Throughout the years, the epilepsy research community has been working in multi-centre collaborations. The joint efforts of stakeholders from different backgrounds, such as industry, healthcare and academia, as well as different countries, is essential to provide better and faster solutions to epilepsy. Some of the collaborations found were iEEG.org, SeizeIT, Epilepsy Ecosystem and EPILEPSIAE. Additionally, other publicly available datasets were found in the literature, which also comprise data from epileptic patients. The description of these datasets is summarised in Table 1 alongside the one proposed in this project. This includes the number of patients, their age, the modalities acquired, the number of seizures, the seizure types, the total duration and the setting.

The iEEG.org online platform ¹ , supported by the National Institutes of Neurological Disorders and Stroke, provides access to over 900 datasets on human and animal epilepsy, as well as analytic and visualisation tools to aid data exploration. By inserting the corresponding keywords, the datasets are filtered to a desired EEG montage, population and seizure types, just to name a few. Nonetheless, the source of the signal is always assumed to be EEG and other physiological signals are not an option when filtering.

SeizeIT ² is another collaborative project within epilepsy, which is supported by the European Institute of Innovation and Technology (EIT Health). This consortium aims to tackle the continuous ambulatory monitoring of epilepsy, in order to accurately register seizures, thereby enabling the optimisation of treatment plans.

The Epilepsy Ecosystem ³ (Kuhlmann et al., 2018) is a crowd-sourcing environment supported through the My Seizure Gauge project, supported by the Epilepsy Foundation of America, the Aikenhead Centre for Medical Discovery at St. Vincent’s Hospital Melbourne, the University of Melbourne, Monash University and Seer Medical. Its goal is to improve the performance of seizure prediction algorithms with the purpose of making them a viable option for PWE. They are responsible for gathering and maintaining two public datasets: NeuroVista Trial (Cook et al., 2013) and My Seizure Gauge Wearable ⁴ . Both datasets are available in the Seer platform requiring only a login setup. The NeuroVista trial (Cook et al., 2013) contains 15 patients that were submitted to an ambulatory intracranial EEG (iEEG) monitoring for more than 80 days. This dataset consists of a set of carefully selected patients whom experienced 2 to 12 monthly seizures, while still having a level of independence compatible with using the device for daily management. Another dataset made available in the Seer platform is the My Seizure Gauge Wearable, which is a collection of wearable data using 3 wearable devices: Empatica E4, Epilog (Frankel et al., 2021) and Byteflies (Swinnen et al., 2021). It comprises 27 patients that were monitored for approximately 4.4 ± 2.4 days. Although this dataset provides unprecedented wearable data in the context of epilepsy, it does not provide any information regarding the positioning of the devices nor some relevant clinical symptoms such as sleep/vigilance state and loss of awareness during the seizure events. Since no information was found regarding the acquisition conditions, we could not assume whether this dataset was acquired during video-EEG or in ambulatory.

The EPILEPSIAE database (Klatt et al., 2012) was created under a joint European project between Epilepsy Center, University Hospital of Freiburg; Center for Informatics and Systems, University of Coimbra; and Epilepsy Unit, CHU Pitie-Salpetriere, Paris. This database contains more than 275 patients that performed Video-EEG monitoring, focusing on the EEG signal. Still, since the ECG is also acquired by the hospital system, it is also present in this database. Several works have used this database both for EEG and ECG analysis (Truong et al., 2019; Gómez et al., 2020; Leal et al., 2021). However, EPILEPSIAE only makes available 30 patients for external researchers, which is always the same subset, without disclosing the type of seizures, or the clinical semiology. Moreover, it does not provide wearable data, only hospital data.

The Siena Scalp EEG Epilepsy Database (Detti et al., 2020) contains the data of 14 patients that performed the video-EEG monitoring in the Unit of Neurology and Neurophysiology of the University of Siena. This dataset is curated to the seizure episodes, hence for some patients, only the peri-ictal period is available to the public. The signal sources present in this database include EEG and ECG, both acquired by the hospital system.

Other epilepsy datasets also available in the literature were included in Table 1 for dimensionality comparison, however, they are out of the scope of this research, since they provide neither non-EEG data or wearable data. For more information in EEG datasets (Wong et al., 2023), provides a comprehensive review.

The majority of available datasets is focused on EEG analysis, and only a few also include the ECG signal acquired by the hospital system. Regarding the acquisition settings, four datasets contain data from long-term or video-EEG monitorings, and four datasets contain short-term EEG acquisitions. The EEG was intracranial in three datasets (indicated with iEEG), and the NeuroVista Trial is the only ambulatory dataset, to the best of our knowledge. The only multimodal wearable dataset found was My Seizure Gauge Wearable data, however it does not provide clear information on seizure types, the acquisition conditions, or the clinical manifestations of seizures. Most events annotated in this dataset have little seizure information, and for some of them it is unclear if they are even seizures or not. Larger datasets are needed to sustain the use of non-EEG data as lens for epileptic activity. Recently, projects such as My Seizure Gauge and Radar-CNS started to create multicenter datasets with hundreds of patients acquired in ambulatory scenarios (Böttcher et al., 2022). However, it was not yet proven the ability of these devices to capture pathophysiological manifestations in non-EEG modalities related to non-convulsive seizures. Hence more evidence based on data collected in a semi-controlled scenario (e.g. as the video-EEG monitoring), is essential towards the use of non-EEG physiological modalities in the prediction and forecast of epileptic seizures, and in the detection of non-convulsive epileptic seizures.

The ever-decreasing size of devices with built-in sensors, internal battery and powerful processors, allow continuous wearable long-term monitoring of movement patterns and physiological variables. A wearable device similar to a smartwatch can be decomposed in four main layers: top cover, electronics board, lithium battery and bottom cover ⁵ . The top cover serves as a protective shield for the other components; the electronics board contains sensors, communication antenna (Bluetooth or WiFi), memory, and MCU; the lithium battery provides a power supply; finally the bottom cover often contains the interface with the patient when necessary (e.g., dry electrodes). This example of a wearable device unfolds the physical constraint associated with each layer: on the one hand, multiple embedded sensors lead to more information, and larger battery sizes lead to longer continuous acquisitions; on the other hand, devices should be small and unobtrusive. Thus, there is a trade-off between the device’s size and its modus operandi.

The devices that can simultaneously record data from multiple sensors are more interesting for epilepsy monitoring since they can increase the chances of capturing manifestations from different seizure types, thus enhancing the performance of detection or early prediction algorithms (Verdru and Van Paesschen, 2020). Some devices are commercially available for the purpose of physiological signal acquisition, which could be used for research purposes. In Table 2, relevant devices found in the literature are described, which were validated and can record multimodal physiological signals. In this table, the devices are compared based on their target use, validation, modalities available, body location, battery duration, sampling frequency (FS) and data accessibility. For comparison, the bottom part of Table 2, contains the characteristics of the devices used in this project.

Table 2 illustrates the characteristics that prove to be the most relevant in wearable devices applied to epilepsy. However, it also uncovers some limitations of current options: from restrictive form factors, to short battery life, lack of validation, limited storage options, or insufficient physiological information. In fact, this last point accurately illustrates one pressing issue in epilepsy research: the extent of the applicability of seizure detection devices is inherently related to their location/signals they acquire, for example, an ACC- or EMG-based device will not be able to detect non-motor seizures. Even though the research community’s intention is in detecting all seizure types, there is an imbalance, in both research and evidence, towards an overrepresentation of convulsive motor seizures (Bruno et al., 2020). Therefore, research efforts should be endorsed for other seizure types. Moreover, additional research is needed towards the applicability of multimodal wearables in epilepsy in ambulatory scenarios, especially combined with seizure risk forecasting.

The benchmark device for epilepsy research is the Empatica E4 smartwatch, since it is discreet, it captures four biosignals and its battery lasts for 32 h. However, the data management is performed in a cloud-based platform, which compromises our privacy protection clauses. Moreover, (Vandecasteele et al., 2017), have reported the higher reliability of ECG, comparatively to PPG, to obtain the heart rate and heart rate variability, which are considered significant autonomic change predictors. These two factors motivated the development of our own devices, designed to ensure the privacy preservation and the recording of relevant physiological modalities, such as the ECG, PZT, PPG, EDA and ACC (for chest motion).

The devices developed in this scope are based on BITalino (Da Silva et al., 2014), which allowed for the experimentation around form-factors and sensors modalities. The BITalino device consists of a core block with a microprocessor (MCU) and a Bluetooth (BT) module for communication, sensors for each physiological modality, and a lithium-ion polymer (LiPo) battery. BITalino has the ability of recording simultaneously up to 6 analog channels (with a FS up to 1000Hz). When all channels are being acquired, the first four channels also allow for 10-bit resolution, while the last two only record with 6-bit resolution.

The first prototypes developed within the PreEpiSeizures project were ArmBIT and ForearmBIT (shown in Figure 1), both consisting of one BITalino core block (with MCU and BT module), and respective sensors, all enclosed within a 3D printed case with an adjustable strap. The ArmBIT, which is displayed in left side of Figure 1, collected the equivalent of lead I ECG and the triaxial ACC of the upper arm. The ForearmBIT, represented in right side of Figure 1, was design to record the bicep’s EMG, EDA on the palm of the hand, fingertip’s PPG, and forearm’s triaxial ACC. Interface with the patient was performed with the use of pre-gelled Ag/AgCl electrodes. These two devices were used simultaneously and were connected by a cable, to enable synchronisation (not shown in Figure 1). Although this montage is acceptable for short-term data acquisitions, it becomes impractical for longer sessions. This is due to the prolonged use of pre-gelled electrodes and connection cables, which can cause discomfort, associated with the fact that wearables are secured to the arm, inhibiting patients from adopting certain sleeping positions. Since these two devices are both on the arm and physically connected by a cable, the remainder of this work uses ArmBIT to refer to the ensemble with both devices.

Given the higher resolution in the first four channels, these were usually used for the more relevant physiological signals (e.g., ECG, EEG, EMG and PPG). Since the ACC possesses three axis, and the human motion has frequencies lower than 5 Hz (Korohoda et al., 2013), two of its axis were usually acquired in the channels with lower resolution.

The second iteration of the PreEpiSeizures wearable devices, consisted on rejecting gel electrodes and finding other ways to collect the same signals. The ForearmBIT was replaced by a BITalino assembled in a wristband (WristBIT), whereas the ArmBIT was replaced by a BITalino mounted in a chestband (ChestBIT). The WristBIT device (shown in Figure 2), is composed of an elastic wristband with a pocket to enclose the BITalino core block, as well as EDA and ACC sensors. The EDA signal was being recorded on the wrist, using two metallic electrodes encapsulated in the strechable fabric. Additionally, the finger strap enclosing the PPG sensor was connected to the wristband.

The ChestBIT is shown in Figure 3, and it was previously applied in other contexts within the research group (Manso et al., 2018). The chestband used was a Polar chest strap ⁶ , with the ECG sensor connected to the metal buttons of the chest strap. The BITalino MCU was glued to the chest strap, alongside the ACC, the Bluetooth module, and a piezoelectric PZT sensor used to record the RESP signal. The entire electronics was protected with a soft sponge and covered with a synthetic leather textile, to improve endurance. The battery remained outside, in a pocket, and a small opening allowed the battery connector to reach the BITalino power module; this enabled easier natter charging and replacement to ensure near-continuous operation. The on/off button was accessible through the side, since the left side of the textile was not fully closed.

To ensure greater battery life, the batteries used in ArmBIT and ForearmBIT were LiPo with 750 mAh and lasted for 12 h, which required a bi-diary battery exchange. However, the new wearables ChestBIT and WristBIT reduce the physical constrain of battery size, reason for which LiPo batteries of 1,600 mAh were used, requiring a single daily battery change.

The quality of all data recorded was assessed both for the hospital system and for the wearable devices, following the approach found in (Böttcher et al., 2022) for data quality evaluation. In this approach, the quality of the acquired data is assessed globally for each dataset, through statistical metrics (mean, median, standard deviation, minimum value and maximum value). These metrics are applied to the patient data variables, namely: duration, data completeness, and each biosignal quality.

The duration variable consists on the time span from the first timestamp marking the beginning of the acquisition until the last timestamp. It is presented in hours, minutes and seconds (H:M:S). The percentage of usable data (i.e., not loss) is measured by the data completeness. This variable is the ratio between the recorded duration and aforementioned duration variable. The recorded duration is given by all the timestamps associated to a data value between the first and last data points.

The ECG quality was evaluated in segments of 10 s and in three levels: low, medium and high. Low quality corresponds to an unusable segment, whether due to the presence of an artifact, or due to the loss of connection with the body. Medium quality (ECG MQ) corresponds to a segment where it is possible to extract rhythm information, in particular the heart rate and heart rate variability. In this scenario, the R peak detection is reliable, but the rest of the ECG waveform can be indistinguishable. Lastly, the high quality (ECG HQ) corresponds to a reliable and consistent depiction of the ECG wave throughout the segment. This qualitative stratification was also described in previous works and gives a perception of possible applications of the acquired data. To reach these levels, each segment is processed according to the following method: the R peaks locations are found and the instantaneous heart rate is calculated. If the heart rate is between 40 and 200 beats per minute, the low quality level is discarded (the low quality is chosen otherwise). Then, the ECG wave cycles are cross-correlated through Pearson correlation (between 0 and 1). If the cross correlation is above 0.8, then the segment is classified as high quality (and medium quality if below the threshold).

For the ACC, (Nasseri et al., 2020), evaluated the quality of Empatica E4 in the context of video EEG-monitoring, proposing as SQI the ratio between human motion activity frequency and the entire frequency range, as given by Eq. 1, where p refers to the power’s spectrum in the indicated frequency band and FN is the Nyquist frequency (half of the sampling frequency). In this paper, the metric was calculated for 4-s non-overlapping segments and averaged across 10-min intervals. S Q I A C C = P 0.5 , 8 P 0.8 , F N (1)

The signal quality analysis of the RESP signal was calculated following the metric proposed by (Charlton et al., 2021) for a respiratory signal acquired through impedance pneumography. This metric consists on the analysis of valid breaths from segments with 32 s of length. For each segment, the peaks are extracted and used to find breath cycles. The segment is valid if the breaths respect some predefined conditions proposed by the authors. This metric results in a binary separation between high quality and low quality.

The EDA quality was retrieved from (Böttcher et al., 2022), where the binary decision high/low quality is based on two factors. The first is the signal amplitude, which has to be above 0.05 μS; then, the signal amplitude change rate is calculated as the ratio between the min-max amplitude and the first of these two extremes to appear. This rate should be below 0.2 in the assessment of EDA for windows with 2 s.

The aforementioned work, evaluates the quality of the PPG also in a binary low/high quality distinction, using the spectral entropy in windows of 4 s. A spectral entropy value should be below 0.8 for the segment to be considered high quality.

The decision to follow the quality assessment approaches taken by (Nasseri et al., 2020; Böttcher et al., 2022) enabled the comparison between our wearable devices and other datasets and data acquired in similar conditions. Both these works acquired wearable data using the Empatica E4.

When a patient is admitted to the hospital stay, they sign an informed consent from the hospital so that their data can be analysed and used for research purposes. After the hospital system setup, the patient is approached to participate in the wearable data collection as well. If accepted, a new informed consent is given for the use of wearable devices during the hospital stay. The proposed acquisition system collects data unobtrusively and only requires a daily short-time interruption for battery replacement. This section describes the process of accessing the wearable and the hospital data, the structure of the dataset, as well as the several formats in which each data source was saved.

The quality of biosignals was studied for all data acquired by the hospital system, the ChestBIT, the WristBIT and the ArmBIT and these results are summarised in Table 3. The hospital data was acquired for 59 patients totalling almost 6 k hours of ECG data. Data completeness is fairly high with a median of 99.21%. The majority of the signal was classified, on average, as high quality (around 70%), however, on average, 15% of the data was of poor quality. Regarding the ChestBIT, 37 patients used this device, with some sessions being more successful than others (this is observable by the minimum and maximum metrics). The data completeness is lower (in mean and median) when compared to the hospital system, which is expected since the system communicates through Bluetooth, hence being more susceptible to connection losses. Nonetheless, the discrepancy between the mean and the median indicates that the majority of acquisitions was mostly successful in terms of data completeness. In ChestBIT, more segments were discarded as poor quality, however, the median of low quality segments is also around 15% (100 - (ECG HQ + ECG MQ)). Interestingly, the median percentage of ECG HQ is higher in the ChestBIT than in the hospital system, and very few segments were classified as medium quality. The quality of RESP signals is on average 60%, which is reasonable.

The WristBIT was only acquired in 9 patients, still, it comprises 500 h of data, with data completeness percentage comparable to the ChestBIT. The quality of the EDA signal was significantly high, however the PPG quality was significantly lower, since on average, only half of the segments have high quality. This was expected, since the PPG is highly susceptible to noise. Moreover, this is comparable to the results reported by (Böttcher et al., 2022) with the Empatica E4, where the mean and median quality of PPG data was in the range [51.5, 63.3]% for the datasets acquired in hospital settings. The SQI _ACC values of the WristBIT are in agreement with those reported by the study of (Nasseri et al., 2020).

The total number of patients in this dataset is 59. From this cohort, 36 performed the video-EEG in Hospital de Santa Maria (HSM), while 23 performed the video-EEG in Hospital Egas Moniz (HEM). This dataset includes biosignals recorded by our wearable devices, as well as records from the hospital. The classification of seizure types follows the guidelines of the 2017 ILAE’s instruction manual (Fisher et al., 2017a; b). Focal seizures were classified as FUAS, whenever the awareness was not registered or tested. The ictal semiology described in this dataset, follows the 2022 ILAE’s glossary of terms (Beniczky et al., 2022).

The patients collected in this database were performing the video-EEG monitoring for presurgical assessment. This section will detail the distribution of onset localisation and lateralisation, seizure types and ictal semiology reported by the clinical team.

In this cohort (59 patients), 64.41% (N = 38) showed temporal lobe activation during the onset of ictal period. The frontal lobe ictal onset activation showed a prevalence of 25.42% (N = 15), closely followed by the parietal lobe with 18.64% (N = 11). Lastly, the occipital lobe ictal onset was reported in 6.78% (N = 4). The ictal onset activation reported in more than one lobe region (e.g., fronto-temporal) was included in both categories. Concerning lateralisation, 38% had a left hemisphere onset seizure, whereas the epileptogenic zone was in the right hemisphere for 44% of patients. Bilateral ictal onsets were described in 4% of patients from the cohort.

In Figure 6, the distribution of seizures per seizure types is shown, regarding the seizure types which were reported in the video-EEG reports. The total number of seizures in the PreEpiSeizures dataset is 348, distributed a across six seizure types (FBTCS = 27, FIAS = 95, FUAS = 35, FAS = 82, subclinical = 89 and unknown = 20). These values are illustrated as a percentage in Figure 6, where the predominance of focal seizures is notable. The classification of seizures as Unknown was chosen when there was no information regarding the type of seizure, whereas FUAS correspond to focal seizures, in which only the awareness was not tested.

Besides the classification of seizure types, the video-EEG reports also contained some details regarding the events occurring during the seizure. In Figure 7, the reported ictal semiology from the dataset is displayed in 6 categories. The Elementary Motor Behaviour was found in 120 seizures and includes Dystonic (N = 23), head version or head orientation (N = 45), eye blinking (N = 12), myoclonic/clonic (N = 14), tonic (N = 6) and tonic-clonic (N = 27). The Complex Motor Behaviour (N = 183) included automatisms and hyperkinetic behaviour (N = 3). The most common automatisms were distal (N = 67) followed by oroalimentary (N = 59) and proximal (N = 16). Other reported automatisms were: verbal (N = 11); vocal (N = 6); mimic (N = 7); nose-wiping (N = 9) and Rinch (N = 5). Sensory phenomena (N = 33) was mostly populated by the somatosensory symptoms (N = 15), followed by cephalic aura (N = 5). Other reported sensory phenomena was gustatory (N = 1) and olfactory (N = 2). The most reported autonomic semiology was epigastric aura (N = 20), followed by hyperventilation (N = 12), tachycardia (N = 11) and piloerection (N = 1). Few seizures reported cognitive ictal phenomena (N = 27), where the most common symptom was autoscopy (N = 12), followed by dejá vu (N = 9), ictal aphasia (N = 3) and dysarthria (N = 3). Only 2 seizures reported emotional phenomena and the symptom was fear.

In Table 3, the total duration of ArmBIT, ChestBIT and WristBIT is shown separately. Figure 8 gives a visual depiction of the number of recorded hours by each modality. The different colours represent the device responsible for each measurement. The least represented modality is EMG with 156 h, 14 min and 54 s of data. Since this measurement required gel electrodes, it was discarded in the new devices. The EDA totals 666 h, 19 min and 2 s of data, which is decomposed in 156 h of using pre-gelled electrodes (with ForearmBIT) and 510 h, 4 min and 7 s of metallic electrodes in the WristBIT. The chest-ACC corresponds to 2,211 h of data, whereas the wrist-ACC totals 510 h. Additionally, the ArmBIT and ForearmBIT add more 312 h to the ACC data. Regarding RESP, 2,211 h were recorded so far, using only the ChestBIT. The ECG was captured using pre-gelled electrodes in 156 h (ArmBIT) and the ChestBIT has so far recorded 2,211 h. The total number of recorded hours by all wearables is 2,721 h, 57 min and 36 s. The broader use of ChestBIT instead of WristBIT was based on three arguments: an easier reproducibility; a higher reliability of ECG instead of PPG to record cardiac activity; and being more discreet, since it is placed under the clothes. This last point is aligned with previously surveyed preferences of PWE (Simblett et al., 2020).

To conclude the overview of the PreEpiSeizures dataset, the aforementioned Table 1 provides a side-by-side characterisation of this dataset with those available within the state-of-the-art. Three of the public databases (Bern-Barcelona (Andrzejak et al., 2012), New Delhi EEG (Swami et al., 2016) and Epilepsy Seizure Recognition (Andrzejak et al., 2001)) only presented segments of a few seconds duration. Bonn-Barcelona micro- and macro-EEG database (Martínez et al., 2020), comprises data from epileptic patients without epileptic seizures.

The comparison of our dataset with EPILEPSIAE (Klatt et al., 2012) and Siena Scalp EEG (Detti, 2020) is performed with regards to the ECG recorded by the hospital system (this is the number in the left side of columns patient and total hours). Our dataset size, both in terms of patients, and in total recorded hours, surpasses its peers.

Regarding wearable data, the only other dataset found with wearable data was the My Seizure Gauge Wearable, which does not contain ECG, PZT or Chest Motion data. Both datasets duration is quite similar (our wearable duration is the right number of column Total Hours in Table 1). Concerning the number of seizures and the variety of seizure types, PreEpiSeizures is similar to EPILEPSIAE and My Seizure Gauge Wearable. Regarding the number of recorded hours, the proposed dataset also closely follows those of the literature.