Scalp high‐frequency oscillations differentiate neonates with seizures from healthy neonates

Abstract

Objective

We aimed to investigate (1) whether an automated detector can capture scalp high‐frequency oscillations (HFO) in neonates and (2) whether scalp HFO rates can differentiate neonates with seizures from healthy neonates.

Methods

We considered 20 neonates with EEG‐confirmed seizures and four healthy neonates. We applied a previously validated automated HFO detector to determine scalp HFO rates in quiet sleep.

Results

Etiology in neonates with seizures included hypoxic‐ischemic encephalopathy in 11 cases, structural vascular lesions in 6, and genetic causes in 3. The HFO rates were significantly higher in neonates with seizures (0.098 ± 0.091 HFO/min) than in healthy neonates (0.038 ± 0.025 HFO/min; P = 0.02) with a Hedge's g value of 0.68 indicating a medium effect size. The HFO rate of 0.1 HFO/min/ch yielded the highest Youden index in discriminating neonates with seizures from healthy neonates. In neonates with seizures, etiology, status epilepticus, EEG background activity, and seizure patterns did not significantly impact HFO rates.

Significance

Neonatal scalp HFO can be detected automatically and differentiate neonates with seizures from healthy neonates. Our observations have significant implications for neuromonitoring in neonates. This is the first step in establishing neonatal HFO as a biomarker for neonatal seizures.

Key Points.

• Scalp HFO can be detected in neonates using an automated framework.

• Scalp HFO rates are higher in neonates with seizures compared to healthy neonates.

• Etiology, status epilepticus, EEG background activity, and seizure patterns do not impact scalp HFO rates in neonates with seizures.

• Our observations have significant implications for neuromonitoring in neonates.

1. INTRODUCTION

High‐frequency oscillations (HFO) in scalp EEG are a new and promising non‐invasive biomarker for the seizure‐generating brain. ¹ , ² Scalp HFO, originally associated with the seizure onset zone in pharmaco‐resistant focal epilepsy, ³ , ⁴ , ⁵ have been recently proposed as EEG biomarkers of (1) seizure risk in a predisposing condition, ⁶ , ⁷ (2) disease severity in various epilepsy syndromes, ⁸ , ⁹ , ¹⁰ and (3) treatment response in both focal and generalized epilepsies. ³ , ⁵ , ¹¹ , ¹² Most studies so far have been conducted in pediatric cohorts, ¹ driven by the higher incidence and severity of epilepsies in early life and benefiting from the higher rates (and, thus, the higher sensitivity) of scalp HFO in this age group. ¹³ Recently, scalp HFO have been described in the absence of epilepsy, ¹⁴ , ¹⁵ and their rates have been reportedly lower in healthy controls than in children with seizures. ¹⁰ , ¹⁶ Considering the wide availability of scalp EEG, scalp HFO may prove a key resource for evaluating children with high seizure risk, starting from the first days of life.

Seizures are more frequent in the neonatal period and often have grave consequences for affected infants. ¹⁷ , ¹⁸ While most neonatal seizures are acute symptomatic related to perinatal brain injury, 15% of neonates present neonatal epilepsy syndromes of structural or genetic origin, ¹⁹ and 18%‐33% of those with acute symptomatic seizures will develop post‐neonatal epilepsy. ²⁰ The chance of treatment success in neonatal seizures decreases with increasing seizure burden, independent of gestational age, etiology, or antiseizure medication (ASM), ²¹ underlining the importance of their early detection and appropriate intervention. While clinical features alone fail to predict neonatal seizures, ²² EEG features, such as background activity, have been more useful alone ²² or as contributors to multifactorial models. ²³ , ²⁴ However, reliable EEG biomarkers for this condition are yet lacking.

Although neonatal seizures offer a suitable paradigm for assessing the value of scalp HFO as a potential EEG biomarker in a vulnerable population, ¹ there is a dearth of data on this topic. This paucity may be attributed to the challenges of HFO detection in neonatal scalp EEG due to maturational changes through different gestational ages, ²⁵ the effect of ASM and hypothermia on the EEG, ²⁶ and the interference from equipment in the neonatal intensive care unit (NICU). ²⁷ So far, only two studies have investigated HFO in the scalp EEG of neonates with seizures, verifying that the neonatal brain can generate HFO, both studies drawing from manual HFO annotation. ²⁸ , ²⁹ However, this HFO detection method is subjective, limited to short time windows, and time‐consuming, and thus cannot be incorporated into clinical routine. ³⁰ , ³¹ Standardized and efficient approaches to scalp HFO detection are required before these can be used as biomarkers in at‐risk neonates.

In the present study, we aimed to investigate (1) whether a previously validated automated detector can capture scalp HFO in neonates and (2) whether scalp HFO rates can help to differentiate neonates with seizures from healthy neonates. Establishing scalp HFO as an EEG biomarker in neonates with seizures will add to our diagnostic repertoire and open new avenues for precision medicine in this vulnerable age group.

2. METHODS

2.1. Patient selection

We retrospectively identified from our institutional database at the University Children's Hospital Zurich preterm (<37 weeks of gestational age) and term neonates (≤30 days of age, corrected gestational age ≤44 weeks) who underwent scalp EEG between January 2021 and December 2022. For this study, we considered those neonates who fulfilled the following inclusion criteria: (1) neonates with EEG‐confirmed seizures or healthy neonates without seizures or any neurologic or systemic disease impacting the central nervous system (CNS), (2) scalp EEG with sampling frequency >1 kHz, (3) identifiable EEG epochs of quiet sleep, and (4) available informed consent. We considered the first available high sampling frequency scalp EEG for each neonate (1) with an overall good recording quality, (2) containing quiet sleep epochs, and (3) including ≥5 min of analyzable quiet sleep data. We collected and analyzed the following clinical features: demographics (sex, gestational age, postnatal age), neonatal seizure etiology, treatment (ASM, therapeutic hypothermia), presence of status epilepticus, and EEG features (background activity, seizure patterns). We classified neonatal seizure etiology based on the current framework for neonatal seizures and epilepsy syndromes ³² as hypoxic‐ischemic, structural vascular (including acute ischemic stroke, hemorrhage, and other vascular‐induced ischemia), structural due to brain malformation, and genetic.

The collection of patient data and their analysis were approved by and performed according to the guidelines and regulations of the local ethics committee (KEK‐ZH PB‐2021‐01245). All parents gave informed general consent to reuse clinical data for research.

2.2. Scalp EEG acquisition and data selection

Neonates underwent scalp video‐EEG with 12 electrodes placed according to the international 10–20 system, ²¹ depending on the infant's head circumference, at a 1024 Hz sampling rate, using the Micromed® EEG recording system (Mogliano Veneto, Treviso, Italy). Impedances were typically ≤5 kΩ. Extracerebral leads were used for respiratory and electrocardiogram recording and surface electromyography. According to our institutional protocol, the EEG recordings in neonates included a complete cycle of awake, quiet, and active sleep states. ²¹ If state changes were indistinguishable, the EEG recording lasted at least 1 h.

Experienced neurophysiologists marked sleep stages and seizure patterns according to the American Clinical Neurophysiology Society criteria. ²⁷ To avoid contamination by muscle artifacts, we considered only quiet sleep, clinically characterized by eye closure, absence of rapid eye movements, and scant body movements. Only interictal epochs were considered for analysis. In scalp EEG containing ictal discharges, we excluded from analysis (1) the full duration of the EEG seizure pattern and (2) additional segments 1 min before and 1 min after the EEG seizure pattern. Scalp EEG intervals with visible artifacts and channels with continuous interference were also excluded from the analysis.

The EEG background was scored ²¹ , ³³ as (1) normal, (2) mildly/moderately abnormal (excess sharp activity, absence or decreased frequency of normal patterns, excessively long low‐voltage periods or overall slightly decreased voltage, asymmetries in voltage or frequencies, asynchrony for age), or (3) severely abnormal (isoelectric or low‐voltage invariant activity, burst‐suppression pattern, permanent discontinuous activity. EEG seizure patterns are characterized by sudden, repetitive, evolving stereotyped waveforms with a beginning and end duration long enough to allow recognition of onset, evolution in frequency and morphology, and resolution of the abnormal discharge. ³²

2.3. Automated scalp HFO detection

Following a common approach in scalp HFO studies, ¹ we used bipolar channels to attain the highest possible signal resolution and, most importantly, to eliminate artifact contamination. We re‐referenced to a bipolar montage using all combinations of neighboring electrodes, ⁴ , ⁵ , ¹³ , ³⁴ thus obtaining 24 bipolar channels (Fp1‐Fp2, Fp1‐FT7, FT7‐TP7, TP7‐CP3, FC3‐FT7, TP7‐O1, Fp1‐FC3, FC3‐CP3, CP3‐O1, Fp2‐FT8, FT8‐TP8, TP8‐CP4, FC4‐FT8, TP8‐O2, Fp2‐FC4, FC4‐CP4, CP4‐O2, FC3‐FC4, FC4‐Cz, FC3‐Cz, CP3‐Cz, CP4‐Cz, CP3‐CP4, and O2‐O1). We segmented quiet sleep data into intervals with a maximum duration of 2 min and a minimum duration of 30 s, which were then considered for further analysis.

We detected HFO in the neonatal scalp EEG using a clinically validated, automated HFO detector, ³ , ⁴ , ⁵ , ¹³ , ³⁴ , ³⁵ previously utilized in pediatric cohorts, including very young infants in the first months of life. The detector operates in three stages ³⁵ and uses Stockwell's algorithm for the time‐frequency transform of the EEG signal. ³⁶ In Stage I, baseline detection, a baseline amplitude threshold is determined in artifact‐free intervals, selecting epochs of high Stockwell entropy in the ripple band (80–250 Hz). Events exceeding this threshold are marked as events of interest (EoI). In Stage II, HFO validation by Stockwell transform, the detector selects those EoI where a high‐frequency peak is isolated from the low‐frequency activity and marks them as HFO. This step reflects the assumption that HFO are brief events with a distinct high‐frequency contribution that stands out from the baseline. ³⁶ Finally, in Stage III, artifact rejection, the detector screens EoI selected in Stage II to eliminate other artifacts. Thus, the detector rejects all EoI with peak‐to‐peak amplitude ≥40 μV or signal‐to‐noise ratio <4.

We optimized the detector for application to neonatal recordings by integrating several artifact rejection steps aiming to evade “false ripples,” that is, artificial oscillations due to bandpass filtering (jump artifacts) or external interference (muscle/eye movement and respirator/pump artifacts). Thus, we used the feature ft_artifact_zvalue from the FieldTrip toolbox ³⁷ to calculate the z‐score of the EEG signal at specific frequency bands and identify artifacts by thresholding the accumulated z‐score. We thus rejected candidate HFO events co‐occurring with (1) jump artifacts identified by a z‐score in the wide band exceeding the median by five times the interquartile range or exceeding the absolute value of 15, and (2) muscle artifacts identified by a z‐score in the 110‐140 Hz band ³⁷ exceeding the median by 1.5 times the interquartile range.

We calculated the HFO rate for each channel by dividing the number of detected HFO by the duration of the analyzed EEG recording in minutes. We calculated the average HFO rate for each patient by dividing the number of detected HFO per minute by the number of analyzed bipolar channels. The average HFO rate was used for statistical comparisons across the patient groups. We calculated the characteristics of each detected HFO event: the root‐mean‐square value (RMS) of the bandpass filtered signal (80‐250 Hz), the duration, and the frequency (defined as the number of peaks divided by the duration of the event).

HFO detection and analysis were performed blinded to clinical features, and the results of HFO analysis were not considered for clinical decision‐making.

2.4. Statistics

We applied descriptive statistics to analyze the study cohort's clinical features. We reported distributions with mean and range or standard deviation (sd). We used Spearman's correlation to investigate the interrelations between HFO event characteristics. To compare the average HFO rates between neonates with seizures and healthy controls, we used Welch's t test and calculated Hedge's g for the effect size. To identify an HFO rate threshold differentiating neonates with seizures from healthy neonates, we calculated the maximum Youden index. To calculate specificity and sensitivity, we considered HFO rates higher than or equal to the HFO threshold in neonates with seizures as true positives and HFO rates lower than the HFO threshold in healthy controls as true negatives. We used one‐way ANOVA to investigate the effect of background activity on HFO rates, independent of etiology. Statistical analysis was performed with R version 4.2.3. Significance was established at P ≤ 0.05.

3. RESULTS

3.1. Patient characteristics

We considered 20 consecutive neonates with EEG‐confirmed seizures (14 male) and 4 healthy neonates (2 male) who fulfilled our inclusion criteria. The indication for performing EEG was suspicion of neonatal seizures in all cases. ³⁸ In healthy neonates, the suspicious episodes were eventually attributed to sleep myoclonia, jitteriness, or mild respiratory distress. 18 neonates were excluded from our study because their EEGs contained no quiet sleep epochs (n = 8), presented an overall poor recording quality (n = 9), and included <5 min of quiet sleep (n = 1; Figure 1). Twenty‐two of 24 neonates were full term, and 2 were preterm. The median postnatal age at EEG was 3.5 days (range 0‐33) in neonates with seizures and 4 days (range 0‐19) in healthy neonates.

FIGURE 1.

Patient selection flowchart.

In neonates with seizures, the etiology was hypoxic‐ischemic encephalopathy (HIE) in 11 cases, structural vascular in 6, and genetic in 3 (including a case of tuberous sclerosis). Six of 20 neonates with seizures presented with status epilepticus. Five neonates had seizure patterns during the analyzed EEG. The EEG background was normal in 6, mildly/moderately abnormal in 11, and severely abnormal in 3 cases. Fourteen of 20 neonates with seizures received ASM shortly before or during the EEG, and 3 underwent therapeutic hypothermia (Table 1). In healthy neonates, the EEG background was normal in all cases, and none received any ASM or CNS medication.

TABLE 1.

Clinical features of neonates with seizures (top panel) and healthy neonates (bottom panel), analyzed EEG duration, and results of HFO analysis.

Pat. Nr.	Sex	Gestational age (w + d)	Weight at birth (g)	Age at EEG (d)	Etiology	Status epilepticus	EEG background	EEG seizure pattern (s)	Therapeutic hypothermia at EEG	ASM at EEG	Analyzed EEG duration (min)	Average HFO rate (HFO/min/ch)
1	M	36 + 4	3000	7	HIE	No	Normal	No	Yes	LEV	7.4	0.20
2	M	36 + 5	2470	1	Genetic; POLG1 mutation	Yes	Severely abnormal	Yes	No	PB + LEV+PHT	39.8	0.12
3	M	37 + 0	2950	33	Genetic; TSC mutation	No	Mildly/moderately abnormal	No	No	No	17.2	0.03
4	M	39 + 0	2960	10	HIE	No	Mildly/moderately abnormal	No	No	No	18.5	0.13
5	F	39 + 0	3040	2	Structural vascular; MCA infarction	Yes	Mildly/moderately abnormal	Yes	No	PB + LEV	35.9	0.05
6	M	39 + 0	3300	3	HIE	No	Normal	No	No	PB	28.5	0.05
7	F	39 + 1	2930	1	HIE	No	Normal	No	No	LEV	18.0	0.23
8	F	39 + 2	2940	1	HIE	No	Normal	No	No	LEV	5.5	0.27
9	F	39 + 3	2790	7	Genetic; KCNT1 mutation	Yes	Mildly/moderately abnormal	No	No	LEV	12.8	0.27
10	M	39 + 3	3390	13	Structural vascular; occipital stroke	No	Mildly/moderately abnormal	No	No	PB + LEV	22.4	0
11	M	39 + 4	3000	4	HIE	No	Normal	No	No	No	13.1	0.14
12	M	40 + 0	3000	3	Structural vascular; MCA infarction	No	Mildly/moderately abnormal	No	No	PB + LEV	37.7	0.01
13	F	40 + 1	3500	3	HIE	No	Severely abnormal	Yes	Yes	PB + LEV+MI	34.9	0.11
14	M	40 + 2	3400	0	HIE	Yes	Severely abnormal	Yes	Yes	PB + MI	33.7	0
15	M	40 + 3	4160	17	HIE	Yes	Mildly/moderately abnormal	No	No	LEV+MI	52.2	0
16	M	40 + 4	2770	6	Structural vascular; MCA infarction	Yes	Mildly/moderately abnormal	no	No	PB + LEV+PHT + MI	14.8	0.01
17	F	40 + 4	3500	10	HIE	No	Normal	no	No	No	15.3	0.04
18	M	40 + 4	3250	2	HIE	No	Mildly/moderately abnormal	no	No	No	19.7	0.17
19	F	41 + 2	3670	4	Structural vascular; left frontal/parieto‐occipital stroke	No	Mildly/moderately abnormal	no	No	PB + LEV	30.9	0.10
20	M	41 + 5	3640	2	Structural vascular; IVH	No	Mildly/moderately abnormal	yes	No	No	15.3	0.02
21	M	38 + 3	2655	0	Healthy neonate	No	Normal	no	No	No	37.7	0.04
22	M	39 + 4	4370	5	Healthy neonate	No	Normal	no	No	No	18.1	0.03
23	M	40 + 4	2910	3	Healthy neonate	No	Normal	no	No	No	5.7	0.07
24	F	41 + 4	4260	19	Healthy neonate	No	Normal	no	No	No	23.7	0.01

Abbreviations: ASM, antiseizure medication; ch, channel; d, day; f, female; HIE, hypoxic‐ischemic encephalopathy; IVH, intraventricular hemorrhage; LEV, levetiracetam; m, male; MI, midazolam; nr, number; Pat, patient, PB, phenobarbital; PHT, phenytoin; w, week.

3.2. Scalp HFO can be automatically detected in the neonatal EEG

In neonates with seizures, the total duration of scalp EEG considered for analysis was 474 min with a median of 19 min per neonate (range 5‐52). In healthy neonates, the total duration of scalp EEG considered for analysis was 85 min, with a median of 21 min per neonate (range 6‐38). In 8 neonates, ≥30 min of scalp EEG were available for analysis; in 13 neonates, >10 min to <30 min were available; in 3 neonates, 5‐10 min were available. Our automated HFO detection is illustrated in Figure 2.

FIGURE 2.

Automated HFO detection. Two exemplary candidate events in a patient with tuberous sclerosis and neonatal seizures associated with a left parieto‐occipital leading tuber (patient 3), depicted in the wide band (top row), in the 80‐250 Hz frequency band (middle row), and the corresponding spectrogram (bottom row). A, This candidate event appeared as a blob on the spectrogram and was localized left parieto‐occipital (CP3‐O1 channel), matching the location of the epileptogenic lesion. The event was identified as an HFO. B, This candidate event appeared as a spanning over the ripple frequency band on the spectrogram and was localized right parieto‐occipital, contralateral to the epileptogenic lesion. The event corresponded to a jump artifact caused by ringing oscillations on the filtered signal and was rejected as an HFO and identified as an artifact.

We visualized the characteristics of the detected HFO events (n = 813) in neonates with seizures. The mean RMS amplitude of the HFO events was 3.00 ± 1.33 μV, the center frequency was 116.11 ± 26.35 Hz, and the mean duration was 34.33 ± 17.27 ms. We found no significant correlation between the center frequency of the HFO events and their RMS amplitude (Figure 3A; Spearman's correlation, r = 0.00, P = 0.91). Other correlations were highly significant: The longer the duration of an HFO, the higher its RMS amplitude (Figure 3B; Spearman's correlation, r = 0.24, P < 0.001) and the lower its center frequency (Figure 3C; Spearman's correlation, r = −0.38, P < 0.001).

FIGURE 3.

Characteristics of the HFO events detected in neonates with seizures. The colors of the heatmaps represent the percentage of events (n = 813) with specific values. A, We identified no correlation between the center frequency of the HFO events (116.11 ± 26.35 Hz) and their mean RMS amplitude (3.00 ± 1.33 μV). B, We identified a positive correlation between the duration of the HFO events (34.33 ± 17.27 ms) and their RMS amplitude. C, We identified a negative correlation between the frequency (116.11 ± 26.35 Hz) and duration (34.33 ± 17.27 ms) of the HFO events.

3.3. Scalp HFO rates are higher in neonates with seizures than in healthy neonates

The average HFO rates were significantly higher in neonates with seizures (n = 20, 0.098 ± 0.091 HFO/min) than in healthy neonates (n = 4, 0.038 ± 0.025 HFO/min; Welch's t test, P = 0.02), with a Hedge's g value of 0.68 indicating a medium effect size. Using a Youden index that identified a threshold value of 0.1 HFO/min/ch led to a specificity of 100% since the HFO rates of all healthy controls were below this threshold, but to a chance‐level sensitivity of 50%, since the HFO rates of only half of the neonates with seizures were above this threshold (Figure 4).

FIGURE 4.

Average scalp HFO rates are higher in neonates with seizures (n = 20) than healthy neonates (n = 4, Welch's t test, P = 0.02). The average HFO rate of each neonate with seizures is depicted by a blue circle, and the average HFO rate of each healthy neonate by a light orange circle, while large dark red circles depict the mean values of each subgroup. The violin plots illustrate the distribution of the HFO rates, while the boxplots illustrate the median and interquartile range.

3.4. Scalp HFO characteristics in neonates with seizures

Scalp HFO showed a focal/regional or bilateral/diffuse spatial distribution (Figure 5).

FIGURE 5.

Spatial distribution of scalp HFO rates on each bipolar channel interpolated on the skull surface for two patients with different etiologies. A, In a patient affected by hypoxic‐ischemic encephalopathy (patient 7), the HFO distribution was bilateral/diffuse, more pronounced over the anterior regions. B, In a patient affected by focal structural epilepsy associated with a leading tuber left parieto‐occipital (patient 3), the HFO distribution was focal/regional.

Among neonates with seizures, HFO rates did not differ (1) according to etiology: HIE (n = 11,0.12 ± 0.09 HFO/min/ch), structural (n = 6, 0.03 ± 0.04 HFO/min/ch), and genetic etiology (n = 3, 0.14 ± 0.12 HFO/min/ch, one‐way ANOVA, P = 0.10), (2) according to the presence of status epilepticus: with (n = 6, 0.08 ± 0.11 HFO/min/ch) vs without status epilepticus (n = 14, 0.11 ± 0.09 HFO/min/ch, Welch's t test, P = 0.53), (3) according to the EEG background activity: normal (n = 6, 0.16 ± 0.10 HFO/min/ch), mildly/moderately abnormal (n = 11, 0.07 ± 0.09 HFO/min/ch), severely abnormal (n = 3, 0.08 ± 0.07 HFO/min/ch) EEG background activity (one‐way ANOVA, P = 0.19), and (4) according to the presence of EEG seizure patterns: with (n = 5, 0.06 ± 0.05 HFO/min/ch) versus without EEG seizure patterns (n = 15, 0.11 ± 0.10 HFO/min/ch, Welch's t test, P = 0.18).

The average HFO rate in neonates with HIE did not differ significantly between the three neonates who underwent therapeutic hypothermia during the EEG (0.10 ± 0.10 HFO/min/ch) and the eight neonates who did not undergo therapeutic hypothermia during the EEG (0.13 ± 0.10 HFO/min/ch, Welch's t test, P = 0.73). Interestingly, HFO rates were above the HFO threshold, as determined by the Youden index, in 2 of 3 neonates who underwent hypothermia.

4. DISCUSSION

Our study is the first to demonstrate that scalp HFO in neonates can be automatically detected in neonatal EEG. The HFO rate was independent of etiology, status epilepticus, EEG background activity, and seizure patterns. As a major result, higher scalp HFO rates in our cohort differentiated neonates with seizures from healthy neonates, suggesting that HFO may indicate dysfunction of the neonatal brain.

4.1. Scalp HFO rates are higher in neonates with seizures than in healthy neonates

Scalp HFO rates in our study were overall higher in neonates with seizures, thus facilitating the crucial distinction between these two states. Our study is the first to include controls, directly comparing scalp HFO rates between neonates with seizures and healthy neonates and describing presumably physiological scalp HFO in the neonatal age group. Despite the relatively small size of our control group, due to the unlikely referral of healthy neonates for EEG recordings, a large effect size has underscored the significance of our findings. Comparisons of scalp HFO between children with seizures and healthy children have only been reported in the context of the infantile epileptic spasms syndrome ¹⁰ , ¹⁶ or of self‐limited epilepsy with centrotemporal spikes, ⁶ with considerably lower rates in neurologically normal children. Previous neonatal studies have considered no healthy neonates without seizures or CNS disease ²⁸ , ²⁹ and could thus draw no conclusions regarding the presence of physiological HFO in the healthy neonatal brain. ²⁸ Identifying the timing during brain maturation when pathological and physiological HFO can be generated, decoding their interrelations with brain development, and clarifying their clinical value is a direction well worth pursuing.

No effect of the underlying etiology on scalp HFO rates has been shown in our study, in line with past neonatal ²⁸ , ²⁹ and pediatric ⁴ , ¹² studies. This observation supports the notion that HFO constitute a universal biomarker for pathological brain activity, irrespective of the specific brain pathology. ²⁸ Furthermore, seizure frequency, particularly status epilepticus, had no significant impact on HFO rates in our study, in contrast to the observation of higher HFO rates at higher seizure frequency in pediatric cohorts. ³ , ⁵ Widely divergent methodologies prohibit comparing the impact of seizure patterns on HFO rates, as a previous neonatal study has focused on ictal segments to facilitate discriminating HFO events from artifacts, ²⁹ while another has focused on interictal segments with >2 h latency from seizures. ²⁸ Finally, EEG background activity did not impact HFO rates in our study, in contrast with the observation of higher scalp HFO rates with abnormal background activity in a neonatal study. ²⁸ However, any findings regarding the impact of seizure substrates, clinical, and EEG features on HFO rates should be viewed cautiously since the limited size of available cohorts so far prohibits definitive conclusions. ³⁹

4.2. Scalp HFO can be automatically detected in the neonatal EEG

Scalp HFO in our study have been captured by an automated detector optimized for application in the artifact‐ridden EEG recordings performed in the NICU. This novel algorithm‐driven approach designates a major breakthrough compared to the former expert‐driven approach of visual identification and manual annotation. ²⁸ , ²⁹ Automatic detection, previously deemed “undesirable” in a neonatal scalp HFO study ²⁹ and “unfeasible” in another ²⁸ due to excessive interference by high noise artifacts, ⁴⁰ offers a substantial benefit for clinical implementation. Integrating several artifact rejection steps in our detector has enabled evading possible sources of interference in the NICU environment ⁴⁰ and reaching meaningful results in a standardized and reproducible manner within a reasonable time frame. While the visual identification of HFO requires expert time and effort, the computational algorithm, once established, can be used by clinicians at any level. This approach paves the way for screening long‐term EEG recordings over critical periods in at‐risk neonates and allowing real‐time scalp HFO detection and, eventually, timely intervention in these neonates.

The reinforcing of artifact rejection in our dedicated neonatal HFO algorithm may have led to the exclusion of HFO events co‐occurring with artifacts ⁴¹ in our study. The lower HFO rates measured in neonates with seizures in our cohort compared to the HFO rates reported in previous pediatric and neonatal studies, including our studies performed with a similar setup, ³ , ⁵ , ¹³ , ²⁸ , ²⁹ , ³⁴ may be at least partly attributed to the stricter selection of HFO events after rigorous artifact removal from the signal. However, the disparities in HFO rates between our study and the two past neonatal studies may also be explained by the differences in the electrode array, ²⁸ since a lower number of electrodes, conform to the currently pertinent guidelines, ²⁷ entails a lower resolution both in HFO generating regions and in artifact polluted regions, hampering the detection of HFO events and their distinction from co‐occurring artifacts. The analysis of longer sleep segments from continuous EEG monitoring yields higher scalp HFO rates, as determined in our previous studies. ¹³ , ³⁴ However, brief EEG segments of 5‐10 min sufficed for meaningful results in one‐third of our neonatal cohort, supporting the feasibility of automated HFO detection in this setting.

4.3. Future directions

While neonatal EEG recordings are subject to interference from NICU equipment, EEG findings may be additionally confounded by the use of sedatives such as propofol in very sick neonates, ⁴² ASM for seizure control in neonates with seizures, ²¹ and therapeutic hypothermia in neonates with HIE that may suppress both EEG background activity and HFO rates. Although available studies have established no impact of these therapeutic interventions on scalp HFO, this issue will have to be investigated further in future studies. In addition, to determine the significance of physiological HFO in healthy neonates, as recently addressed in pediatric cohorts, ¹⁴ , ¹⁵ a much larger number of EEG recordings collected longitudinally from healthy neonates through different gestational ages is required.

Given our encouraging results, neonatal scalp HFO could be applied in large representative samples of neonates with seizures and normal neonates to confirm the reliability and generalizability of our findings and enable the delineation of their determinants. Longitudinal studies of neonates with seizures will address whether scalp HFO can predict the development of post‐neonatal epilepsy, in line with a recent work demonstrating that scalp HFO after a first unprovoked seizure can predict the development of pediatric epilepsy. ⁴³ The utility and validity of scalp HFO as an independent determinant of neonatal seizures and neonatal‐onset epilepsy will have to be assessed by multifactorial clinical prognosis scores integrating clinical, EEG data—such as EEG background activity and interictal spikes—and imaging data.

Considering that HFO detection in scalp EEG is affected by poor SNR and by the presence of artifacts, the implementation of recent and exciting approaches ⁴⁴ , ⁴⁵ aiming at the elimination of false positives and thus increasing HFO detection accuracy is bound to facilitate the detection and interpretation of neonatal scalp HFO. These approaches, drawing from intracranial human and animal data, will, however first have to be adapted and validated on scalp EEG. Furthermore, dedicated HFO detection devices ⁴⁶ may facilitate future long‐term scalp HFO monitoring.

Establishing scalp HFO as a key resource for assessing at‐risk neonates may pave the way to precision medicine approaches aiming to stop epileptogenesis in this vulnerable patient group.

5. CONCLUSION

Our study provides a robust and reliable framework for neonatal scalp HFO detection that may facilitate their implementation as an EEG biomarker for seizure risk, thus enabling early intervention in this vulnerable age group. Neonatal scalp HFO can be detected automatically, systematically, and reproducibly in carefully selected intervals, and scalp HFO rates can differentiate neonates with seizures from healthy neonates. Our observations have significant implications for neuromonitoring in at‐risk neonates and contribute to the growing body of evidence establishing scalp HFO as an epilepsy biomarker, particularly in early life.

AUTHOR CONTRIBUTIONS

GG, DC, and GR wrote the manuscript and prepared the figures and tables. GR designed the study. BA, GG, CD, and GR gathered and prepared the data. BA, GG, DC, TF, AR, and GR analyzed the data. All authors reviewed, edited, and approved the manuscript.

CONFLICT OF INTEREST STATEMENT

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

ETHICS 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.

Cserpan D, Guidi G, Alessandri B, Fedele T, Rüegger A, Pisani F, et al. Scalp high‐frequency oscillations differentiate neonates with seizures from healthy neonates. Epilepsia Open. 2023;8:1491–1502. 10.1002/epi4.12827

DATA AVAILABILITY STATEMENT

Data used in this study are available from the corresponding author upon reasonable request.