Pulse‐based photoplethysmography quality assessment improves wearable seizure detection performance

Abstract

Objective

Wearable photoplethysmography (PPG) is increasingly used for seizure detection due to its ability to unobtrusively estimate heart rate (HR). However, the vulnerability of PPG to body movements often leads to unreliable HR estimates, compromising accurate seizure detection. Although this limitation is well recognized, methods for rejecting artifactual HR measures—and their impact on seizure detection performance—have not yet been studied.

Methods

We sourced data from the PROMISE trial, a home‐based study in children with refractory epilepsy that evaluated the performance of the NightWatch (NW) in detecting nocturnal motor seizures. We applied a novel pulse‐based quality assessment (PQA) method that evaluates the quality of each individual PPG pulse and excludes compromised ones prior to HR estimation. We compared the performance of PQA with two alternatives: HR‐based quality assessment (HRQA), rejecting extreme outliers in consecutive inter‐beat intervals, and no quality assessment (noQA). We assessed the impact of PQA, HRQA, and noQA on seizure detection performance by applying HR increase thresholds from 30% to 80% relative to the baseline. A permutation test with bootstrapping was applied for all statistical comparisons.

Results

We analyzed 741 NW alarms, comprising 135 true seizures and 606 non‐seizure events in 28 children (46% female; age: 9.1 ± 3.3 years). Averaged across all evaluated detection thresholds, both PQA and HRQA significantly improved seizure detection performance compared to noQA (p < 0.005). PQA led to mean increases of 0.248 in sensitivity, 0.025 in positive predictive value (PPV), and 0.082 in F1‐score. HRQA showed mean increases of 0.081 in sensitivity, 0.040 in PPV, and 0.056 in F1‐score. At clinically relevant sensitivity levels (0.8, 0.85, and 0.9), PQA showed a relative improvement over HRQA, with F1‐scores increasing by 4%–15% and PPV by 5%–19%.

Significance

Implementing PPG quality assessment—particularly pulse‐based—improves the performance of wearable HR‐based seizure detection.

Plain Language Summary

This study aimed to improve seizure detection in wearable devices by introducing a new method for assessing the quality of the heart rate (HR) data derived from a light sensor called photoplethysmography (PPG). By applying our method to data from children with epilepsy, we showed that discarding low‐quality PPG data before HR estimation leads to better seizure detection, allowing more seizures to be correctly identified. This improvement is important because it enhances seizure detection performance through a software‐only upgrade.

Pulse‐based photoplethysmography quality assessment improves wearable seizure detection performance.

Key points.

• Incorporating data quality assessment significantly enhances the performance of wearable HR‐based seizure detection.

• The morphology and amplitude consistency of the PPG pulse are robust indicators for quality assessment.

• Pulse quality assessment prior to HR estimation offers superior results for seizure detection compared to post‐hoc HR quality assessment.

1. INTRODUCTION

Given the unpredictable and potentially life‐threatening nature of seizures, ¹ there is a pressing need for reliable home‐based seizure detection and intervention systems. This is especially true for high‐risk convulsive seizures (e.g., focal to bilateral or generalized tonic–clonic seizures), particularly during sleep. ¹ , ² Wearables can help address this challenge as they enable continuous, unobtrusive vital data collection outside of clinical settings. Photoplethysmography (PPG) has emerged as an attractive modality due to its inherent noninvasiveness, low cost, and remarkable ease of integration into a wide array of wearable devices ³ offering an estimate of heart rate (HR), an important feature to detect seizures. ⁴

Nonetheless, the utility of wearable PPG in real‐world, ambulatory settings is often compromised by its inherent susceptibility to various confounding factors, such as loss of sensor‐skin contact, environmental noise, and body movements. ⁵ , ⁶ These disturbances introduce artifacts that directly corrupt the raw PPG waveform, ⁷ leading to the inaccurate extraction of crucial physiological parameters ⁸ and diminishing their value as reliable indicators for seizure detection, especially in the presence of body movements. ⁹

Although the problem of poor PPG data quality is well recognized in the context of seizure detection, ⁶ , ¹⁰ , ¹¹ its impact on detection performance has not been systematically studied. ¹² Here, we propose a pulse‐based quality assessment (PQA) method that evaluates the quality of each individual pulse wave based on its morphological characteristics and amplitude stability, enabling the exclusion of compromised PPG pulses before HR estimation. We assessed the validity of PQA by comparing its impact on HR‐based seizure detection performance with two alternatives: HR‐based quality assessment (HRQA), which scores each second of the HR estimates on a 0–100 scale, and noQA, where all HR estimates are used without quality filtering.

2. MATERIALS AND METHODS

2.1. Data description & selection

The data utilized in this study were collected as part of the PROMISE trial (NCT03909984), a long‐term, prospective, home‐based investigation to evaluate the performance of the NW device in detecting nocturnal major motor seizures in children with refractory epilepsy. ¹³ The cohort comprised 51 children, aged 4 to 16 years, recruited from three Dutch epilepsy centers, all of whom experienced at least one nocturnal major motor seizure per week. Participants wore the NW device on their upper arm, which continuously recorded 3D accelerometry (ACC) for movement and PPG for HR monitoring.

The NW combines HR and movement to detect seizures. HR is continuously updated from PPG intervals, while ACC captures motion and rhythmic movements; both are analyzed in parallel when HR quality is sufficient; otherwise, only movements are used. All alarms generated by the NW were annotated through blinded review of video recordings, which served as the gold standard. ¹³ Alarms corresponding to major motor seizures—defined as generalized tonic–clonic, focal to bilateral tonic–clonic, generalized tonic seizures lasting over 30 s, focal onset clonic, generalized onset, and “tonic–clonic–like” bilateral events without a classic focal‐to‐bilateral pattern—were classified as true alarms, while alarms triggered by other seizure types or non‐seizure‐related events were classified as false.

For each NW alarm, we selected a 10‐min PPG segment for analysis, spanning 5 min before and 5 min after the seizure onset for true, and alarm onset for false alarm. This window ensured that enough data was available for reliable HR baseline estimation. Although the NW device normally transmits continuous PPG data wirelessly for storage, brief connection losses sometimes interrupt this process. As a result, some alarms had incomplete PPG recordings. Although the gaps or missing samples did not impact NW's real‐time seizure detection performance, they could jeopardize the accuracy of the retrospective data analysis. We therefore applied a strict gap‐filtering criterion and only included segments with a fully stored, continuous 10‐min PPG signal.

2.2. The proposed PQA method

The proposed PQA method to assess the quality of individual PPG pulses is exemplified in Figure 1. PQA combines morphology‐based template matching with amplitude stability checks and assigns a binary label to each pulse, that is, 1 for good quality, 0 for poor quality.

FIGURE 1.

The overview of the proposed pulse‐based quality assessment (PQA). The method processes raw photoplethysmographic (PPG) signal through six sequential stages: (1) pre‐processing and pulse segmentation; (2) generation of a reference template from first 20s of the signal; (3) pulse‐by‐pulse morphological comparison against the template; (4) evaluation of amplitude consistency across pulses; (5) flagging individual PPG pulses as poor quality (red or black) or retained as high quality (blue); and (6) updating the reference template using recent high‐quality pulses to accommodate signal variations over time.

2.2.1. Stage 1: Pre‐processing and pulse segmentation

The raw PPG signal is first band‐pass filtered using a zero‐phase fourth‐order Butterworth filter with cut‐off frequencies of 0.2 and 5 Hz, to suppress low‐frequency drift and high‐frequency noise components that lie outside the frequency range of typical cardiac cycles. Next, PPG pulse segmentation is performed by identifying pulse peaks and onsets from the band‐pass filtered signal. Within 10‐s sliding windows, the signal is smoothed using a 100 ms moving average filter to reduce residual noise. ¹⁴ Peaks are then detected as local maxima, identified by zero‐crossings of the first derivative, refined using an adaptive 25% amplitude threshold, and validated by enforcing a minimum inter‐peak distance of 200 ms. ¹⁵ Finally, pulse onsets are identified as the global minima between two consecutive peaks in the band‐pass filtered signal.

2.2.2. Stage 2: Template generation

Template matching assesses the morphological similarity of individual PPG pulses to a predefined representative waveform, referred to as the reference template, in order to identify pulses that deviate from a typical clean pattern. ¹⁶ The reference template is generated from the first 20‐s segment of data. Each pulse is time‐normalized by linearly interpolating it to the median pulse length (in samples) to ensure consistent alignment. The time‐normalized pulses are then averaged pointwise to produce the final template, which represents a typical clean PPG pulse within the signal.

2.2.3. Stage 3: Template matching

Each PPG pulse is resampled to match the duration of the reference template. The Pearson correlation coefficient is then computed between the interpolated pulse and the reference template. If the correlation exceeds a predefined threshold, the pulse is considered morphologically valid; otherwise, it is rejected.

2.2.4. Stage 4: Amplitude consistency

Pulses that have passed the template matching stage are further evaluated for amplitude consistency. This is done to account for corrupted pulses that can have a similar morphology to the reference template but exhibit abnormally large amplitudes due to motion artifacts. ¹⁷ Each morphologically valid pulse's amplitude is compared to the local median amplitude of the last ten valid pulses. A pulse is considered valid only if its relative amplitude deviation from the local median is within a predefined threshold (e.g., ±50%).

2.2.5. Stage 5: Pulse flagging

The final binary quality flag for each pulse is set to 1 (good) if the pulse passes both the template matching and amplitude consistency checks; otherwise, it is labeled 0 (poor). This two‐stage procedure ensurse robust rejection of pulses affected by motion artifacts, clipping, or abnormal morphology.

2.2.6. Stage 6: Template updating

To account for slow physiological changes in PPG signal morphology (e.g., vasoconstriction), ¹⁸ the reference template is adaptively updated every 10 pulses. Within each 10‐pulse window, the pulse with the highest correlation to the current template—provided the correlation exceeded 0.9—is selected as the new template. If no pulse meets this criterion, the previous template is retained.

2.3. Setting PQA thresholds

PQA employs two thresholds that must be tuned prior to running: (1) the morphological threshold, which assesses the similarity (correlation) between each PPG pulse and a reference template, and (2) the amplitude consistency threshold, which evaluates the magnitude variation of each pulse relative to the preceding ten valid pulses. To tune these thresholds, we randomly selected ten PPG segments, each with a fixed duration of 10 min, from the total pool of analyzed data. These samples were chosen to manually regulate the PQA thresholds based on visual inspection. The resulting final thresholds were then fixed and applied uniformly across all data used in the subsequent analysis.

2.4. Methods for comparison

To evaluate the performance of PQA, we compared it against two alternatives: HRQA method developed for the NW, and a baseline approach (noQA) where all estimated HR values regardless of quality were used. HRQA evaluates each second of the HR estimates based on two main criteria: the stability of HR and the presence of motion. To assess stability, the current HR is compared to the median of the five most recent HR values, with a difference of less than 10% suggesting a stable rhythm that increases the score. Conversely, a difference exceeding 35% decreases the score, indicating potential signal corruption. HRQA also considers the number of motions, interpreting more than two zero‐crossings per second as prominent movements that may compromise the HR estimate. Based on these criteria, HRQA generates a score on a scale of 0–100, where HR values with scores below 80 discarded as unreliable. However, these criteria are contextual. When HRQA registers a period of low‐quality HR, it requires multiple consecutive reliable estimates before the score recovers fully. Thus, even if a new HR estimate is reliable, the quality score only increases gradually. As a result, it may take several seconds for HRQA to re‐engage.

2.5. Impact on seizure detection performance

We evaluated the impact of PQA, HRQA, and noQA on the performance of the HR‐based seizure detection method, where HR increase relative to the HR baseline was an indicator of seizure. ¹⁹ An initial baseline HR was established from the first 120 s of data. The method then iterated through HR data in nonoverlapping 12‐s segments. Within each segment, the average HR was computed, and a potential seizure event was detected if this average HR exceeded the current dynamic baseline by a relative increase (varied from 30% to 80%). To account for sustained physiological changes, ²⁰ if a detection occurred, the method also checked the subsequent window for a continued increase above the baseline (Figure 2). The baseline itself was dynamically updated at 12‐s intervals where no seizure was detected. This ensures the baseline is consistently updated throughout the data where no event is occurring.

FIGURE 2.

Illustration of the heart rate (HR)‐based seizure detection process, which evaluates relative increase (Th) in mean HR across two consecutive windows (HRw1, HRw2) compared to a baseline HR (HRb).

The evaluation of these quality assessment methods followed a three‐step process. First, we assessed the impact of both PQA and HRQA on seizure detection performance by comparing them to noQA. We designed this step to determine if data quality assessment, in general, has a positive influence on seizure detection performance. Second, we directly compared the performance of PQA and HRQA to determine which of the two methods was superior. As a final step, we compared F1‐score, positive predictive value (PPV), and false alarm rate (FAR) between PQA and HRQA for clinically relevant sensitivities from 0.8 and beyond. ¹³

2.5.1. Performance metrics

For each detection threshold and each quality assessment method, we calculated sensitivity, which is the proportion of true positives among all actual seizure events, and PPV, defined as the proportion of true positive detections among all positive detections. Given the typically imbalanced nature of seizure detection datasets—where seizure events are much less frequent than non‐seizure periods—we used the F1‐score, defined as the harmonic mean of PPV and sensitivity, as a single scalar measure of performance. For comparison at clinically relevant sensitivity levels, we also computed FAR, defined as the number of false detections divided by the total number of gap‐free 10‐min PPG segments containing an NW alarm.

2.5.2. Statistical analysis

We employed the permutation test with bootstrapping to compare the paired performance of HR‐based seizure detection between PQA versus noQA, HRQA versus noQA, and PQA versus HRQA. To account for the multiple comparisons performed, we applied the Benjamini–Hochberg procedure to control the false discovery rate (FDR).

3. RESULTS

3.1. PPG data selection

Of the 51 enrolled children, 19 were excluded because they did not experience major motor seizures. The remaining 32 children generated 1544 NW alarms. We excluded 803 alarms due to gaps or incomplete recordings. The final dataset consisted of 741 gap‐free PPG segments from 28 children (46% female; age: 9.1 ± 3.3 years), including 135 true seizures and 606 non‐seizure events detected by the NW.

3.2. PQA thresholds tuning

We tuned PQA thresholds for the identification of poor PPG pulses on a subset of 10 PPG measurements. Through visual inspection of the flagged pulses from these measurements and confirming accurate identification of corrupted pulses, we selected a threshold for template matching of 0.7 and a threshold for amplitude consistency of 70%. These settings correctly identified over 98% of the corrupted cycles in these signals. Figure 3 provides a visual illustration of PQA's operation, showcasing the raw PPG signal (Figure 3A) and the resulting HR estimates using PQA, HRQA, and noQA on a representative data segment (Figure 3B).

FIGURE 3.

An example of a PPG signal with the applied PQA and the corresponding estimated HR. (A) shows the raw PPG signal and flagged pulses, (B) HR estimates based on the no quality assessment (noQA), HR quality assessment (HRQA), and PQA.

3.3. Seizure detection performance

Figure 4 shows seizure detection performance using HR estimates from PQA, HRQA, and noQA across the range of detection thresholds for the relative HR increase.

FIGURE 4.

Comparison of HR‐based seizure detection using PQA (dark blue), noQA (gray), and HRQA (green), across relative HR increase thresholds from 30% to 80% from baseline. (A) sensitivity, (B) positive predictive value (PPV), and (C) F1‐score as functions of the detection threshold.

3.3.1. Performance improvement of PQA and HRQA over noQA

As shown in Table 1, incorporating either PQA or HRQA led to a significant improvement in all three performance metrics compared to noQA across all thresholds. The magnitude of these improvements, however, differed between the two quality assessment methods. Applying PQA led to a greater mean gain in sensitivity (∆ = 0.248) and F1‐score (∆ = 0.082) than the HRQA method (∆ = 0.081; and ∆ = 0.056, respectively). Conversely, employing HRQA produced a greater mean gain in PPV (∆ = 0.040) than PQA (∆ = 0.025). All of these differences were found to be statistically significant after controlling for the FDR.

TABLE 1.

Paired comparison of seizure detection performance metrics across all thresholds for pulse‐based quality assessment (PQA) and HR quality assessment (HRQA) versus no quality assessment (noQA).

Metric across thresholds	Method	Δ	0.95 CI	r	p‐value
Sensitivity	PQA	0.248	[0.207, 0.288]	0.87	<0.001
	HRQA	0.081	[0.057, 0.105]	0.65	<0.001
PPV	PQA	0.025	[0.008, 0.041]	0.48	<0.005
	HRQA	0.040	[0.025, 0.057]	0.46	<0.001
F1‐score	PQA	0.082	[0.051, 0.113]	0.48	<0.001
	HRQA	0.056	[0.036, 0.077]	0.49	<0.001

Note: ∆ is mean difference, 95% CI is 95% bootstrap confidence interval of the mean difference, r is effect size, p is p‐value from the permutation test, and PPV is positive predictive value.

3.3.2. Comparison between PQA and HRQA

Applying PQA led to a more robust and consistent sensitivity, maintaining levels above 0.9 across 17 thresholds, compared to only two for HRQA (Figure 4A). Across all thresholds, applying PQA yielded a higher sensitivity than HRQA (∆ = 0.167, p < 0.001, 95% CI: [0.148, 0.185], r = 0.87). In contrast, HRQA generally outperformed PQA in terms of PPV, achieving higher values across most thresholds, particularly between 30% and 60% (Figure 4B). When analyzed over all thresholds, HRQA maintained significantly higher PPV than PQA (∆ = 0.015, p < 0.001, 95% CI: [0.013, 0.018], r = 0.54). For the F1‐score, incorporating PQA led to superior performance, reaching a higher peak F1‐score of 0.48 compared to HRQA's 0.42 (Figure 4C). Notably, the maximal F1‐score for PQA occurred at a sensitivity of 0.71, substantially higher than HRQA's corresponding sensitivity of 0.54. When compared across relative HR thresholds, applying PQA significantly outperformed HRQA in F1‐score (∆ = 0.025, p < 0.001, 95% CI: [0.014, 0.036], r = 0.49). All differences remained statistically significant after FDR correction. When considering clinically relevant sensitivity levels (0.8, 0.85, and 0.9), applying PQA resulted in higher F1‐scores and PPV, and lower FAR, compared to HRQA (Table 2).

TABLE 2.

Comparative analysis of the F1‐score, PPV, and FAR for PQA and HRQA at three clinically relevant sensitivity levels for gap‐free 10‐min PPG segments containing an NW alarm. The improvement column highlights the percentage gain of applying PQA over HRQA.

Sensitivity	Metric	HRQA	PQA	Relative improvement (%)
0.8	F1‐score	0.411	0.440	7
	PPV	0.277	0.303	9
	FAR	0.380	0.334	12
0.85	F1‐score	0.409	0.425	4
	PPV	0.268	0.282	5
	FAR	0.422	0.394	7
0.9	F1‐score	0.350	0.405	15
	PPV	0.216	0.260	19
	FAR	0.596	0.465	22

4. DISCUSSION

We investigated whether incorporating PPG quality assessment improves wearable HR‐based seizure detection. Our study introduced a novel method, PQA (pulse‐based quality assessment), and compared it with two alternatives—HRQA (post‐hoc HR outlier rejection) and noQA (using all unfiltered HR estimates)—to evaluate their impact on seizure detection performance. The obtained results suggested that applying either PQA or HRQA led to a significant improvement in seizure detection performance compared to noQA.

PQA showed significantly better sensitivity and F1‐score compared to HRQA. HRQA, in contrast, yielded a generally higher PPV due to a reduction in false positives along with lower sensitivity. This behavior can be linked to the conservative nature of HRQA: when several consecutive HR estimates are marked unreliable, it takes time to re‐engage. Having fewer HR estimates available will lower both true and false positives, thereby resulting in an artificially inflated PPV. PQA retains more usable pulses, leading to higher sensitivity, at the cost of a slightly lower overall PPV. From a clinical perspective, preserving sensitivity is far more important than marginal gains in PPV. Crucially, when comparing the methods at clinically relevant sensitivity levels (0.8, 0.85, and 0.9), PQA achieved a higher F1‐score and PPV, as well as a lower FAR, compared to HRQA. This suggests that PQA is more robust, likely because it removes low‐quality PPG pulses before HR estimation rather than relying on post‐hoc HR outlier rejection. By addressing signal quality at the pulse level, PQA more effectively mitigates artifacts that often degrade wearable HR‐based seizure detection.

PQA differs from existing approaches, ⁶ , ¹⁰ , ¹¹ which score signal quality using fixed window lengths (e.g., 4 s), by evaluating quality pulse by pulse. Segment‐based methods may require discarding entire windows even when only a small portion is corrupted. In contrast, PQA targets individual pulse waveforms, excluding only the affected pulses. This is especially beneficial during periods of frequent motion artifacts, such as major motor seizures, where retaining even a small proportion of clean data (e.g., 20% of pulses) can help maintain detection accuracy and continuity. A further strength of this study is its use of data from a home‐based trial, where real‐world artifacts naturally occurred.

The improvements in detection performance achieved by incorporating PQA must be viewed from a clinical adoption and patient risk perspective. The primary clinical relevance lies not only in the magnitude of the effect but in the mechanism of its realization: performance enhancement was achieved entirely through improved PPG data quality assessment. This represents a substantial practical advantage, as it enables meaningful gains in detection accuracy without imposing any additional user burden, hardware costs, or changes to established clinical workflows. For children at high risk, even a marginal increase in the number of successfully detected nocturnal major motor seizures provides a tangible benefit, offering additional opportunities for caregiver intervention and potentially improving safety outcomes. Therefore, these improvements are clinically and practically meaningful as it leverages software innovation—enhanced signal quality management—to improve an existing, critical monitoring tool.

This paper has some limitations. The proposed method requires an initial 20‐s data segment to generate its first template. This introduces a cold start period during which pulse quality cannot be assessed. It also relies on the assumption that this initial segment is sufficiently clean to form a reliable representative template, which might not always be the case. Nonetheless, averaging across multiple cycles typically produces a robust and reliable template. Moreover, since the reference template is updated every 10 cycles, any initial noise in the first reference template may have only a short‐lived influence on the overall quality assessment. Another limitation is that PQA thresholds were tuned on a small data subset, which may introduce an optimistic bias. Future work should validate PQA on a larger, independent cohort to confirm its generalizability. However, the study's primary goal was to demonstrate the efficacy of this novel pulse‐based approach, and the observed improvements in seizure detection performance serve as strong evidence for its potential. Finally, our analysis focused exclusively on gap‐free events detected by the NW system. Because baseline data (data with no‐events) were excluded from this subset, the reported FAR is inherently higher than it would be in a continuous, unselected dataset. Nonetheless, this study is intended as a conceptual evaluation of how data quality assessment impacts seizure detection, rather than a clinical validation of a standalone detection algorithm.

While PQA showed promising results for improving HR‐based seizure detection, it is crucial to acknowledge the inherent limitations of PPG‐based HR as a single biomarker. During the ictal phase of major motor seizures, convulsive movements can obscure the pulse waveform, ⁶ making it impossible to determine HR. Adding other more robust modalities such as electrodermal activity or accelerometry could help to overcome this limitation. ¹ , ⁶ , ²¹ Some evidence also suggests that multimodal approaches improve detection performance. ¹ , ²¹ , ²² , ²³ Alternatively, other PPG features beyond HR could be applied to improve seizure detection performance. Morphological features, such as pulse amplitude, increasing slope, or the area under the curve, could be explored to identify early pre‐ictal patterns ⁹ , ²⁴ , ²⁵ that may be leveraged for proactive seizure detection.

AUTHOR CONTRIBUTIONS

Study Conception and Design: The PQA method was developed by M.SH, while the study's overall concept and framework were jointly conceived and designed by M.SH and R.D.T., who together refined the research objectives, hypotheses, and approach. Data Collection: A.T.B. was responsible for gathering the data required for the study, including designing and implementing the data collection process, ensuring accuracy, and maintaining the integrity of the dataset. Analysis and Interpretation of Results: The analysis and interpretation of the data were jointly carried out by all authors. This included statistical evaluations, synthesis of findings, and drawing meaningful conclusions in the context of the study's aims. Draft Manuscript Preparation: M.SH, A.T.B. and R.D.T. took the lead in drafting the manuscript, organizing the findings, and presenting them in a coherent narrative that aligned with the study objectives and methods. Review and Approval: All authors contributed to the critical review of the manuscript, ensuring the accuracy and completeness of the results and interpretations. Each author provided feedback and revisions, culminating in approval of the final version for publication.

FUNDING INFORMATION

This work was supported by EpilepsieNL (2023–04) and the Anna Teding van Berkhout Stichting. The funders had no role in the study design, data collection, analysis, interpretation, or decision to submit for publication.

CONFLICT OF INTEREST STATEMENT

Roland D. Thijs (R.D.T.) reports lecture and consultancy fees from Medtronic, UCB, Angelini Pharma, Theravarance, Zogenix, Novartis, LivAssured, and Arvelle, as well as grants from Medtronic and NewLife Wearables. None of the other authors have any conflicts of interest to disclose. The authors are part of the Dutch TeleEpilepsy Consortium. LivAssured, the company developing the NightWatch device, has an exclusive license to use the data for commercial purposes in exchange for a percentage of the revenue directed to the University Medical Center Utrecht and Kempenhaeghe. The Dutch Tele‐Epilepsy Consortium will receive research funds if this project generates profits. The authors declare no financial interests or shares in LivAssured and have not received, nor will they receive, any income or compensation from the NightWatch sales. LivAssured had no role in the study design, analysis, or decision to submit the manuscript.

ETHICS STATEMENT

The PROMISE trial was approved by the Medical Research Ethics Committee of the University Medical Center Utrecht.

ETHICAL PUBLICATION STATEMENT

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.

PATIENT CONSENT STATEMENT

Written informed consent was obtained from the participants or assent from parents or legal guardians.

CLINICAL TRIAL REGISTRATION

The PROMISE trial was registered at the Dutch Trial Registry (NCT03909984).

DATA AVAILABILITY STATEMENT

The data that support the findings of this study are available from the corresponding author upon reasonable request.