A multi‐feature method for real‐time seizure detection in pediatric intensive care unit

Abstract

Purpose

Continuous electroencephalogram (cEEG) monitoring is an important technique used in detecting electroclinical seizures in the pediatric intensive care unit (PICU). This study developed an artificial intelligence method for the real‐time automatic detection of seizures in the PICU.

Methods

We designed an artificial intelligence method to analyze EEG, electromyography (EMG), and electrocardiography (ECG) signals, detecting features from multiple dimensions, extracting candidate signal fragments in real time, and analyzing the relevant indicators of these fragments to determine whether they indicate electroclinical seizures. We tested the sensitivity and specificity of the detection system on patients with seizures who were hospitalized in the PICU of Peking University First Hospital and received cEEG monitoring.

Results

A total of 28 PICU patients were collected, including 17 boys and 11 girls, with a median age of 4.30 (1.02–7.00) years. Sixteen patients had convulsive status epilepticus, of which eight were generalized and eight were focal. Twelve patients had cluster seizures, of which seven were generalized and five were focal. A total of 218.73 h of EEG data were collected from all 28 EEG records. The electroencephalography physician annotated a total of 1561 seizures, whereas the algorithm detected a total of 1095 seizures. The overall detection sensitivity was 94%, and the overall false detection rate was 0.18 (0.03–0.28)/h. There was no statistical difference in the sensitivity and false detection rate between focal and generalized seizures.

Conclusion

The detection system has high sensitivity and specificity, suggesting great potential for future real‐time automatic detection of electroclinical seizures in the PICU.

Plain Language Summary

We developed a computer program that helps doctors quickly detect seizures in critically ill children by analyzing brain activity, muscle movements, and heart rate signals. The system was designed specifically for the challenging environment of pediatric intensive care units, where timely seizure detection is particularly important but often difficult. The program reliably identified seizures that showed both brain activity and physical signs, matching doctors' diagnoses. This technology marks an important step toward better monitoring and care for critically ill children with seizures.

Key points.

• This study developed and validated a real‐time seizure detection algorithm for critically ill pediatric patients in PICU settings.

• The multimodal algorithm integrates EEG, EMG, and ECG signals to automatically identify clinically relevant seizure events.

• With 94% overall sensitivity and a low false detection rate of 0.18/h, the system demonstrates high diagnostic accuracy in PICU settings.

• This automated approach enables earlier intervention and has strong potential to improve clinical outcomes in critically ill children.

1. BACKGROUND

Continuous electroencephalogram (cEEG) is currently the gold standard for seizure detection in critically ill children in the PICU. ¹ , ² Seizure burden is associated with worse outcomes, including increased mortality. ³ However, seizure detection requires a high degree of expertise and is time‐consuming. The majority of PICUs do not have enough electroencephalographers or EEG technologists available to continually interpret cEEG. Further, when available, cEEG is often only investigated for seizures twice daily, which can lead to delays in seizure identification. ⁴ Therefore, real‐time automatic seizure detection technology is necessary to reduce the workload of clinicians and improve the timeliness of detection. ⁵

The existing seizure detection technologies do not distinguish seizure waveform and the pseudo‐difference waveform well, with a high false detection rate. Although seizures in the PICU can be frequent and prolonged, there are limited real‐time seizure detection technologies that are specifically developed and validated for use in this environment. Previous studies have proposed various methods for automatic detection of seizures. One of the earliest seizure detection methods was proposed by Gotman, which divided EEG signals into half‐waves and extracted three features of average amplitude, duration, and coefficient variation for the detection of seizures. ⁶ The widely used Reveal algorithm has been reported to have a clinical sensitivity of 76% and a false positive rate of 0.11/h, ⁷ which results in excessive non‐seizure events being flagged and may contribute to alarm fatigue for clinicians. Additionally, some algorithms may also have a significant false negative rate, leading to true seizures being missed. Therefore, it is clear that highly sensitive and specific non‐patient‐specific seizure detection methods are needed clinically.

The aim of this study is to develop and evaluate a multimodal real‐time seizure detection algorithm using continuous EEG, ECG, and EMG recordings in the pediatric intensive care unit (PICU). We hypothesize that the integration of multimodal physiological signals will improve the accuracy and reduce the false alarm rate of seizure detection.

The basic innovation presented in this paper is the addition of machine learning (ML) capabilities to the system. We achieve this on two levels by collecting data in real time and on a deeper level, by detecting suspicious electroclinical EEG signals combined with ECG and EMG signals to determine a real seizure. We implemented a real‐time inspection system program where we did not need to provide training data sets prior to the initial operational state. Our system picks up its own seizure events and uses them to improve its performance. The novelty presented in this paper includes: (1) introducing and implementing an algorithm for real‐time detection of electroclinical seizure detection parameters; (2) comprehensive analysis of three different signals, including EEG, ECG, and myoelectric, has a good detection effect on different seizure types.

2. METHODS AND MATERIALS

2.1. Study population

This prospective research was conducted from January 2020 to April 2022 on patients who underwent cEEG monitoring in the PICU of Peking University First Hospital, China. This study was approved by the Ethics Committee of the Peking University First Hospital, China (No. 2021‐384).

2.2. Methods

The workflow of the automated seizure detection system in the PICU is shown in Figure 1A. Real‐time acquisition of EEG, ECG, and EMG signals from patients with seizures is input into the PICU seizure detection system. Preprocessing is applied independently to each signal, with the processed data shown in Figure 1B. The preprocessed data segments are then divided into fixed‐duration fragments. An initial screening of EEG segments is performed based on energy thresholds for high‐ and low‐frequency signals. The initially screened suspicious segments are shown in Figure 1C. Subsequently, feature extraction is performed on EEG, ECG, and EMG signals within suspicious segments. As illustrated in Figure 1D, the extracted features include EEG Time‐domain correlation coefficients (TDCC), ECG Heart rate (HR) and heart rate bursts (HRB), as well as EMG bursts and EMG bursts onset latency. Figure 1D demonstrates the intrinsic alignment of these features in a confirmed seizure, supporting their clinical relevance. Each suspected seizure segment is assigned a weighted score based on feature‐specific thresholds. Classification as an electroclinical seizure is determined by whether the aggregated score exceeds the predefined threshold. As shown in Figure 1E, the blue‐marked segment (scores below threshold) is classified as a non‐electroclinical event, while the red‐marked segment (scores exceeding threshold) is confirmed as a true electroclinical seizure.

FIGURE 1.

Workflow of the Automated Seizure Detection System (A) System flowchart for real‐time seizure detection in the PICU. (B) Preprocessed multimodal bio signals: EEG (1.6–70 Hz), ECG (5.3–40 Hz), and EMG (53–120 Hz). (C) Representative EEG segments flagged during initial screening (high/low‐frequency energy thresholds). (D) Temporally aligned seizure‐associated features. (E) Classification outcomes: Red segments (score ≥ threshold) = confirmed seizures; blue segments (score < threshold) = non‐electroclinicalevents.

2.2.1. Signal acquisition and preprocessing‌

Signals were acquired through clinical‐grade amplifiers operating at a sampling rate of 1000 Hz, capturing a 19‐channel EEG configured via the 10–20 system montage alongside ECG and EMG recordings. To optimize computational efficiency while preserving critical physiological information, all signals were down‐sampled to 200 Hz. Powerline interference was mitigated through notch filtering targeting 50 Hz and its harmonics. Subsequent bandpass filtering was tailored to each modality: EEG signals were filtered to a range of 1.6–70 Hz to attenuate baseline drift and high‐frequency artifacts, ECG signals were filtered to a range of 5.3–40 Hz to enhance cardiac waveform clarity, and EMG signals were filtered to a range of 53–120 Hz to isolate myoelectric activity. Global re‐referencing of EEG data was achieved by subtracting the mean signal across all channels from each electrode's raw voltage, resulting in the denoized re‐referenced EEG signal S(t), which minimizes common‐mode noise and enhances spatial resolution.

2.2.2. Suspicious segment detection‌

Suspicious segment detection was implemented through a sliding window protocol utilizing 20‐s epochs with 1‐s steps. Candidate segments were identified based on EEG patterns demonstrating either excessively high‐frequency oscillations relative to baseline or pathologically dominant low‐frequency activity. Given the primacy of abnormal EEG discharges as the definitive biomarker for electroclinical seizures, with ECG and EMG serving ancillary diagnostic roles, the classification framework prioritized EEG criteria: only segments surpassing predefined EEG abnormality thresholds were preliminarily flagged as suspected seizures. Secondary validation was then performed through systematic analysis of ECG rhythm irregularities and EMG burst patterns.

A segment is classified as a suspected seizure episode only if it satisfies two stringently defined electrophysiological criteria: (1) the amplitude of the 50 Hz‐denoized EEG signal S(t) must stabilize within the 300–800 μV range and (2) the line‐length metric $L_l$ calculated using Equation 1 must exceed 8 times the median value derived from 20‐min artifact‐free interictal baseline recording. This dual‐threshold framework leverages amplitude filtering to exclude extreme artifacts while employing line‐length analysis to capture sustained electrophysiological evolution patterns characteristic of epileptiform activity. The 8× multiplier statistically corresponds to 99.7% confidence intervals (3σ) when calibrated against interictal reference recordings, ensuring specificity against transient physiological fluctuations.

$$L_l=\frac{1}{N_{\mathit{ll}}}\sum _{t=1}^{N_{\mathit{ll}}-1}\mathrm{abs}\left[S\left(t-1\right)-S\left(t\right)\right]$$ (1)

where $S\left(t\right)$ denotes the 50 Hz‐denoized electrophysiological signal, represents the time sampling point, $N_{\mathit{ll}}$ denoting the computational window width.

2.2.3. Feature extraction

Temporal Correlation Coefficient (TMCC) Calculation

Neuronal hyper synchronization during electroclinical seizures manifests as abrupt increases in cross‐channel temporal correlations, establishing this metric as a critical biomarker for seizure detection. To quantify this phenomenon, we compute pairwise cross‐electrode Pearson correlation coefficients across 17 EEG channels (excluding frontal polar electrodes FP1/FP2 to minimize ocular artifact contamination). For each candidate segment, the correlation coefficient matrix is calculated through Formula 2.

$$R_{\mathit{ij}}=\frac{C_{\mathit{ij}}}{\sqrt{C_{\mathit{ii}}}\times \sqrt{C_{\mathit{jj}}}}$$ (2)

where $R_{\mathit{ij}}$ represents the temporal correlation matrix, $C_{\mathit{ij}}$ represents elements of the voltage covariance matrix. The segment's correlation coefficient is calculated by averaging the upper triangular elements of $R_{\mathit{ij}}$ through Formula 3

$$\text{corr}=\sum _{i=0}^n\sum _{j=i+1}^n{R}_{\mathit{ij}}$$ (3)

where n represents the matrix dimension (17 × 17). A seizure detection threshold is established at five times the median correlation value from baseline data, effectively distinguishing pathological synchronization (typically >7 times baseline) from normal interictal fluctuations (1–3 times baseline).

ECG feature extraction

Electroclinical seizures are frequently associated with transient alterations in cardiac autonomic regulation, characterized by abrupt deviations in HR and heart rate variability (HRV). A 20‐minute artifact‐free ECG recording establishes baseline thresholds through Formula 4

$$\text{thresh}=0.8*\text{mean}*\mathrm{std}$$ (4)

where mean and std represent the mean and standard deviation of the line‐length transformed baseline signal, respectively. For suspicious segments, R‐wave peaks are identified via a modified Pan‐Tompkins algorithm with adaptive thresholding, and instantaneous HR is computed from inter‐peak intervals (IPIs). HRV is derived as the difference between successive HR values. Outliers exceeding five times the median absolute deviation are excluded to suppress transient artifacts (e.g., R‐wave misdetection or motion‐induced signal corruption). The denoized HR/HRV time series is further smoothed using a 5‐beat Savitzky–Golay filter to attenuate high‐frequency noise while preserving temporal dynamics. HRB detection utilizes the Cumulative Sum (CUSUM) algorithm, where statistically derived thresholds determine significant HRB. Following grid search optimization, our cardiac signal detection thresholds were finalized as follows: The HR anomaly threshold was set at 120 bpm, requiring HR exceeding this value for ≥10% of the candidate EEG epoch duration to trigger detection. For HRV, the threshold was defined as 1.3× the median baseline HRV (calculated from historical non‐seizure segments); any HRV measurement surpassing this value was classified as an HRV burst event.

EMG feature extraction

Electroclinical seizures often manifest muscle clonic activity, characterized by abrupt EMG amplitude surges that typically lag behind EEG synchrony abnormalities by 0.5–3 s. This temporal dissociation motivates using both EMG bursts and EMG burst onset latency as seizure detection biomarkers. The analytical pipeline comprises three hierarchically processed stages: (1) Signal Conditioning‌ ‐ raw EMG undergoes noise reduction via a 200 ms moving average filter to suppress high‐frequency artifacts; (2) Feature Engineering‌ ‐ RMS envelope extraction (500 ms window, 80% overlap) converts the smoothed signal into energy‐modulated trajectories; (3) Bursts‐Delay Analysis‌ ‐ EMG activation onsets are detected by the CUSUM algorithm, and temporal delays are calculated between EMG bursts and EEG signal mutation time. Following grid search optimization, the EMG detection parameters were configured as follows: The EMG burst threshold was set at 150 μV, with any amplitude after the CUSUM algorithm exceeding this value classified as a detected EMG burst. Concurrently, the EMG delay criterion was defined as 0.5 s, meaning an EMG burst is considered temporally correlated if it occurs within 0.5 s after the onset of abnormal EEG activity.

2.2.4. Seizure verification via multimodal feature scoring

Seizure verification is implemented through a multimodal feature scoring protocol that classifies suspected electroclinical episodes by integrating five physiological biomarkers: TMCC, HR, HRV, EMG bursts, and EMG bursts onset latency. A weighted scoring mechanism assigns threshold‐derived points: +4 for TMCC exceeding predefined thresholds, +1 for sustained HR elevation >120 bpm exceeding 10% of the analysis window, +4 for detected HRB, +2 for detected EMG bursts, and +1 for detected EMG bursts onset latency. Seizure confirmation is triggered when cumulative scores reach ≥6 points. Threshold parameters and weighting coefficients were optimized via grid search optimization across predefined parameter ranges. Validation studies confirmed system robustness, demonstrating <5% variation in sensitivity and false alarm rate under ±20% threshold perturbations, thereby ensuring stable performance across heterogeneous clinical datasets.

2.2.5. Multimodal seizure detection framework

This multimodal framework enables real‐time electroclinical seizure monitoring in PICUs through synchronized analysis of multichannel EEG, ECG, and EMG signals processed in 20‐s sliding windows. During real‐time monitoring, physiological features are continuously extracted and evaluated against the optimized detection thresholds. When one or more features exceed their respective thresholds, this threshold activation triggers the assignment of feature‐specific scores according to our predefined weighting system. Finally, through decision fusion, these scores are aggregated across modalities, and any signal segment achieving a cumulative score ≥6 points is classified as a seizure event. This integrated process enables robust multimodal interpretation of ambiguous electrophysiological patterns.

2.3. Algorithm evaluation

The algorithm was evaluated based on the following parameters: true positive (TP) indicated EEG seizures labeled by the electroencephalographer and detected by the algorithm, false negative (FN) indicated EEG seizures labeled by clinicians but not detected by the algorithm; sensitivity (SEN) referred to the ability of the algorithm to detect seizures, calculated using Formula 5.

$$\mathrm{SEN}=\mathrm{TP}/\left(\mathrm{TP}+\mathrm{FN}\right)$$ (5)

and specificity (SPE) was the false‐positive alarm rate (FAR) per hour. In the following, all of the FAR values mentioned refer to the false alarm rate per hour.

2.4. Statistical analysis

Statistical analysis was performed using SPSS v26.0 and GraphPad Prism v9.0. The Kolmogorov–Smirnov test was used to check the normal distribution of the data, whereas an F‐test was used to check the homogeneity of variances. Data were presented as mean ± standard deviation (SD) conforming to a normal distribution; otherwise, they were expressed as median (25th–75th percentile). For two or three groups of normally distributed data, a two‐tailed t test or one‐way ANOVA was used; otherwise, the Mann–Whitney U and Kruskal–Wallis tests were performed. p < 0.05 was considered to be statistically significant, and ns means no statistical significance.

3. RESULTS

3.1. Demographic characteristics of the study population

A total of 28 patients were enrolled in the population, 11 (39.29%) girls and 17 (60.71%) boys. The median age of the enrolled population was 4.29 (0.17–17.42) years. According to the etiologies of seizures in our cohort, according to the 2017 ILAE classification system, ⁸ specifically, among our patients, 14 were classified as having genetic etiologies, 4 as infectious, 3 as immune, 2 as structural, and 5 cases remained of unknown etiology. All patients received cEEG with a total duration of EEG data of 218.73 h and a median duration of 4.78 (3.09–10.43) h per patient. The seizure events were identified and marked by a professional electroencephalographer. Patients were classified according to the seizure type observed during the monitoring period: patients with only convulsive status epilepticus (CSE) episodes and patients with only epileptic seizures (without SE). In this study, only seizures with both EEG and electroclinical seizures were included in the reference standard. Nonconvulsive, EEG‐only (subclinical) seizures were not annotated or analyzed (Table 1).

TABLE 1.

Clinical information of 28 patients.

Items	CSE	Seizures	Total (%)
Number/n
N	16	12	28
Sex/n (%)
Male	10 (62.50)	7 (58.33)	17 (60.71)
Female	6 (37.50)	5 (41.67)	11 (39.29)
Age/years (min‐max)	4.87 (0.25–17.42)	2.92 (0.17–9.00)	4.29 (0.17–17.42)
Etiology 8 /n (%)
Structural etiology	2 (12.50)	0	2 (7.14)
Genetic etiology	9 (56.25)	5 (41.67)	14 (50.00)
Infectious etiology	2 (12.50)	2 (16.67)	4 (14.28)
Immune etiology	2 (12.50)	1 (8.33)	3 (10.71)
Unknown etiology	1 (6.25)	4 (33.3)	5 (17.24)
Seizure types 9 /n (%)
Focal
Atonic	1 (6.25)	0	1 (3.57)
Clonic	6 (37.50)	4 (33.30)	10 (35.71)
Epileptic spasms	1 (6.25)	0	1 (3.57)
Myoclonic	1 (6.25)	1 (8.33)	2 (7.14)
Generalized
Tonic–clonic	2 (12.50)	1 (8.33)	3 (10.71)
Clonic	5 (31.25)	4 (33.30)	9 (32.14)
Tonic	3 (18.75)	1 (8.33)	4 (14.28)
Myoclonic	3 (18.75)	3 (25.00)	6 (21.43)
Seizure onset localization/n (%)
Frontal lobe	1 (6.25)	1 (8.33)	2 (7.14)
Temporal lobe	1 (6.25)	2 (16.67)	3 (10.71)
Occipital lobe	3 (18.75)	0	3 (10.71)
Multifocal lobe	6 (37.50)	2 (16.67)	8 (28.58)
Generalized	5 (31.25)	7 (58.33)	12 (42.86)
Anti‐seizure medications a
0	2 (12.50)	0	2 (7.14)
1	1 (6.25)	1 (8.33)	2 (7.14)
2	3 (18.75)	3 (25.00)	6 (21.43)
3	5 (31.25)	6 (50.00)	11 (39.29)
≥4	5 (31.25)	2 (16.67)	7 (25.00)
Narcotics b
0	1 (6.25)	12	13 (46.43)
1	7 (43.75)	/	7 (25.00)
2	6 (37.50)	/	6 (21.42)
≥3	2 (12.50)	/	2 (7.15)
Data duration/h
Total	148.68	70.05	218.73
Median (25th–75th)	7.03 (3.89–15.29)	3.47 (2.59–6.80)	4.78 (3.09–10.43)

^a Anti‐seizure medications include: valproic acid, levetiracetam, perampanel, clobazam, baclofen, topiramate, oxcarbazepine, clonazepam, nitrazepam, lacosamide, lamotrigine, zonisamide, et al.

^b Narcotics include: midazolam, diazepam, propofol, fentanyl, et al.

3.2. Algorithm evaluation

We analyzed 218.73 h of EEG data recorded in 28 patients, and the median duration was 4.78 (3.09–10.43) h. A total of 1561 seizures were labeled by the electroencephalographer, whereas a total of 1095 seizures were detected by the algorithm, with an average sensitivity of 94% and an average FAR of 0.18/h (Table 2).

TABLE 2.

Seizures and algorithm evaluation of 28 Patients.

Patient	EEG data duration/h	Number of electroencephalographer detected seizures	Number of algorithm‐detected seizures	Sensitivity	FDR/h
1	27.73	86	35	1.00	0.02
2	6.66	14	37	1.00	0.18
3	5.88	219	3	1.00	0.00
4	15.68	114	62	0.98	0.04
5	7.40	98	33	0.75	0.26
6	17.50	308	150	1.00	0.25
7	14.12	183	36	0.94	0.18
8	15.73	72	31	1.00	0.11
9	5.59	8	42	1.00	0.53
10	11.25	104	105	1.00	0.33
11	1.98	23	12	1.00	0.02
12	7.95	23	19	1.00	0.23
13	1.97	2	6	1.00	0.38
14	3.88	1	25	1.00	0.43
15	1.40	36	4	1.00	0.00
16	3.95	2	3	1.00	0.16
17	3.22	5	5	1.00	0.09
18	3.75	40	35	1.00	0.10
19	27.36	80	284	1.00	0.20
20	7.78	56	87	1.00	0.06
21	1.97	2	3	1.00	0.19
22	7.82	7	41	1.00	0.42
23	3.05	17	9	1.00	0.41
24	3.37	4	4	1.00	0.13
25	3.56	46	4	1.00	0.00
26	3.86	6	5	0.00	0.01
27	1.87	3	7	0.67	0.02
28	2.43	2	8	1.00	0.28
Total	218.73	1561	1095	0.94	0.18

3.3. Algorithm evaluation of different seizure types

In this study, we further analyzed the algorithm evaluation for focal and generalized seizures. The overall sensitivity for focal seizure detection was 0.97 ± 0.09, and the overall FAR was 0.11 (0.04–0.20)/h. The overall sensitivity for generalized seizure detection was 0.92 ± 0.26, and the total FAR was 0.19 (0.05–0.27)/h. There was no statistical difference in the sensitivity and false detection rate of the algorithm in detecting seizure types (both p > 0.05; Table 3).

TABLE 3.

Seizures and algorithm evaluation of different seizure types.

Seizure type	Focal seizure			Generalized			Compare
	Seizures	Convulsive status epilepticus	Total	Seizures	Convulsive status epilepticus	Total	Statistical value	p
n	5	8	13	7	8	15
EEG data duration/h	48.05	89.46	137.51	22.00	59.22	81.22
Number of electroencephalographer detected seizures	151	502	653	117	791	908
Number of algorithm‐detected seizures	424	261	685	68	342	410
Sensitivity	0.93 ± 0.15	0.99 ± 0.02	0.97 ± 0.09	0.86 ± 0.38	0.97 ± 0.09	0.92 ± 0.26	0.678	0.148
FAR/h	0.09 (0.04–0.31)	0.14 (0.03–0.33)	0.11 (0.04–0.20)	0.13 (0.01–0.28)	0.24 (0.04–0.31)	0.19 (0.05–0.27)	0.207	0.836

4. DISCUSSION

In this study, we designed and evaluated an algorithm that can be used to detect real‐time electroclinical seizures automatically in the PICU. Acquiring features of EEG, EMG, and ECG signals in multiple dimensions, extracting candidate signal segments in real time, and analyzing correlation indicators of candidate signal fragments enabled us to determine whether candidate signal fragments indicate seizures. Continuous seizure detection is necessary to assist pediatricians in providing a clinical diagnosis basis to prevent serious seizure complications. Our results suggest that this approach has potential for future application in real‐time seizure detection and notification in the PICU. ¹⁰

Because seizures are caused by a variety of neurological and systemic diseases, the use of an automatic seizure detection system is more likely to result in low sensitivity, a high false detection rate, and limited clinical application in PICU. ¹¹ It is currently a challenge to develop effective algorithms for automatic electroclinical detection. The algorithm is required to be able to detect various episodic EEG patterns and to identify organized EEG rhythm patterns caused by sleep, narcotic effects, and various encephalopathies, as well as repetitive, organized, or rhythmic patterns caused artificially. There were few studies on algorithms applied in the PICU. Chris et al. ¹² reported a novel ICU automated seizure detection algorithm that used signal amplitude to detect significant changes in amplitude, detected rhythmic changes based on the basis of repetitive signal patterns, and calculated the maximum signal frequency and amplitude in the high‐frequency range. The average sensitivity of the algorithm was 90.4%, and the average FAR was 0.066/h. Moreover, Herta et al. ¹³ used the NeuroTrend algorithm to screen video EEG data with frequencies greater than 4 Hz in 68 patients hospitalized in the neurological ICU. Based on the ACNS standardized critical care EEG terminology analysis, the algorithm had an overall detection sensitivity of 94%, of which the sensitivity to periodic discharges was 80%, the sensitivity to rhythmic delta waves was 82%, and the overall specificity was 67%.

Signal data commonly used for seizure detection or prediction systems and algorithms include EEG, EMG, ECG, accelerometer (ACM), electrical skin activity, and optical volumetric tracings. The current methods for seizure detection rely on several of these signals or various combinations. Seizures, particularly generalized tonic–clonic seizures, are frequently accompanied by significant alterations in cardiac autonomic function, such as sudden increases in heart rate and changes in HRV. These cardiac changes may precede or coincide with electroclinical seizure onset, providing an additional physiological marker for seizure detection. ¹⁴ Many seizures, especially motor seizures, are characterized by sudden bursts of muscle activity, which can be captured as abrupt increases in EMG amplitude. EMG is particularly valuable for distinguishing true motor seizures from artifacts and for detecting ictal events with prominent muscle involvement, even when EEG changes may be subtle or masked by movement artifacts. ¹⁵ Research has indicated the different effects of sensitivity and specificity of using single and multiple signals to detect seizures. ¹⁶ Single signal data such as ECG ¹⁷ , ¹⁸ and EEG ¹⁹ , ²⁰ , ²¹ have a sensitivity of 33.2–100% and FAR of 0.096–14.8/day. Multiple signal combinations, such as ECG combined with ACM, ²¹ , ²² and EEG combined with ECG, ²³ have a sensitivity of 51–100% and FAR of 0.12–17.7/day. The artificial intelligence detection method in this study used a combination of EEG, EMG, and ECG signal data, instead of a single EEG‐based algorithm. This resulted in superior feature classification accuracy of suspicious fragments and more specific detection of electroclinical seizures. There have been reports indicating that if acceleration detection and increased HR in electromyography can explain the increase in specificity threshold, then these can be marked as electroclinical seizures. ²² To the best of our knowledge, B. Olmi's study proposed a multi‐feature ECG‐based NSD system to investigate the usefulness of HRV analysis to detect neonatal seizures for the first time, ²⁴ the proposed system gave results that sensitivity was 47%, and specificity was 67%. ²⁴ Hence, the implementation of a comprehensive detection method utilizing multiple signal data and non‐brain physiological signals exhibits significant potential, particularly for non‐motor seizures and seizures accompanied by autonomic nervous changes. In the present study, the data from 28 patients with seizures were analyzed, and the overall average sensitivity was 94%, with an average false detection rate of 0.18/h. These metrics surpass the capabilities of existing PICU‐oriented automated seizure detection systems. ²⁵ , ²⁶ The tri‐modal fusion architecture significantly reduces false positive alerts and enhances sensitivity compared to conventional EEG‐only approaches, thereby establishing the clinical efficacy of cross‐domain physiological feature integration for discriminating ictal events from artifact‐prone critical care environments. ²⁷

There are certain differences in the classification of different types of electroclinical seizures in studies, which are generally divided into different seizure types and seizure origins. For instance, some studies divide recorded electroclinical seizures into convulsive and nonconvulsive types, ¹⁷ , ²⁸ whereas others categorize them as focal or generalized seizures, and even further into specific brain regions (such as temporal, frontal, or occipital lobe epilepsy). ²² , ²⁹ A recent study summarized the sensitivity range of detecting generalized tonic–clonic seizures as 83.64–100%, with a FAR range of 0–1.8/day. ¹⁶ Most studies analyzed motor and non‐motor electroclinical seizures, with an overall sensitivity range of 32–100% and a FAR range of 0–50/day. ¹⁹ No significant differences were observed between different detection algorithms, and there were no clear differences in features and algorithm selection in different research environments. In this study, both generalized and focal seizures were analyzed simultaneously, with an average sensitivity of 97% and FAR of 0.11 (0.04–0.20)/h for focal seizures, and an average sensitivity of 92% and FAR of 0.19 (0.05–0.27)/h for generalized seizures. Similar to previous research results, this algorithm indicated no statistical difference in sensitivity and detection error rates between different types of seizures.

A major limitation of this study is that only electroclinical seizures—those with both clinical and EEG manifestations—were included in the analysis, while subclinical (EEG‐only) seizures, which are common and clinically important in critically ill pediatric patients, were not systematically annotated, detected, or evaluated. As a result, our algorithm was trained and validated specifically for the detection of electroclinical seizures, and its ability to identify the full spectrum of seizure burden, particularly subclinical events or those occurring after electroclinical uncoupling, remains untested. In addition, the relatively small sample size, the lack of detailed analysis of seizure type heterogeneity, and the absence of algorithm optimization for distinguishing among different seizure types further limit the generalizability and practical validation of our findings. Future studies with larger cohorts, systematic annotation of all seizure types including nonconvulsive EEG‐only seizures and algorithmic adaptation for comprehensive seizure detection are needed to fully assess and expand the clinical utility of our approach in the PICU setting.

5. CONCLUSION

We have explored the value of an artificial intelligence system for the automatic detection of electroclinical seizures, suitable for real‐time monitoring of seizures in the PICU. Compared with previous detection algorithms, the seizure detection algorithm used in this study combines different features from EEG, ECG, and EMG, exhibiting considerable sensitivity and specificity. Its potential for future application in the automatic detection of seizures in the PICU is significant.

AUTHOR CONTRIBUTIONS

ST, DJ, and ZT contributed equally to this manuscript, and they were co‐first authors. ST contributed to the study design, collection, evaluation of the data, and drafting and writing the manuscript. DJ and ZT contributed to the study design, collection, evaluation of the data, and drafting and writing the manuscript. GQ and ZQ contributed to the collection, evaluation, and interpretation of the data. LYQ and YYX contributed to the algorithm calculations, and corrections., HB, WNN, and JYW contributed to the algorithm updates. WY contributed to revising the manuscript and approving the final version.

FUNDING INFORMATION

No funding.

CONFLICT OF INTEREST STATEMENT

None of the authors has any conflict of interest to disclose. 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.

ETHICS STATEMENT

The studies involving human participants were reviewed and approved by the Medical Ethics Committee of Peking University First Hospital [No. 2021‐384].

INFORMED CONSENT STATEMENT

The written informed consent form has been obtained from the patients by the parents and published this paper.

Sang T, Deng J, Zhao T, Zhang Q, Guan Q, Lei Y, et al. A multi‐feature method for real‐time seizure detection in pediatric intensive care unit. Epilepsia Open. 2026;11:112–122. 10.1002/epi4.70171

DATA AVAILABILITY STATEMENT

The original contributions presented in the study are included in the article and supplementary material; further inquiries can be directed to the corresponding authors.