Introduction

Preterm birth is associated with an increased risk of neurodevelopmental disorders,1,2,3,4,5,6 with susceptibility rising as gestational age at birth decreases.7 Early identification of potential atypical brain development in preterm infants can significantly improve individualized neonatal care by facilitating timely interventions. Therefore, the identification of reliable signatures of cerebral maturation goes beyond theoretical and fundamental research and has critical practical implications in the clinic, notably for at-risk infants.

Recent studies on the development of brain activity have utilized semi-quantitative and quantitative analyses of EEG to characterize spontaneous endogenous neural activity,8,9,10,11,12,13,14,15,16,17 using features such as relative power at different frequency bands9,10 and density and duration of bursts of activity and the intervals between them.11 This evidence reflects the fast evolution of EEG with increasing gestational age. The rapid structural and functional neurodevelopment during the third trimester of gestation18,19 is orchestrated not only by spontaneous neural activity20,21,22,23 but also by sensory-driven neural activity,24,25 with ‘richer’ processing of exogenous information as the brain matures.26,27 Therefore, a comprehensive assessment of early neurodevelopment needs to incorporate both spontaneous neural activity at rest and the reactivity of developing neural networks to exogenous stimuli. This is particularly important during this period of heightened plasticity, when the encoding of external information undergoes rapid developmental changes. Brain responsiveness to exogenous stimulation has proven to be a robust measure of complex brain states in adults.28 However, assessment of cerebral maturation is currently primarily based on spontaneous neural activity.

Previous studies have demonstrated the maturation of neural tracking of auditory rhythmic sequences in the delta frequency range26 as well as the increase in the amplitude of evoked neural activity in response to tactile stimulation.29 Although preterm maturational patterns in neural encoding of exogenous and notably auditory stimulation are reported,27,30,31 high inter-individual variability of response characteristics limits their reliability as markers of cerebral maturation. Building on the formulation of the perturbational complexity index (PCI) — originally developed to assess the distributed cortical response to transcranial magnetic stimulation for evaluating levels of consciousness32,33 — we designed a measure that would estimate the global effects of auditory stimulation on the preterm brain during routine EEG examinations. The hypothesis is that this measure enables the evaluation of how distributed cortical networks encode exogenous perturbation. This approach is grounded in the principle that the newborn brain’s ability to process exogenous stimulation relies on its capacity to integrate information across distributed cortical circuits. As cortical auditory dorsal and ventral pathways develop34 and neural interactions mature,35,36,37 the encoding of auditory input evolves.38,39 This maturation reflects increasing functional integration and functional differentiation in thalamocortical networks — key aspects of brain complexity40 — and therefore should manifest in richer neural responses to stimulation with increasing gestational age at birth.26,27

Classic assessment of cerebral maturation is based on the evaluation of the ongoing neural activity independent of stimulation. During the third trimester of gestation, brain activity is predominantly ‘discontinuous’, where high amplitude slow electrical bursts with nested faster oscillations interrupt low-voltage ‘silence’ intervals. The dynamics of these bursts of oscillations have been shown to be predictive of neurodevelopmental outcome in extremely premature newborns.16 The development of the interplay between excitation and inhibition in the cortical pathways leads to the emergence of rapid oscillations independent of exogenous stimulation. This would, in turn, impact the spectral content of EEG signals. In order to assess the emergence and evolution of neural interactions35,41 and oscillatory regimes, we quantified the aperiodic component of neural activity, linked to inhibitory-excitatory interactions42,43 and shown to represent systematic neurodevelopmental trends.35,44,45,46,47

We propose an integrative approach to predict cerebral age by combining endogenous and exogenous dimensions of neural activity, using high-density EEG data from 74 premature neonates, born between 27 and 35 weeks of gestational age (wGA) and tested within their first week of life during rest and auditory stimulation. We evaluated the spatiotemporal complexity of the neural response to auditory stimulation with increasing gestational age. By integrating this measure with spontaneous neural activity, our approach aims to capture complementary aspects of cerebral maturation and to explore whether the inclusion of sensory information processing can enhance the estimation of brain age during this critical period of neurodevelopment.

Materials and methods

Participants

We revisited 74 high-density EEG data (33 male) previously acquired in our studies on auditory rhythm encoding during the last trimester26,48 with a mean gestational age at birth of 31.67 ± 2.03 weeks (age range 27–35 wGA). Detailed information about the birth conditions is provided in Table S1. All EEGs were recorded during sleep about one week after birth at the neonatal intensive care unit of the Amiens University Hospital, France. This hospital serves as a regional reference center for neonatal care, and premature infants born there may come from both urban and rural areas across the departments of Somme, Aisne, and L’Oise. The neonates included in the study had normal birth weight, size, and head circumference for their gestational age. They also had normal auditory (otoacoustic emission and Auditory Brainstem Response) and clinical (Denver Developmental Screening) assessments at the equivalent age of term. None of the infants was considered at risk based on these assessments. One or both parents were informed about the study and provided their written informed consent. The ethics committee CPP Ouest I approved the study (ID-RCB: 2019-A01534-53).

Experimental procedure

Figure 1 summarizes the methodology and analysis flow of our cross-sectional study, dividing it into two complementary aspects of neonatal brain activity: auditory-evoked EEG and spontaneous EEG. Each experimental session began with a 20-min silent period to record the neonate’s spontaneous neural activity at rest. Following this resting period, auditory rhythmic patterns were presented to the sleeping infants to assess brain responses to external sensory input. Auditory-evoked EEG responses were analyzed to quantify spatiotemporal complexity using the stimulation complexity index (SCI; see also the flowchart in Fig. S1 for analysis steps), while spontaneous EEG bursts were characterized to extract inter-burst intervals (IBIs) and compute 1/f exponent, hereafter referred to as the linear slope (LS; see also Fig. S3). The extracted features were then fed into different regression models for cerebral age prediction. The following sections describe the methods used to analyze and extract various features from EEG data, as well as the development and evaluation of different models for cerebral age prediction.

Fig. 1: Visual overview of the study.
Fig. 1: Visual overview of the study.The alternative text for this image may have been generated using AI.
Full size image

HD EEG recordings were obtained cross-sectionally from premature newborns during auditory stimulation, to assess exogenous neural responses, and during the resting state, to analyze spontaneous neural activity. The exogenous feature extraction block (top) examined brain responses to auditory stimuli, quantifying spatiotemporal complexity with SCI. The endogenous feature extraction block (bottom) quantified intrinsic brain activity, where bursts of neural signals were detected to measure IBIs and to calculate the LS from the EEG power spectrum. The obtained features (SCI, IBI, and LS) were incorporated into different regression models to predict cerebral age, followed by the evaluation and comparison of their performance.

Data acquisition and preparation

EEG signals were recorded with either a 124-channel or 64-channel HydroCel Geodesic Sensor Net (Electrical Geodesics, Inc., Eugene, OR), selected based on the neonate’s head circumference, and amplified using the Electrical Geodesic NetAmps 200 system. Signals were digitized at a 1000 Hz sampling rate, with Cz as the reference electrode, and processed in Net Station software (version 5). The EEG data were analyzed in MATLAB (R2022b, The MathWorks) using FieldTrip (version 20240307),49 EEGLAB (version 2023.0),50 and custom MATLAB scripts. A two-pass 1–18 Hz finite impulse response filter was applied to remove artifacts, followed by down-sampling to 500 Hz. To standardize electrode configurations, 124-channel data were spatially down-sampled to a 64-electrode montage. Electrodes located in the outer ring were excluded from further analyses due to poor signal-to-noise ratio.

Auditory stimuli

We evaluated the neural response to single-tone (vibraphone, piano, guitar, and percussion) stimulation in a 600 ms post-stimulation window delivered through a speaker at 65 dB SPL, which was located at the feet of the neonates (see Supplementary information for more details).26,48 This was carried out by revisiting recordings where premature newborns were exposed to an auditory rhythmic sequence consisting of six events in the pattern [tone, rest (silence), tone, tone, tone, rest], with each event lasting 300 ms, and repeated without any silence in between sequences.26,48,51,52 This resulted in the first tone being preceded and followed by a tone at 600 ms intervals. In this study, we focused specifically on the 600 ms window preceding the first tone. Additional details regarding the auditory stimuli and experimental procedures, which are not central to the main objectives of this study, can be found in our previous publications.26,48,53

Quantifying the stimulation complexity index

EEG signals were prepared for extracting the SCI. First, data were epoched from the onset of the first tone in 600 ms windows. Then, individual channels were excluded if, in the epoched intervals, (1) the absolute amplitude exceeded 50 μV, (2) the standard deviation of amplitude exceeded 25 μV within the moving window of 200 ms, with a 50% overlap, or (3) a local amplitude jump exceeding 30 μV happened between ten consecutive time points. Finally, channels rejected in over 70% of epochs were excluded from the entire session, epochs with more than 15% rejected channels were discarded, and the EEG signals were re-referenced to the average of all electrodes. The same procedure was applied to the EEG data recorded during the resting state, which served as the pre-stimulus window. All the remaining electrodes for each participant were included in calculating the SCI. An average of 638.12 ± 184.37 trials per participant were retained after artifact rejection (61.94% ± 19.81% of the original trials). Each epoch was baseline-corrected to the average of the entire epoch, and the average response over each electrode was calculated as the average across all the remaining epochs.

The spatiotemporal complexity of EEG responses to auditory stimuli was quantified using PCI state-transition (PCIst),54 which measures the number of state transitions (NST) present in the principal components of the signal, reflecting both the spatial differentiation-integration and temporal dynamics of the brain’s response. This measure, which was originally introduced to evaluate consciousness through characterizing the neural response to transcranial magnetic stimulation33 at the cortical source level, was later modified for sensor-level quantification of neural spatiotemporal dynamics in response to transcranial magnetic stimulation.54 Thus, we developed an index analogous to the PCIst, but applied to sensory stimulation rather than magnetic stimulation to perturb the cortex and quantify how the cortical networks encode auditory stimulation. All parameters and functions were kept identical to those developed by Comolatti et al.54 Briefly, starting from average epochs in response to stimulation, singular value decomposition was performed in order to effectively reduce the spatial dimension of the EEG. Through singular value decomposition, the principal components that explained at least 99% of the variance present in the response were kept for further analyses (with an average of 2.87 components retained after dimensionality reduction). We then quantified to what extent the number of state transitions during the response differed from baseline for each component. The procedure was as follows: First, amplitude distances (pointwise differences) between all time points were computed separately for the baseline and response periods. Next, a binarization threshold was set to maximize the weighted difference between the number of state transitions in the response and those in the baseline (ΔNST). The final measure, SCI, was then calculated as the sum of these significant state transitions (ΔNSTn) across all selected principal components (n = 1 … NC) of the evoked signal (\({{{\rm{SCI}}}}=\,\mathop{\sum }_{n=1}^{{{{\rm{N}}}}{{{\rm{c}}}}}\Delta {{{\rm{NST}}}}\)n), indicating the spatiotemporal complexity of EEG responses to auditory stimuli. These steps are summarized in Fig. S1. SCI is higher when a response shows multiple, linearly independent components (indicating spatial differentiation), each of which contributes significant state transitions (reflecting temporal complexity).

Characterizing the spontaneous neural activity

This analysis was performed over the rest of the EEG. To prepare the signals for evaluation of bursts of spontaneous activity, data were first segmented into 2 s epochs specifically for preprocessing. Individual channels were excluded if (1) their absolute amplitude exceeded 300 μV or (2) their average absolute amplitude exceeded four standard deviations from the mean. Channels rejected in over 70% of trials were excluded for the entire session, and trials with more than 15% rejected channels were discarded. Finally, the data were re-referenced to the average of all electrodes. After preprocessing, resting-state EEG recordings retained an average of 13.72 ± 5.12 min (87.38% ± 5.47% of the whole recording) per participant for further analysis.

To detect and quantify the properties of bursts, we used a detection algorithm similar to Chini et al.35. EEG signals were averaged separately across electrodes corresponding to three frontal, parietal, and bilateral temporal regions (Fig. S2). The z-scored average signal from each region was then processed using a moving average filter with a 500 ms window, followed by the application of a hysteresis threshold — a two-level thresholding method in which a high threshold is used to identify the core of an event, and a lower threshold is used to extend the event to include neighboring points above this lower level. Burst activities were initially detected as oscillatory peaks exceeding a relative (high) threshold of 2 standard deviations and lasting for more than 500 ms. These periods were then extended to include all neighboring time points that exceeded a lower threshold of 1 standard deviation. This relative thresholding approach ensures adaptability to a wide range of signals, from the highly discontinuous brain activity observed in younger infants to the more continuous activity of older infants. Neighboring bursts with inter-burst intervals shorter than 1 s were merged, and burst periods lasting less than 1 s were discarded. Finally, the duration of bursts and inter-burst intervals (IBIs) was extracted.

For calculating the EEG power spectrum, all the detected burst intervals were concatenated. The power spectrum was computed using the FOOOF package,55 implemented in FieldTrip,49 over 10 s windows with 50% overlap, using the multi-taper method with a 2 Hz smoothing window. The spectrum was first calculated at each electrode location and then averaged across electrodes belonging to the three frontal, parietal, and temporal regions. LS corresponding to the aperiodic component of neural activity was extracted on the 1–18 Hz frequency range (Fig. 1, bottom block, and Fig. S3). For visual evaluation, the power spectrum of each infant was normalized by the area under the curve.

Cerebral age prediction model

The relationship between gestational age and each EEG-derived feature was first analyzed using Pearson correlation analysis to determine the strength and direction of associations (normality was tested using the Lilliefors test).56 This analysis provided initial insights into how each feature (SCI, IBI, and LS) varied with gestational age. Next, we developed and evaluated seven linear regression models, each incorporating different combinations of these features. These included single-feature models (SCI, IBI, and LS), as well as models combining multiple features in different manners: one endogenous marker and the exogenous marker (SCI-IBI and SCI-LS), two endogenous markers (IBI-LS), and all three features combined (SCI-IBI-LS). To evaluate the predictive accuracy of these models, we employed LOOCV. This approach ensured that the models were trained on data from all but one participant and tested on the not seen participant. Repeating this process for all infants provided a robust assessment of model performance. The performance metrics included R², which represents the proportion of variance explained by the model, the mean absolute error (MAE), and root mean square error (RMSE), which measure the average prediction error in days. Higher R² and lower RMSE/MAE values indicate better predictive accuracy and model fit.

The statistical significance of all LOOCV results was evaluated using a permutation test. For each model, the gestational age values were randomly permuted across subjects 10,000 times, and linear regression models were fit to these randomized age values. A p-value was calculated as the proportion of permutations in which the R² value exceeded the observed R² value, divided by the total number of permutations. Only models with p-values < 0.05 were considered reliable.

Statistical tests were conducted to compare the performance of the models and assess the significance of differences between them. Paired t-tests were performed to compare RMSE values across models, assessing whether the differences in prediction errors were statistically significant. To account for multiple comparisons, we applied the Benjamini-Hochberg False Discovery Rate correction (FDR)57 to adjust the p-values across all models. An adjusted p-value < 0.05 was considered statistically significant.

Results

Evolution of brain complexity during the third trimester of gestation in response to auditory stimulation

To estimate the global cortical processing of auditory stimulation on the preterm brain, we evaluated the spatiotemporal complexity of the neural response in a short window following stimulation. SCI evolution showed a developmental trend, manifesting in significant positive correlation with increasing gestational age at birth (Pearson correlation, r = 0.64, p < 0.001; Fig. 2a), reflecting cerebral maturation toward more complex spatiotemporal dynamics in neural responses to stimulation. Figure 2b reveals the emergence of response in the butterfly of auditory-evoked potentials to auditory stimulation in two premature neonates at 28 and 36 wGA, with distinct spatial patterns, resulting in more spatiotemporal diversity of the neural response with increasing gestational age at birth.

Fig. 2: Quantification of exogenous and endogenous EEG characteristics for the prediction of cerebral maturation.
Fig. 2: Quantification of exogenous and endogenous EEG characteristics for the prediction of cerebral maturation.The alternative text for this image may have been generated using AI.
Full size image

a Scatter plot shows the positive correlation between SCI and gestational age (Pearson correlation, r = 0.64, p < 0.001), indicating that SCI increases with age. Each gray dot corresponds to one participant, with the green and yellow markers corresponding to the two infants presented at (b) illustrating their alignment with the overall trend. b EEG butterfly of responses to auditory stimuli and the spatial distribution at different time points for two representative premature newborns at 28 and 36 wGA. Heatmaps display EEG activity across different brain regions over time. c Scatter plot illustrating the negative correlation between IBI and age, reflecting increased continuity in neural activity with maturation (Pearson correlation, r = −0.63, p < 0.001). d Normalized log-log EEG power spectra for all infants, demonstrating an increase in relatively high-frequency content, leading to a flattening of the spectra in older infants. The scatter plot (right) shows a significant negative correlation between LS and age (Pearson correlation, r = −0.55, p < 0.001). e EEG traces showing bursts and IBIs for the same infants at 28 and 36 wGA. Shorter IBIs with an increase in the relatively high-frequency content are evident in the older infant. f Comparison of the best-performing model (SCI-IBI-LS) and other models for predicting cerebral age evaluated using LOOCV. Stars indicate models that are significantly different from the best-performing model. g Scatter plots show actual versus predicted cerebral age for individual feature models (SCI, IBI, and LS) and the best-performing model (SCI-IBI-LS). The solid black line indicates the ideal prediction line (y = x), where predicted and actual ages align. Each point represents one participant. Vertical lines at each point show the confidence intervals for the predictions, representing the uncertainty associated with each estimate. The shaded area around the prediction line marks a ±2-week zone. The best-performing model (SCI-IBI-LS) demonstrates a closer alignment of predicted ages with the ideal prediction line and tighter clustering within the ±2-week margin, indicating higher accuracy and reliability.

Quantification of endogenous EEG characteristics during the third trimester of gestation

To quantify cerebral maturation based on spontaneous endogenous activity, we computed the mean IBI and the LS of the EEG spectrum corresponding to bursts of activity over frontal, parietal, and temporal regions. Initial evaluation showed that the LS of the EEG power spectrum did not show a developmental trend over the temporal regions (Fig. S4). Therefore, these regions were excluded from further evaluation of cerebral maturation, which was hereafter carried out over the average of frontal and parietal regions (see Supplementary information Fig. S4 for analyses separately conducted over frontal and parietal regions). Statistical analysis at the population level revealed a systematic decrease in IBI with increasing gestational age at birth (Pearson correlation, r = −0.63, p < 0.001, Fig. 2c). Superposition of normalized log-log power spectra corresponding to the whole population (Fig. 2d, left) showed a systematic evolution, with an increase in the relatively high-frequency content, resulting in flattening of the spectrum. The LS of the EEG power spectrum also showed a clear developmental trend manifesting in a systematic decrease with increasing gestational age at birth (Fig. 2d, right, Pearson correlation, r = −0.55, p < 0.001). As shown from the two representative examples (Fig. 2e), the patterns of burst and IBI activity differed between younger and older infants, with ‘richer’ bursts and shorter IBI in older premature newborns.

Prediction of brain maturation using a multimodal regression model integrating brain complexity and spontaneous EEG characteristics

With developmental trends revealed in both exogenously induced and endogenous neural activity, we next aimed to create a comprehensive model for cerebral age prediction, integrating these diverse yet tangled neural markers. Toward this, we evaluated seven linear regression models to predict cerebral age at the time of the recording. Table 1 presents the performance of the models in descending order of predictive accuracy, from the best-performing to the worst-performing model.

Table 1 Performance metrics for different models predicting cerebral age.
Full size table

The best-performing model was SCI-IBI-LS, which integrated all three features and achieved an R² of 0.67, an RMSE of 8.05, and an MAE of 6.72. This model explained 67% of the variance in cerebral age with the lowest prediction error, making it the most accurate and robust approach. Models that incorporated at least one endogenous feature (IBI or LS) along with SCI performed close to the best-performing model (SCI-IBI: R² = 0.64, RMSE = 8.36, MAE = 7.2; SCI-LS: R² = 0.62, RMSE = 8.6, MAE = 6.93). SCI, which used only the exogenous feature (brain complexity), achieved an R² of 0.56, an RMSE of 9.36, and an MAE of 7.69. While less accurate than the multimodal models, SCI alone exhibited better performance compared to endogenous-based models (IBI: R² = 0.38, RMSE = 11.2, MAE = 9.07; LS: R² = 0.27, RMSE = 12.04, MAE = 10.06). The predictive performance of all models was significantly different from chance, as confirmed by a permutation test (all p-values < 0.001).

We conducted statistical tests to compare each model with the SCI-IBI-LS model, which incorporated all three features, to investigate whether the observed differences in performance were significant. Paired t-tests were conducted to compare the prediction errors of different models obtained from leave-one-out cross-validation (LOOCV). The results (Fig. 2f) showed that the SCI, IBI, LS, and IBI-LS models had significantly higher prediction errors compared to SCI-IBI-LS (SCI vs. SCI-IBI-LS: t = 2.45, p = 0.043; IBI vs. SCI-IBI-LS: t = 3.04, p = 0.009; LS vs. SCI-IBI-LS: t = 4.28, p < 0.001; IBI-LS vs. SCI-IBI-LS: t = 2.38, p = 0.003; all p-values were FDR-corrected), indicating that these models were less accurate in predicting the cerebral age. However, the SCI-IBI and SCI-LS models, although higher in RMSE and lower in R2, did not show a significant difference in prediction error compared to SCI-IBI-LS (SCI-IBI vs. SCI-IBI-LS: t = 1.70, p = 0.11; SCI-LS vs. SCI-IBI-LS: t = 0.57, p = 0.56; all p-values were FDR-corrected). Figure 2g displays individual-level predictions of cerebral age for models relying on a single feature (SCI, IBI, or LS) and for the SCI-IBI-LS model. The alignment between predicted and actual cerebral age reflects the model’s ability to capture developmental trends reliably across the entire cerebral age range, with most predictions falling within the ±2-week error margin. The SCI-IBI-LS model distributed more narrowly around the identity line, whereas other models showed greater variability beyond the shaded area. These results suggest that while individual features provide useful information, their isolated use limits the model’s ability to generalize across the population.

Discussion

We used a measure of complexity (stimulation complexity index, SCI) for assessing cerebral maturation in premature newborns, leveraging the spatiotemporal complexity of neural response to single-tone auditory stimulation. This approach captured how the developing brain processes exogenous auditory inputs in distributed cortical networks. We demonstrated that the SCI evolves with cerebral maturation, showing increasing temporal complexity and spatial richness in neural responses. Crucially, our findings underscore the necessity of integrating both endogenous and exogenous dimensions for a more comprehensive evaluation of neurodevelopment in premature newborns. More precisely, while predictive models based on only spontaneous neural activity (IBI, LS, and IBI-LS) provided acceptable performance in estimating the cerebral age of premature newborns, incorporating the brain’s response to exogenous stimuli significantly improved prediction accuracy compared to models that rely solely on spontaneous neural activity. Interestingly, two-feature models that included one endogenous and one exogenous feature (SCI-IBI and SCI-LS) performed comparably to the three-feature model (SCI-IBI-LS), with no significant difference in prediction accuracy, highlighting the important role of the exogenous feature and that a reduced set of complementary features can still capture most of the relevant information for estimation of cerebral maturation. These results underscore the complementary nature of spontaneous and stimulus-driven brain activity, reinforcing the idea that a multidimensional approach — considering both intrinsic and evoked neural dynamics — can provide a more robust estimation of cerebral maturation.

SCI provides a measure that can quantify cerebral age in single newborns. The increase in SCI with gestational age suggests that the neural response to exogenous stimulation does not simply become more spatially widespread, but instead includes spatial differentiation and integration. A key advantage of SCI is its independence from precise component extraction, a limitation inherent to evoked potential analysis. While the latency of evoked potentials evolves with neural maturation,31,58,59 accurately detecting latency requires clearly formed peaks, which are not necessarily well-defined in young infants at the individual level, due to poor signal-to-noise ratios. This challenge is compounded by the need to predefine regions of interest or apply spatial filters. In contrast, SCI captures both the temporal dynamics of neural responses and the involvement of distributed cortical regions, offering a robust and promising metric to evaluate the brain’s response to exogenous auditory stimulation in clinical applications. SCI can be influenced by the ongoing development of thalamocortical loops, which are still maturing at this stage of neurodevelopment.18,60,61 Consequently, SCI may also reflect the interplay of cortical and subcortical networks.

Mean IBI has previously been introduced as a biomarker of neurodevelopment.7,62,63 The observed decreasing trend of IBI with increasing gestational age is in agreement with previous reports of cortical activity becoming more continuous as the brain matures, resulting in shorter quiescent periods between bursts.63,64,65,66 The EEG spectrum is a general measure that provides information about the intrinsic activity of the brain.67,68 From the practical point of view, the EEG spectrum, once calculated and averaged over relatively long recording periods, is not impacted by short-lasting artifacts and does not depend on sophisticated automatic detection techniques. Animal and computational models relate the 1/f exponent of the EEG spectrum to the ratio of excitation and inhibition in cortical circuits.69,70 Interestingly, individual deviations in the 1/f exponent have been linked to several pathologies.46,71,72,73,74,75,76 A major benefit of calculating the 1/f exponent for evaluating neurodevelopment is that it treats the power spectrum as a continuous dimension rather than requiring a priori assumptions about specific spectral bands. These characteristics make the 1/f exponent of the EEG spectrum an ideal candidate for evaluating the resting-state neural activity for automatic clinical evaluation of neurodevelopment. In this study, the 1/f exponent calculated over the full-head EEG decreased between 28 and 36 wGA. This was in agreement with the process taking place between the first 7 months of life,44 but in contrast to observations during the first two months after birth.35 The discrepancies might be related to underlying neurodevelopmental mechanisms and the rapid evolution of the inhibitory/excitatory balance around the age of term.35 However, considering the difference in the developmental periods, the origin of these discrepancies is difficult to explain.

In conclusion, this study demonstrates the potential of assessing the premature brain’s response to exogenous stimulation through response complexity as an objective and automated predictor of cerebral age. Furthermore, we highlight the complementarity of endogenous and exogenous dimensions in providing a more comprehensive evaluation of neural maturation in premature newborns, paving the way for earlier identification of at-risk populations. Both endogenous and exogenous measures proposed in the current study capture global trends in neural activity, making them promising candidates for developing automatic bedside tools to evaluate cerebral maturation. PCI has previously been demonstrated as an objective and robust measure of consciousness in adults. Our proposed SCI offers a potential alternative for neonates by quantifying the complexity of EEG responses to sensory, rather than electromagnetic, perturbations. To establish its clinical utility, the proposed index should first be validated in a preclinical EEG setting, assessing its performance as a function of electrode density and stimulation duration. Moreover, these measures could be extended beyond term age, offering potential applications for monitoring longer-term brain development.

One limitation of the present study is that sleep stages were not separated, and bursts of activity were detected across the entire recording of spontaneous activity. Although it is unlikely that this affected the developmental trend of the linear slope of the aperiodic component, a sleep state–specific analysis of burst activity could provide a more detailed understanding of the evolution of spontaneous neural activity. In addition, because quiet sleep is characterized by more discontinuous and “immature” EEG patterns than active sleep and wakefulness, incorporating vigilance-state information may improve model accuracy. The present study was conducted exclusively on normally developing premature newborns, preventing the demonstration of the prognostic value of our measures. Future research should thus precisely investigate deviations from normal cerebral maturation to identify premature newborns at risk of neurodevelopmental disorders, for which our results provide a valuable normative model.