics between the patient and healthy groups was not statistically significant (p > 0.05). These results suggest that the autonomic activity rates of the healthy subjects and the epilepsy patients are not significantly different during interictal periods, where there is no effect from the interictal discharges.
The false positive rate of the
Q statistic was significantly high in patient A. The data length of the extracted interictal periods of patient A was shorter than other patients because the patient had more seizures and interictal discharges than others. The insufficient length of the interictal data made the control limit unreliable. In patient C, two of the four false positives occurred while eating in the
Q statistic, while three of the four false positives in
T2 occurred while eating. Patients A and C had temporal lobe epilepsy (TLE), the severity of which may be related to abnormal heart rate regulation [
49,
50]. Considering preictal episodes A3 and C2, which were not predicted by the
Q statistic, the interictal period should be carefully selected for appropriate control limit definition. However, the extracted interictal episodes were shorter than those in previous studies [
5], because the severe exclusion criteria were defined to strictly exclude the epileptiform activity from the learning and CL tuning dataset [
22]. Further work is required to determine sufficient durations of the interictal episodes for tuning of CLs and to develop a MSPC model specific to TLE patients.
Patient F showed a significantly high false positive rate in the
T2 statistic. According to the results of the video-EEG assessment, four of thirteen false positives occurred during a cognitive function test and seven occurred when the patients were focused on their hobbies (e.g., making plastic models), which suggests that HRV changes occurred due to the mental workload, as seen from the
T2 statistic. This hypothesis coincides with our previous results, which showed that the
T2 statistic exerted a better discrimination performance than the
Q statistic in preventing drowsy driving accidents using HRV analysis and MSPC [
51].
This study has several limitations that should be focused on in future work. Only Japanese patients with limited epilepsy syndrome (focal epilepsy) were selected, which should be addressed in future work. The developed system was examined under limited conditions at the hospital, as precise evaluation of the prediction sensitivity requires the determination of accurate seizure onset and because the false positive rate requires reliable evidence of isolation from the effect of interictal discharges or sleep. Hence, we selected this limited situation that could facilitate simultaneous video-EEG monitoring, and the results were compared with the healthy controls in unlimited situations. However, evaluation of the proposed system considering practical situations in a patient’s daily life will be performed in future work to demonstrate the efficacy of a wearable seizure alert system. The ictal and interictal data while sleeping were excluded from the analysis. As suggested in the clinical trial in a seizure detection device [
33], the development of an alternative MSPC model constructed only from the interictal sleep data may have the ability to predict sleep seizures, which remains to be developed and evaluated in future work. In addition, the intervals that were difficult for labeling preictal or interictal phases due to artifact contamination in the EEG data were removed from analysis retrospectively, although this offline preprocessing cannot be achieved in real-world situations. This may lead to the underestimation of the number of false positives and should be addressed in the upcoming phase 3 trial.