Seizure classification using a multimodal seizure monitoring system (Nelli) in Dravet and Lennox–Gastaut syndromes: A non‐randomized, single‐center feasibility study

Abstract

Objective

This study aimed to assess the performance of the Nelli seizure monitoring system in detecting and classifying seizures during sleep or while at rest in bed in patients with Lennox–Gastaut syndrome (LGS) and Dravet syndrome (DS).

Methods

We conducted a non‐interventional, single‐center feasibility study from August 2023 to March 2024, involving 20 patients aged ≥2 years diagnosed with DS or LGS. Participants used Nelli for home‐based seizure monitoring during sleep or while at rest in bed for 4 weeks. Seizures were detected and classified by Nelli, and results were compared to epileptologist reviews and seizure diaries.

Results

Of 20 enrolled patients, 14 (70%) who experienced seizures at rest were included in the analyses. Among them, Nelli detected 368 seizures, with an accuracy of 97.8%, as confirmed by independent reviewers. Eight seizures (2.2%) detected by Nelli were false positives, identified as part of a single seizure episode. Of the 14 patients, only 35.7% reported experiencing seizures in their diaries, and only 26.1% of the seizures were documented. Seizure durations ranged from 6 to 396 s, with considerable variation. Nelli demonstrated high accuracy in seizure classification (Gwet agreement coefficient [AC1] = .81–1.00) in nine of 14 cases. However, in three of 14 patients, moderate accuracy (AC1 = .41–.60) was observed due to challenges in classifying seizures in patients with high seizure frequency or suboptimal device positioning. The average classification accuracy of Nelli for tonic–clonic seizures was .99 (150/152 seizures), tonic seizures .55 (102/186), clonic seizures 1.00 (3/3), focal motor seizures .89 (16/18), and myoclonic seizures 1.00 (1/1).

Significance

Nelli demonstrated high sensitivity and classification accuracy for detecting and categorizing seizures in bed in patients with DS and LGS, outperforming seizure diaries and providing a reliable tool for seizure monitoring in home settings.

Key points.

• The Nelli seizure monitoring system offers a potential solution for seizure detection, classification, and monitoring in both home and hospital settings.

• This non‐randomized feasibility study assessed Nelli's accuracy in detecting seizures at rest in patients ≥2 years old with Dravet or Lennox–Gastaut syndrome.

• Fourteen of 20 patients had seizures at rest. Nelli detected 368 seizures, 360 confirmed by reviewers. Only 94 were recorded in seizure diaries.

• Seizure durations varied across patients, ranging from 6 to 396 s.

• Nelli showed high accuracy (AC1 = .81–1.00) in nine of 14 patients, substantial (AC1 = .61–.80) in two of 14, and moderate (AC1 = .41–.60) in three of 14.

1. INTRODUCTION

Lennox–Gastaut syndrome (LGS) and Dravet syndrome (DS) are rare developmental and epileptic encephalopathies that typically emerge in early childhood. ¹ , ² LGS is an electroclinical diagnosis that encompasses various etiologies, including structural, genetic, infectious, metabolic, immune, and other causes. ³ , ⁴ In contrast, DS is primarily a genetic epilepsy, most often linked to mutations in the SCN1A gene, ⁵ which distinguishes it from the more heterogeneous nature of LGS.

Both conditions are characterized by multiple, frequent, difficult to treat seizures, creating significant challenges in management. ⁶ , ⁷ Seizure types in these syndromes include tonic, atonic, atypical absence, myoclonic, focal, clonic, tonic–clonic, and hemiclonic seizures. ¹ , ² , ⁶ , ⁸

The global incidence and prevalence per 100 000 individuals vary widely (DS: incidence = 2.2–6.5, prevalence = 1.2–6.5; LGS: incidence = 14.5–28.0, prevalence = 5.8–60.8). ⁹ These syndromes typically result in motor, cognitive, and behavioral abnormalities with detrimental effects on quality of life (QoL), social and personal relationships, education, career, and finances of patients. ¹ , ² In addition, caregivers of these patients often face significant time burdens, reduced health‐related QoL, and work/activity impairment, leading to physical and emotional strain. The resulting loss of work productivity, increased health care costs, and financial strain contribute to broader socioeconomic consequences, placing a heavy burden on both caregivers and society. ¹⁰ Improved seizure management, leading to reduced frequency and additional seizure‐free intervals, enhances the QoL for both patients and caregivers. ¹¹

Although treatment resistance in DS and LGS is primarily driven by their underlying complex etiologies, a significant challenge in managing seizures more broadly is the inaccurate reporting of seizure frequency, classification, intensity, and duration. ¹² On average, up to half of motor seizures go unnoticed, with even greater underreporting for nocturnal seizures, which are documented only 15% of the time. ¹³ Such underreporting complicates the assessment of treatment effectiveness and may hinder optimal clinical decision‐making. ¹⁴

Video‐electroencephalography (VEEG), also called long‐term video‐electroencephalographic monitoring, remains the gold standard for seizure detection, but it necessitates hospital admission for these patients. Seizure diaries and patient seizure counts often prove unreliable due to postictal amnesia and patients'/caregivers' inability to observe and accurately describe all seizures. ¹⁴ , ¹⁵ , ¹⁶ Wearable devices have also been developed to assist in tracking and characterizing seizures, but their clinical accuracy lacks strong evidence, and they may not be well tolerated, particularly in children or those with intellectual disabilities. ¹² Additionally, for long‐term use, these devices can be challenging to use, and patients may become nonadherent over time. This highlights the need for automated seizure detection devices, especially those that integrate video monitoring in patients' homes, to characterize seizures during both sleep and wakefulness in patients with DS or LGS. Recently, there has been growing interest in smartphone videos recorded by family members or caregivers. These videos, being user‐friendly, cost‐effective, and capable of capturing extensive information on seizure characteristics, hold promise in aiding seizure classification and differential diagnosis, including psychogenic nonepileptic seizures. ¹⁷

The Nelli seizure monitoring system, hereafter referred to as Nelli, is a hybrid system utilizing computer vision and machine learning to detect and classify major motor seizures, while also identifying kinematic data commonly linked with seizures. This system, augmented by human experts' visual assessment, offers a potential solution for enhancing seizure detection, classification, and monitoring in both home and hospital settings. ¹² , ¹³ , ¹⁶ It assesses treatment necessity and can monitor treatment effectiveness. ¹⁸ If successfully validated and adapted for DS and LGS, it could support diagnosis of disease and treatment monitoring, safety assessments, and real‐world evidence generation for various use cases, with particularly strong potential for LGS. ¹⁹ This non‐randomized, single‐center feasibility study in pediatric and adult patients with DS and LGS aimed to evaluate the accuracy of Nelli in detecting major motor seizures at rest of various types in pediatric and adult patients with DS and LGS for Nelli performance assessment and future algorithm improvement.

2. MATERIALS AND METHODS

2.1. Study design

This non‐interventional, non‐randomized, open‐label, single‐center, single‐arm feasibility study, conducted between August 2023 and March 2024 at the Danish Epilepsy Center Filadelfia, included pediatric and adult patients (aged 2 years and older) with a definite clinical diagnosis of DS or LGS.

2.2. Study population

Patients were eligible for inclusion if they, or a legally authorized representative, provided written informed consent. Additionally, they had to be aged 2 years or older and have a clinical diagnosis of DS or LGS, established at an experienced epilepsy center (Filadelfia Hospital, Denmark) based on comprehensive diagnostic evaluation following standard clinical practice. For DS, this included—but was not limited to—genetic testing results, whereas for LGS, electroencephalographic (EEG) findings were considered among other clinical features. All included patients had previously documented seizures, which occurred during sleep/at rest or at bedtime through VEEG or another reliable method. Furthermore, they were required to have an average seizure frequency of at least one major motor seizure per week, supported by adequate documentation. Major motor seizures in this study were defined similarly to our previous study ²⁰ as seizures containing a motor component with a behavioral duration lasting more than 10 s. This threshold was selected based on literature documenting clinically relevant ictal phenomena and widely accepted standards for electrographic seizures. ²¹ The category of “focal onset, motor seizures” included seizures that began with localized motor activity (e.g., focal limb jerking or posturing) but did not meet the criteria for generalized seizure types such as tonic or clonic seizures.

Patients also needed to have been on stable anti‐seizure medication or other treatment for at least 4 weeks prior to enrollment, with no planned medication changes during the study period. Furthermore, they should have agreed to daily use of the device for monitoring seizures for up to 4 weeks and have the capability to understand and comply with study requirements.

Patients were excluded if they had unstable housing conditions or an active severe behavioral illness that could potentially interfere with compliance with study procedures and monitoring using the device, as determined by the investigator. Additionally, exclusion criteria included comorbidities that might disrupt sleep or exacerbate seizures, as determined by the investigator, or current participation in another clinical drug or device trial.

2.3. Nelli system

2.3.1. Identification and description of device

Nelli was designed and manufactured by Neuro Event Labs, located at Biokatu 10, 33520 Tampere, Finland, to complement seizure monitoring for both adults and children in either home or medical settings while they are at rest. The device relies on automated analysis of audio and video (A/V) data obtained through the personal recording unit (PRU) hardware accessory. Its purpose is to identify both epileptic and nonepileptic seizure events characterized by positive motor activity. The system furnishes A/V data along with objective summaries of the semiological aspects of identified events. This assists physicians in categorizing seizures and related peri‐ictal occurrences.

Nelli serves as a non‐EEG physiological signal‐based system for detecting and quantifying seizures. It operates as a semiautomated (hybrid) system, wherein potentially pertinent recording periods are chosen for later review by a human expert. Additionally, Nelli incorporates an artificial intelligence algorithm, which detects potentially significant clinical episodes, categorizes their seizure type, and provides information about their duration and, if applicable, relative intensity.

2.3.2. Device components

Nelli consists of software and hardware components (Figure 1). Nelli software has three components: (1) the analysis pipeline, which is a multilayered algorithmic model that extracts physiological signals from A/V to identify events indicative of seizure activity; and (2) the user interface, which is accessible to authorized medical personnel to view patient information, patient reports and modify event annotations (the user interface is accessible from any standard browser with an internet connection after proper secure login); and (3) Nelli firmware, which is used to control the commercial off‐the‐shelf components that comprise the PRU. The PRU is an accessory to Nelli software. It is composed of a small form factor industrial computer configured with a hardened embedded Linux‐based operating system and Nelli firmware and software. These are used to control commercial off‐the‐shelf hardware components that make up the PRU. This includes a stereoscopic near‐infrared camera with an illuminator, a microphone, wired and/or wireless interfaces for data communication, and a solution for mounting the PRU so that the patient's bed fills the camera's field of view during periods of rest. The version of Nelli used in this study included software version 7.5.0 and PRU models 1.8 and 1.9.

FIGURE 1.

Nelli components and interactions. PRU, personal recording unit.

2.3.3. Conformité Européenne marking and medical device classification

Nelli has received Conformité Européenne (CE) marking and is approved as a medical device in the European Union (EU; registration No.: FI‐CA01‐2020‐0612). In the EU, Nelli conforms to the class I EU MDD 93/42/EEC requirements and is transitioning to a class IIa medical device in accordance with EU MDR 2017/745 by May 27, 2024. In the United States, Nelli is a class II medical device under the US Food and Drug Administration's regulatory name “Non‐Electroencephalogram (EEG) Physiological Signal Based Seizure Monitoring System” (product code: POS, classification: 882.1580). The CE‐marked Nelli hybrid system has been recommended for clinical use in Finland by the National Coordinating Group for Drug‐Resistant Epilepsy. ¹²

2.4. Study procedure

Patients underwent a 4‐week monitoring period at home using the Nelli device. The PRU was placed above the patient's bed, mounted on the ceiling or in a device holder. A trained technician installed the device to ensure that the device was placed in accordance with specifications required for reliable recording of seizure events. The patient or caregiver ensured that the device was on before periods of sleep or rest to initiate the recording. The patient did not need to turn off the device during short periods away from the device, for example, during bathroom breaks. The A/V recordings were collected via device and uploaded to the cloud. The Nelli software algorithm automatically detected motor events and categorized seizures at bedtime. International League Against Epilepsy (ILAE) seizure classifications were then designated by a trained Nelli annotator (medical technician; Figure 1). Concurrently, caregivers of the patients were requested to maintain a seizure diary to document seizure occurrences.

All events interpreted as seizures and classified based on ILAE seizure classification by the NELLI hybrid system were manually evaluated and reclassified by two experienced independent epileptologists (hereby referred to as reviewers). This also served as a reference standard for this study according to standards for testing and clinical validation of seizure detection devices. ²² Following the 4‐week monitoring period, the outcomes assessed included (1) seizure count and types recorded by Nelli compared to those validated by two independent epileptologists; (2) seizure count recorded by Nelli versus those documented in the seizure diary; (3) duration of each seizure recorded by Nelli, along with the duration (range) for each seizure type; and (4) individual seizure classifications as recorded by Nelli compared to those validated by the two independent epileptologists.

2.5. Statistical analysis

The baseline demographics and seizure types of the study population were characterized using descriptive statistics. Baseline characteristics included sex, age category (pediatric: ≥2 to ≤12 years; adolescent: >12 to <18 years; adult: ≥18 years), type of disease, and type of setting. The data are presented as mean with SD, median with ranges (minimum–maximum), proportions, and frequencies.

Accuracy evaluations compared Nelli seizure type classification performances against seizure classification by two experienced independent epileptologists. When seizure type classification disagreement occurred between the two experienced independent epileptologists, consensus review was done to determine the seizure type.

Gwet agreement coefficient (AC1) was used to assess agreement between the epileptologists and semiautomatic Nelli's performance as a measure of accuracy (sensitivity and positive predicted value as a method of specificity measure). ⁸ , ²³ The following classification was used to assign accuracy as an interpretation of the degree of AC1 agreement: <0, no agreement/no accuracy; .01–.20, slight agreement/slight accuracy; .21–.40, fair agreement/fair accuracy; .41–.60, moderate agreement/moderate accuracy; .61–.80, substantial agreement/substantial accuracy; and .81–1.00, almost perfect agreement/high accuracy.

2.6. Ethical statement

To ensure the quality and integrity of the research, this study was conducted in accordance with the principles outlined in ISO 14155:2020 (Clinical Investigation of Medical Devices for Human Subjects—Good Clinical Practice), EU MDR 2017/745, and the Declaration of Helsinki and its amendments, as well as any applicable national guidelines. Ethics committee approval was obtained on May 30, 2023 from the De Videnskabsetiske Medicinske Komitéer. All participants or their caregivers in the study provided written informed consent, ensuring that they understood the scope of the study and voluntarily agreed to take part in the study.

3. RESULTS

3.1. Demographic and baseline characteristics of study population

A total of 20 patients were included, with a balanced gender distribution (55.0% males and 45.0% females). The average age of the patients was 20 (SD = 9.9) years, and median age was 16.5 years (range = 7–43 years; Table 1). Based on predefined age groups, three patients (15.0%) were classified as pediatric (≥2 to ≤12 years), eight patients (40.0%) as adolescents (>12 to <18 years), and nine patients (45.0%) as adults (≥18 years). Among the 20 patients, 13 (65.0%) had DS and seven (35.0%) had LGS. Thirteen patients (65.0%) utilized Nelli in a home setting, whereas the remaining seven (35.0%) used Nelli in an institutional setting (Table S1).

TABLE 1.

Demographic and baseline characteristics of patients with DS and LGS (N = 20).

Patient characteristics	n (%)
Age, years, mean (SD)	20 (9.9)
Age, years, median (range)	16.5 (7–43)
Sex
Males	11 (55.0%)
Females	9 (45.0%)
Syndrome
DS	13 (65.0%)
LGS	7 (35.0%)
Setting
Home	13 (65.0%)
Institution	7 (35.0%)

Abbreviations: DS, Dravet syndrome; LGS, Lennox–Gastaut syndrome.

Of the 20 patients enrolled, 14 (70%) experienced seizures at rest during the monitoring period. The mean (SD) age of these patients was 21.9 (10.5) years, and median age was 18.0 (range = 8–43) years, with 42.9% females and 57.1% males. Based on predefined age groups, eight patients (57.1%) were adults (≥18 years), five (35.7%) were adolescents (>12 to <18 years), and one (7.1%) was a pediatric patient (≥2 to ≤12 years). Among them, nine (64.3%) had DS and five (35.7%) had LGS. Of the 14 patients, nine (64.3%) were monitored in a home setting, whereas five (35.7%) were monitored in an institutional setting (Table S1).

3.2. Seizure counts and types recorded by Nelli compared to those validated by two independent epileptologists

Nelli detected a total of 368 seizures at rest, and independent epileptologists confirmed 360 of these (97.8%). Eight seizures (2.2%) identified by Nelli were not confirmed by independent epileptologists because they were identified as part of a single seizure episode rather than distinct, independent events and therefore were considered false positives in Nelli (Tables 2 and 3). These events included tonic seizures (n = 4) in Patient 11, tonic seizures (n = 3) in Patient 12, and focal motor seizures (n = 1) in Patient 17, where the seizure classification differed between Nelli and the reviewers. However, only one of these was an actual additional (false positive) detection by Nelli; in the remaining cases, the classification—not the presence—of the seizure was in disagreement. Additionally, in Patient 13 and Patient 16, Nelli detected clonic and focal motor seizures (respectively), although the number of seizures counted by Nelli and the epileptologists remained the same (Table S2 and Table 2).

TABLE 2.

Type of seizures detected by Nelli and confirmed by reviewers.

Deidentified patient serial number	Type of seizures detected by Nelli			Type of seizures confirmed by reviewers
1	Tonic–clonic			Tonic–clonic
3	Tonic–clonic	Tonic		Tonic–clonic	Tonic
4	Tonic–clonic	Clonic		Tonic–clonic	Clonic
7	Tonic–clonic			Tonic–clonic
8	Tonic–clonic			Tonic–clonic
9	Tonic–clonic			Tonic–clonic	Tonic a
11	Tonic–clonic	Tonic b		Tonic–clonic
12	Tonic–clonic	Focal motor	Tonic b	Tonic–clonic	Focal motor
13	Tonic–clonic	Tonic	Clonic b	Tonic–clonic	Tonic
14	Tonic–clonic	Tonic		Tonic–clonic	Tonic
16	Tonic	Focal motor b		Tonic
17	Tonic–clonic	Tonic	Focal motor b	Tonic–clonic	Tonic
18	Tonic–clonic	Tonic		Tonic–clonic	Tonic
20	Tonic–clonic	Tonic	Myoclonic	Tonic–clonic	Tonic	Myoclonic

^a Seizure not detected by Nelli but noted by reviewers.

^b False positive.

TABLE 3.

Number of seizures each patient experienced as reported by Nelli, seizure diary, and reviewers.

Deidentified patient serial number	Seizures reported by Nelli	Seizures reported by reviewers	Seizures reported in seizure diary	AC1 score (0–1) in terms of Nelli's ability to identify the types of major motor seizures the patient was experiencing	AC1 score (0–1) in terms of Nelli's ability to classify each seizure event for that given patient compared to reviewers
1	25	25	0	1.00	1.00
3	9	9	2	1.00	1.00
4	10	10	2	.70	.70
7	4	4	0	1.00	1.00
8	1	1	0	1.00	1.00
9	49	49	0	1.00	.82
11	16	12	0	.75	.75
12	25	24	0	.96	.88
13	12	12	24	1.00	.83
14	39	39	25	1.00	.90
16	8	8	0	1.00	.88
17	65	62	0	.95	.54
18	44	44	0	1.00	.52
20	61	61	41	1.00	.49
Total	368	360	94	–	–

Abbreviation: AC1, Gwet agreement coefficient.

For tonic–clonic seizures, Nelli recorded 242, and the independent epileptologists confirmed 152. For tonic seizures, Nelli documented 102, whereas the epileptologists noted 186. Nelli observed five clonic seizures, and the epileptologists confirmed 3. Both Nelli and the epileptologists recorded 18 focal motor seizures. Finally, for myoclonic seizures, both Nelli and the epileptologists documented one each (Table S2). Nelli correctly detected tonic seizures (n = 102), tonic–clonic seizures (n = 150), focal motor seizures (n = 16), clonic seizures (n = 3), and myoclonic seizures (n = 1). Nelli correctly detected tonic–clonic seizures in 13 patients, tonic seizures in seven patients and myoclonic, focal motor, and clonic seizures in one patient each. These classifications were all validated by independent epileptologists, demonstrating 100% sensitivity of Nelli in identifying these types of major motor seizures experienced by each patient. However, tonic seizures in Patient 9, as classified by the epileptologists, were correctly detected as seizures by Nelli but misclassified as tonic–clonic seizures, resulting in an 85% sensitivity rate (Tables 2 and 3).

3.3. Seizure count recorded by Nelli versus those documented in the seizure diary

Of the 14 patients who experienced seizures at bedtime according to Nelli, caregivers of only five (35.7%) patients (three with DS and two with LGS) reported seizures in their seizure diary (Table 3). Only 94 of 360 (26.1%) seizures were recorded in a seizure diary, and 266 of 360 (73.9%) of seizures detected by Nelli were missing in the seizure diaries. No seizures documented in a seizure diary were missed by Nelli (Table 3).

3.4. Seizure duration measured using Nelli

The data for seizure durations across various seizure types, as detected by Nelli, showed a wide range of durations depending on the type of seizure, with some patients exhibiting longer durations in specific seizure types. Overall, seizure durations varied across patients, ranging from 6 to 396 s. For tonic–clonic seizures, the duration ranged from 21 to 396 s. The longest tonic seizure duration was recorded for Patient 14 (8–255 s), whereas the shortest tonic seizure was observed in Patient 16 (6–100 s). Clonic seizures were experienced by Patient 4 only, with a duration range of 74–123 s. Similarly, focal motor seizures were recorded only for Patient 12, with a duration range of 19–86 s. A myoclonic seizure was recorded for Patient 20 only, with a duration of 15 s (Table 4).

TABLE 4.

Seizure duration recorded by Nelli.

Deidentified patient serial number	Range of seizure duration, s a
	All seizure types	Tonic–clonic	Tonic	Clonic	Focal motor	Myoclonic
1	73–256	73–256
3	74–159	74–159	145
4	74–277	91–277		74–123
7	98–257	98–257
8	363	363
9	53–396	63–396
11	21–180	21–180
12	19–333	213–333			19–86
13	23–174	174	23–145
14	8–255	134–206	8–255
16	6–100		6–100
17	13–171	65–171	13–127
18	8–273	100–273	8–182
20	15–266	145–266	43–155			15

Note: These data only represent the duration of seizures correctly detected by Nelli.

^a Ranges are presented as minimum–maximum.

3.5. Individual seizure classifications as recorded by Nelli versus those validated by two independent epileptologists

Nelli reported very high classification accuracy (AC1 = .81–1.00) in the majority (9/14) of cases, substantial accuracy (AC1 = .61–.80) in two of 14 patients, and moderate accuracy (AC1 = .41–.60) in three of 14 patients (Table 3). Average classification accuracy of Nelli for tonic–clonic seizures was .99 (150/152 seizures), tonic seizures .55 (102/186), clonic seizures 1.00 (3/3), focal motor seizures .89 (16/18), and myoclonic seizures 1.00 (1/1). No common root cause was identified for the moderate ratings received by three patients. Two patients (Patients 17 and 20) were rated moderately due to the similarity between tonic–clonic and tonic seizures, compounded by the high frequency of seizures in those patients. These factors made it difficult to classify each seizure accurately, for both the Nelli system and the two epileptologists. Another patient (Patient 18) received a moderate rating due to suboptimal camera angle and distance during installation, which created additional challenges for both Nelli and the independent epileptologists in classifying the seizures at rest.

4. DISCUSSION

This feasibility study on patients with LGS and DS demonstrates clinical utility of Nelli. The system achieved a seizure detection accuracy of 97.8%, with only 2.2% false detections, that is, seizures identified by Nelli but not confirmed by independent epileptologists. The Nelli system was able to detect tonic–clonic seizures, tonic seizures, focal seizures, clonic seizures, and myoclonic seizures. This aligns with the expected seizure profile for patients with LGS and DS, as outlined in studies by previous literature. ⁸ , ²⁴ , ²⁵ , ²⁶ The findings suggest that Nelli can effectively complement effective methods of seizure evaluation, such as the gold standard (VEEG) and expert review by epileptologists.

The identification of seizure types using Nelli in nine of 14 patients was highly accurate, with a Gwet AC1 coefficient > .8, indicating strong agreement between Nelli's detection and independent expert review. The accuracy for identifying most seizure types—such as tonic–clonic, clonic, and focal motor seizures—was high, consistent with the previous studies. ⁸ , ¹⁶ However, there was a single instance of missed detection of tonic seizures by Nelli. This high level of accuracy suggests that Nelli is a reliable tool for seizure type classification, although the missed tonic seizure may be attributed to factors such as the specific characteristics of the seizure or limitations in detection under certain conditions. Overall, these results demonstrate Nelli's potential to provide accurate seizure data, which can complement clinical assessments and improve patient monitoring.

False detection of seizures by the Nelli system was likely due to factors such as obscurity, uncertainty, or insufficient evidence in the detection process. In Patient 11, Nelli detected four tonic seizures that were not confirmed by reviewers. In Patient 12, Nelli identified three tonic seizures that the reviewers considered part of a single seizure episode rather than distinct, separate events. Similarly, in Patient 17, one focal motor seizure detected by Nelli was viewed by reviewers as part of a single seizure episode rather than a distinct, individual event.

In this study, seizure diaries were maintained by caregivers, as patients with DS and LGS typically cannot self‐report due to cognitive or developmental impairments. When seizure count recorded by the Nelli system was compared with that documented in the seizure diaries, no seizures documented in the diaries were missed by Nelli, but the data in the seizure diaries were neither accurate nor complete. It is clinically important that of the 14 patients with seizures detected by the Nelli, only 35.7% reported them in their diaries. Overall, of 360 seizures, only 26.1% were documented in the diaries, whereas 73.9% were missed. These findings regarding seizure diaries are consistent with the previous studies conducted in epilepsy populations. ¹⁸ , ²⁷ , ²⁸ A scoping review by Egenasi et al. highlights that although seizure diaries remain useful for epilepsy management in resource‐poor settings where modern devices for objective seizure detection are unavailable or unaffordable, challenges persist. Issues such as patient compliance, the reliability of patient‐reported data, and the validity of seizure diary entries are significant obstacles in clinical practice. ²⁹

The low reporting rate of seizures in diaries in our study may reflect both the nocturnal timing of events and the challenges caregivers face in consistently recognizing and documenting seizures. Kerling et al. highlighted that the information provided by patients with epilepsy regarding the number of seizures they experience is often unreliable, particularly in individuals where events occurred during sleep or originated in (or propagated to) the left temporal lobe. ²⁸ Other potential reasons could include impaired awareness during seizures or cognitive challenges that prevented the patients from maintaining accurate records. Independent of the reasons for the gaps, relying on seizure diaries alone would have provided an incomplete and potentially misleading picture of the patients' seizure frequency and semiology. Previous studies have shown that Nelli is effective at detecting seizures at rest and offers a more reliable alternative to traditional seizure diaries. ²⁰ , ³⁰

Home VEEG telemetry (VT) is now an established tool in epilepsy care. ³¹ , ³² However, a few limitations have been highlighted in recent literature. A scoping review by Milne‐Ives et al. reported several challenges, including inferior video quality compared to hospital‐based systems, technical difficulties in setting up equipment in patients' homes, and usability issues for patients and caregivers.³³ Additionally, VT may not be appropriate for individuals requiring sleep deprivation or antiepileptic drug withdrawal prior to EEG, due to safety concerns. Furthermore, event capture rates in home settings may be reduced, as close supervision is more difficult during remote, extended monitoring. ³¹ , ³³ Moreover, unlike VT, which requires EEG setup, Nelli provides continuous, noninvasive video monitoring in the home environment over extended periods. These complementary roles suggest potential for future comparative studies, particularly in populations with complex epilepsy and cognitive impairment where EEG placement may be challenging; moreover, the objective video data that Nelli provides enable specialists to better categorize patient risk and tailor treatment plans without having to monitor patients in a hospital environment. ⁸ , ¹² , ¹³

Given the severe cognitive impairment and treatment resistance in DS and LGS, seizure identification is particularly difficult. In this context, Nelli provided more actionable data than seizure diaries and performed comparably to epileptologist review. Moreover, no safety concerns have been identified with Nelli's postmarket use. This reinforces the clinical value and safety of Nelli. ¹²

Overall, the data collected provided valuable insights into the performance of Nelli in a real‐world context. The accuracy of seizure detection and classification was highly dependent on the correct installation of the device, emphasizing the importance of proper setup for optimal results. Additionally, privacy concerns were carefully addressed, and thorough documentation was provided to institutions, which facilitated their participation and ensured a smooth process. However, there were a few limitations. First, the small sample size may affect the generalizability of the findings. Second, enrollment was limited by institutional circumstances, which may have influenced patient representation. Third, as a feasibility study, there was no direct comparison between Nelli and VT in terms of clinical utility or cost‐effectiveness. Additionally, atypical absences and other nonmotor seizures—common in both LGS and DS—were beyond the detection scope of the Nelli system, which relies on visible motor features. As such, nonmotor events may have gone undetected.

Previous research has also identified limitations in the Nelli system's ability to detect subtle motor seizures, such as single myoclonic jerks, epileptic spasms, and other very short seizures. These types of seizures were excluded from analyses in some algorithm studies ⁸ , ²⁰ due to their brief duration and minimal motor manifestations, which pose challenges for video‐based detection systems. However, in the present hybrid study, both algorithm/device outcome and subsequent medical review, based on device outcome, were employed. This approach allowed for inclusion of subtle motor seizures, as false positives triggered comprehensive video review by clinicians.

5. CONCLUSIONS

The A/V‐based Nelli seizure detection and classification system demonstrated high accuracy in detecting seizures and substantial accuracy in classifying seizure types in patients with DS and LGS. It outperformed traditional seizure diaries, offering a reliable and effective tool for seizure monitoring and management in these complex epilepsy syndromes, particularly in home settings. The ability of Nelli to detect seizures at rest, often missed by caregivers and traditional documentation methods, further underscores its potential to improve patient care and optimize treatment. This approach offers a promising solution for enhancing seizure monitoring, supporting clinical decision‐making, and ultimately improving the QoL for both patients and caregivers.

AUTHOR CONTRIBUTIONS

Line Kønig Wilms: Conceptualization; writing—review & editing. Morten I. Lossius: Conceptualization; writing—review & editing. Kaapo Annala: Conceptualization; methodology; software; formal analysis; writing—review & editing. Jonas Abdel‐Khalik: Conceptualization; writing—review & editing. Lena Fanter: Conceptualization; writing—review & editing. Kaisa Elomaa: Conceptualization; funding acquisition; writing—review & editing. Jukka Peltola: Conceptualization; methodology; writing—review & editing.

FUNDING INFORMATION

The study was conducted by Neuro Event Labs (NEL) and Filadelfia Hospital and was fully funded by Takeda Pharma, Lindhagensgatan 20, 112 51 Stockholm. The funding covered the work of the principal investigator and NEL employees, as well as payments to NEL for providing devices, installation, documentation, analysis efforts, and the development of the finalized study report. Manuscript writing services were provided by Quantify Research, part of the Athagoras group, and were funded by Takeda Pharma. The article processing charges for publication were also funded by Takeda Pharma.

CONFLICT OF INTEREST STATEMENT

L.K.W. has no conflicts of interest. M.I.L. reports personal fees for lectures from Eisai, Jazz, UCB, and Arvelle, outside the submitted work. K.A. is an employee and shareholder of Neuro Event Labs, the company that provided the equipment and technology used in the study. J.A.‐K. has no conflict of interest to disclose. He was an employee of Takeda Nordics during the conduct of the study. L.F. is an employee and shareholder of Takeda Germany. K.E. is employed by Takeda and holds Takeda stocks/stock options. J.P. has received clinical trial grants from Eisai, UCB, and Bial, research grants from Angelini Pharma, Eisai, Jazz Pharma, Medtronic, UCB, and LivaNova, speaker's honoraria from LivaNova, Angelini Pharma, Eisai, Jazz Pharma, Medtronic, Orion Pharma, and UCB, and travel support from LivaNova, Eisai, Medtronic, and UCB; and has sat on the advisory boards for LivaNova, Angelini Pharma, Jazz Pharma, Eisai, Medtronic, and UCB. He is the medical director and a shareholder of Neuro Event Labs. We confirm that we have read the Journal's position on issues involved in ethical publication and affirm that this report is consistent with those guidelines.

Supporting information

Table S1.

Wilms LK, Lossius MI, Annala K, Abdel‐Khalik J, Fanter L, Elomaa K, et al. Seizure classification using a multimodal seizure monitoring system (Nelli) in Dravet and Lennox–Gastaut syndromes: A non‐randomized, single‐center feasibility study. Epilepsia. 2026;67:164–174. 10.1111/epi.18640

DATA AVAILABILITY STATEMENT

Due to the lack of patient consent on data sharing and compliance with the General Data Protection Regulation, the data (e.g., videos, patient names, or other identifiers) cannot be made available.