Sex differences in prenatal development of neural complexity in the human brain

The complexity of neural activity is a commonly used read-out of healthy functioning in cortical circuits. Prior work has linked neural complexity to the level of maternal care in preterm infants at risk for developing mental disorders, yet the evolution of neural complexity in early human development is largely unknown. We hypothesized that cortical dynamics would evolve to optimize information processing as birth approaches, thereby increasing the complexity of cortical activity. To test this hypothesis, we conducted the first ever study relating prenatal neural complexity to maturation. MEG recordings were obtained from a sample of fetuses and newborns, including longitudinal data before and after birth. Using cortical responses to auditory irregularities, we computed several entropy measures which reflect the complexity of the MEG signal. Despite our hypothesis, neural complexity significantly decreased with maturation in both fetuses and newborns. Furthermore, we found that complexity decreased significantly faster in male fetuses for most entropy measures. Our surprising results lay the groundwork for the first ever mapping of how neural complexity evolves in early human development, with important implications for future efforts to develop predictive biomarkers of psychiatric disorders based on the complexity of perinatal MEG signals.


Introduction 7
The entropy or "complexity" of neural signals has been pro-8 posed as a useful readout of cortical circuit functioning. Across 9 a wide variety of clinical contexts, including neuropsychiatric 10 disorders (1-3) and neurodegenerative disorders (4, 5), low 11 entropy has been associated with, or even predictive of (1), 12 impaired functioning. Similarly, in healthy individuals, low 13 entropy is associated with states with a low capacity for infor-14 mation processing, such as non-rapid eye movement (NREM) 15 sleep (6, 7) or general anesthesia (8,9). On the other hand, 16 high entropy is sometimes associated with abundant flexibil-17 ity in circuits, e.g., in healthy individuals under the influence 18 of psychedelic drugs (10, 11). A closely related measure of 19 cortical dynamics, known as criticality, also follows a simi-20 lar pattern from general anesthesia to the psychedelic state 21 (12). The proper interpretation of neural entropy is open to 22 some debate, yet these lines of evidence, combined with analo-23 gous findings from other organ systems (13,14), suggest that 24 optimal levels of physiological complexity might be general 25 signatures of adaptability, flexibility, and/or efficient informa-26 tion processing, with potential for predicting mental health 27 outcomes (1,15). 28 In order to explore neural complexity in a developmental 29 context, the meaning of neural signal entropy in prenatal and 30 neonatal development should be better understood. Prior stud-31 ies have examined preterm infants with this goal, given that 32 preterm births are several times more likely to be associated 33 with a later neuropsychiatric disorder (16-18). Two early elec-34 der conditions of controlled thalamocortical perturbation may 81 be valuable for probing circuit development (28). Our current 82 study addresses these research directions by examining neural 83 entropy from event-related MEG signals in human fetuses and 84 newborns, with the hypothesis that MEG signal complexity, 85 like anatomical complexity (29), should increase with age both 86 before and after birth. We measured cortical responses to au-87 ditory irregularities using 101 MEG recordings that passed 88 rigorous quality control in a cohort with prior evidence of 89 P300-like neural responses (30, 31), including 43 fetuses and 90 20 newborns; 16 subjects gave longitudinal data both before 91 and after birth. Because sensory "oddballs" generate Bayesian 92 prediction errors, we reasoned that they should more strongly 93 perturb the thalamocortical system than repetitive trains of 94 identical stimuli. Our approach thus takes inspiration from 95 the perturbational complexity index (often referred to as PCI) 96 (6, 8) , which emphasizes causal influences within a system 97 (32) [however, our method also features important differences, 98 see (28)].
99 Surprisingly, we found evidence that neural entropy declines 100 with maturation in both fetuses and newborns, and that this de-101 cline cannot be easily explained by noise differences according 102 to a decision tree for classifying neural dynamics (33). The 103 entropy decline occurs faster for male fetuses, corroborating 104 earlier work suggesting sex differences in sensory evoked fetal 105 MEG signals (34). The evolution of neural complexity in early 106 development and its relationship to fetal sex should be under-107 stood prior to efforts toward using neural complexity as an in 108 utero marker of prematurity and neuropsychiatric risk. P r e p r i n t Each auditory sequence consisted of four tones of 200 ms duration each, separated by a 400 ms intertone interval. The entire sequence, from the onset of the first tone to the offset of the fourth tone, was 2000 ms in duration. The fourth tone of each sequence varied during the test phase. After averaging across trials within each condition, we analyzed signals starting from 200 ms prior to the onset of the first tone to 1000 ms following the offset of the fourth tone (3200 ms duration total). (B) For fetal recordings, the expecting mother-to-be must position her abdomen within the concavity of the MEG the sensor array, with a sound balloon placed in between her body and the SARA device to deliver auditory tones. (C) Fetal MEG signals are recorded noninvasively in response to auditory tones. To correct for the influence of fetal head orientation and size on MEG signal amplitude, all signals were normalized as percent maximum response (PMR) relative to the maximum amplitude recorded during the earlier exposure phase. The average across all recordings for global and standard deviant trials are shown in magenta and gray, respectively. (D) After birth, a subset of subjects returned to the laboratory as newborns and were recorded from after being placed in a cradle oriented head-first toward the SARA device's SQUID magnetometer array. In order to safely deliver auditory stimuli to the neonatal brain, the newborn wore infant-friendly headphones. (E) The SARA device records cortical signals noninvasively from newborns; again, the average of all global deviant trials is shown in magenta, and the average of all gloabl standard trials is shown in gray. Note that B and D are adapted from (28).
increase with age in both fetuses (using gestational age) and 125 newborns. Additional predictors such as stimulus sequence 126 type (i.e, 'ssss' or 'sssd'), block rule, heart rate variability 127 (i.e., the standard deviation of heart beat intervals, SDNN, as 128 an index of arousal), sex, PMA at birth, maternal age, and 129 maternal body mass index (BMI) prior to pregnancy were (magnetocardiography or MCG), the latter of which was re-140 tained separately to measure HRV [one of the main parameters 141 used to classify fetal behavioral and sleep states (35)] and, by 142 proxy, arousal. The dataset consisted of 81 usable recordings 143 of cortical and cardiac signals from 43 fetuses (gestational age 144 range: 25 -40 weeks) which passed strict MEG quality control 145 (74.3% of recordings retained). To correct for the confounding 146 influence of differences in fetal head size and orientation, we 147 normalized fetal MEG signals as percent maximal response 148 (PMR) relative to MEG recorded during the earlier exposure 149 phase of the experiment (Fig. 1 ings from 43 fetal subjects were used in fetal entropy models 158 regardless of whether they contained usable MCG. 159 We additionally included 20 recordings from 20 newborns 160 (age range: 13 -59 days) acquired with the same MEG system 161 and which also passed strict MEG quality control (60.1% of 162 recordings retained); 16 newborns were also recorded from 163 prior to birth in the fetal cohort. All newborns in our sample 164 had full-term births. 165 We estimated the complexity of MEG signals averaged 166 across trials and channels according to six different entropy 167 measures (see Materials and Methods; note that channel-168 averaging is not principally motivated by low SNR but, rather, 169 by variable head positions between recordings, which pre-170 cludes direct comparisons of data from a specific channel be-171 tween recordings, even within the same subject, as the fetal 172 or neonatal head may have been oriented differently in each 173 instance. Two entropy approaches, Lempel cating whether data were recorded from fetuses or newborns.

274
All six entropy measures significantly declined with PMA 275 (P < 0.0001) and were smaller for newborns than for fetuses 276 (P < 0.01). Although sex was entered into the model as a fixed 277 effect, it did not significantly predict entropy for the pooled 278 data (but note a trend of lower PermEn32 in males, P = 0.03).

279
For exact P-values, see Table 4.  and CTW in newborns did not yield significant differences 312 between amplitude and non-amplitude signal components (but 313 note a trend-level effect for neonatal CTW, P < 0.05). in separate models fit to fetal data. In all cases, we found no 322 significant effect of mediation, even without accounting for 323 multiple testing (see Table S1).

Evidence of cortical stochasticity in fetuses and 325
neonates. To assess the degree to which cortical MEG signals 326 differed from noise, i.e., surrogate signals with the same ampli-327 tude distribution, we next tested whether the entropy estimates 328 for cortical signals were significantly different from the entropy 329 of surrogates. In each case, we modeled entropy separately in 330 fetuses and newborns using surrogacy as a predictor in LMMs 331 alongside the same fixed and random effects as were used in 332 the previous entropy models. Surrogacy did not significantly 333 predict any entropy measure except for mSampEn in newborns 334 (P < P crit , higher entropy in surrogate signals; see Table S2 for 335 exact P-values); note also a trend-level effect for PermEn32 336 in fetuses (P < 0.05, lower entropy in surrogate signals). This 337 generally null result suggests that MEG signals lack nonlinear 338 P r e p r i n t Red data points (fetuses) and blue data points (newborns) are taken from each session and rule/stimulus condition, with thin dotted lines connecting longitudinal data from the same subjects. All neonatal data were filtered at the same frequency as fetal data such that neonatal signal entropy estimates could be compared with fetal entropy estimates in the same space, revealing a continuous entropy decline with postmenstrual age. Time of normal birth is indicated in cyan. We averaged data both within-and between-subjects at the level of one-week time bins (magenta circles) and computed the variance explained (R 2 ) according to the least-squares fit of the smoothed data. Because of the GA x sex interaction found in fetal data, we plotted data separately for males (first and third columns) and females (second and fourth columns). The above data show that signal entropy follows a smooth, continuous decline ranging from gestation to after birth. In all cases, postmenstrual age explains the majority of variance in signal entropy.
components that affect entropy.

339
Given this likely lack of nonlinearity, at least some degree 340 of stochasticity, or intrinsic randomness, appears to be present 341 in MEG signals. Using the Toker decision tree algorithm (33), 342 we found that the proportion of signals with deterministic dy-343 namics was significantly larger in fetuses than in newborns 344 (i.e., neonatal signals were more likely to show intrinsically 345 random behavior; χ 2 = 28.6, P = 9.1 · 10 −8 ); note that each 346 recording was treated as an independent sample). When we 347 applied the same predictor from Table 1 and Table 2  recordings from the same subjects (Table 3). For this reason, 371 fetal correlations were estimated using the beta coefficeints of 372 LMMs with random intercepts (see Materials and Methods).

373
As expected, in both fetal and neonatal data, entropy mea-   mately 400 -750 ms after the onset of the fourth tone (Fig. 3B).  within-group differences in neural complexity based on level 528 of maternal care (21, 22) and between-group differences in neu-529 P r e p r i n t  (Table 1 ), as lower 551 entropy signals should show a more pronounced evoked re-552 sponse which is easier to detect than responses in high-entropy 553 waveforms, whose complexity may mask an evoked response 554 (Fig. 4A).

555
Besides MEG, functional magnetic resonance imaging 556 (fMRI) has also been used to detect prenatal sex differences in 557 human brain activity. The maturation of ERF components, which introduce struc- the study focused on differences between diagnostic groups 624 and did not report the relationship between entropy and age.

625
Our finding of a net decline in complexity with maturation 626 was robust across six different entropy measures (       it is reasonable to expect that boys and girls will, on average, 695 show differences in neural entropy during childhood and even 696 infancy. However, it remains unknown whether these factors 697 indeed explain the sex finding in our data.

698
Although our sample of infants and fetuses was recruited 699 from healthy mothers/expecting-mothers with normal-risk 700 pregnancies, subjects still differed along relative risk factors 701 whose correlations with neural variables may inform future 702 efforts to derive risk biomarkers. Besides the higher relative 703 risk incurred by male fetal sex (59), which corresponded to 704 the risk factor of BMI (53-55). Maternal BMI did correspond 706 to lower synchronization in neonatal ERSPs (Fig. 3G) to relate these data to later mental health outcomes due to 754 the limitations of current technology (e.g., inflexible sensory 755 placement, limited SNR). However, we expect that this may 756 change in the near future, especially for newborns, due to 757 advances in optically pumped magnetometers that will allow 758 for optimal sensor placement and spatial coverage (28, 82). (GA = 30 weeks), though this case of very preterm birth was 882 due to the mother's health rather than that of the fetus. These 883 subjects were not studied after birth, as only healthy, full-term 884 infants were recruited for the neonatal cohort.

885
All recordings from newborns were cross-sectional. As alyzed. Note that we did not examine shorter subsegments 925 of trial-averaged signals (e.g., after the onset of the fourth 926 tone) due to the lowpass filtering which limits the usefulness 927 of shorter data segments (e.g., in fetal data lowpass filtered 928 at 10 Hz, a 1000 ms subsegment would contain at most 10 929 oscillatory cycles).

930
Heart rate variability (HRV   Accordingly, we identified n = 12 fetuses with data from both 1135 tertiles delineated above. After decomposing entropy changes 1136 into amplitude, phase, and amplitude x phase interaction com-1137 ponents, we summed the differences in each entropy measure 1138 attributable to phase and phase x amplitude interactions to cre-1139 ate a single "non-amplitude" quantity Data from all four condi-1140 tions (i.e., global rule x stimulus combinations) were used, and 1141 P r e p r i n t we modeled the entropy differences ∆H using LMMs with a fixed effect of condition and random intercepts for subjects; we 1143 then added a fixed effect of the signal property (amplitude or 1144 non-amplitude) and used a log-likelihood ratio test to evaluate 1145 whether adding this term significantly improved the model fit. 1146 We also applied the entropy decomposition on neonatal 1147 data; however, due to the smaller sample size of newborns, 1148 we performed a median split on the data (rather than a tertile 1149 split), grouping all newborns that were recorded from at 34 using the same log-likelihood ratio test approach described 1183 above for fetal data. To evaluate whether a given measure differed between cortical 1209 and surrogate signals, we used LMMs (see "Statistical analysis" 1210 above) and evaluated the surrogacy term in each model. ings (see Table S4), we z-scored MEG measures from fetal data 1218 and used the normalized beta coefficients from random inter-1219 cept regression models. Because standardized betas in LMMs 1220 depend on the variance of the random effect and are thus gen-1221 erally asymmetrical (i.e., β i, j ̸ = β i, j ), we used the mean of β i, j 1222 and β i, j to represent the correlation between entropy measure 1223 i and j (Fig. S1). This was done using LMMs which predicted 1224 and random intercepts, and vice versa (reversing the roles of