The genesis of OS in the preterm brain is multifactorial, with several key enzymatic and non-enzymatic sources contributing to the ROS/RNS burden. Mitochondria are primary contributors, as electron leakage from the electron transport chain, particularly during the rapid but often inefficient oxidative phosphorylation required for neurodevelopment, generates superoxide anions (14,15). This is compounded by the immature regulation of this process. Beyond mitochondria, the nicotinamide adenine dinucleotide phosphate (NADPH) oxidase (NOX) family of enzymes plays a significant role. Investigations have shown that NOX enzymes are activated in neuroinflammatory conditions, with neuronal NADPH oxidase 1 contributing to injury in models of cerebral ischemia (16).

Following hypoxia-ischemia, the enzyme xanthine oxidase (XO) is another major source of ROS, generated during the reperfusion phase when molecular oxygen is reintroduced to tissues with accumulated hypoxanthine (17). The phase III ALBINO trial (NCT03162653), an investigator-initiated, randomized, placebo-controlled, double-blinded, multi-national trial, is currently evaluating allopurinol as an adjunct to therapeutic hypothermia for hypoxic-ischemic encephalopathy; the detailed statistical analysis plan has been published, with the primary endpoint being long-term outcome determined as survival with neurodevelopmental impairment vs. death vs. non-impaired survival at 2 years of age (18). As Martini et al (19) and Hu et al (20) have detailed, free iron catalyzes the Fenton reaction, converting hydrogen peroxide into the highly toxic hydroxyl radical, thereby amplifying oxidative damage. The convergence of these diverse sources suggests that effective neuroprotection may require a multi-pronged approach rather than targeting any single enzymatic pathway, as the dominant source of ROS likely varies with the phase of injury and individual patient characteristics.

The preterm brain possesses an inherent vulnerability to OS that distinguishes it from the mature brain. A critical factor is the underdevelopment of its antioxidant enzyme systems. As Tataranno et al (21) and Kletkiewicz et al (22) have discussed, the expression and activity of key enzymes such as superoxide dismutase, catalase and glutathione peroxidase are low in the preterm period, limiting the capacity to neutralize ROS. This is further exacerbated by the interruption of the maternal-fetal transfer of essential antioxidant cofactors like selenium, zinc and copper (21).

The cellular composition of the preterm brain also heightens its risk. Oligodendrocyte progenitor cells (pre-OLs), which are abundant in the periventricular white matter during the critical window of 24–32 weeks gestation, are exquisitely sensitive to OS. Pandya et al (23) demonstrated that hemoglobin exposure induces OS and mitochondrial dysfunction in these cells, while Sunny et al (24) showed that hyperoxia-induced damage in oligodendrocyte progenitor cells (OPCs) is mediated by specific molecular pathways, highlighting their fragility. Wellmann et al (25) have explored this concept, framing the vulnerability of white matter as a consequence of either too much or too little oxygen, both of which can disrupt myelination. Even neuronal subpopulations are not spared; Scheuer et al (26) provided evidence that neonatal OS, from relative hyperoxia, impairs cortical synapse formation and γ-aminobutyric acid (GABA) homeostasis in parvalbumin-expressing interneurons, potentially explaining long-term neuropsychiatric sequelae. This developmental window of susceptibility underscores that neuroprotective strategies must be timed precisely, as interventions effective in the mature brain may be ineffective or even harmful when applied during this period of unique cellular fragility and ongoing maturation.

The unmitigated production of ROS and RNS in the vulnerable preterm brain leads to widespread damage to essential cellular macromolecules. Lipid peroxidation is a hallmark of this injury, as the brain's high content of polyunsaturated fatty acids makes it a prime target. The formation of F2-isoprostanes, stable end-products of lipid peroxidation, has been extensively studied. Negro et al (27) and Stolwijk et al (28) have independently shown that elevated levels of F2-isoprostanes in cord blood or postnatally are predictive of brain damage in term and preterm infants, respectively. Coviello et al (29) further linked higher plasma isoprostane levels with decreased functional brain activity on amplitude-integrated electroencephalography (aEEG) in extremely preterm infants, providing a direct correlation between oxidative injury and early neurophysiological dysfunction.

Proteins are also susceptible, with ROS/RNS causing carbonylation and nitration, which can impair enzyme function and disrupt cellular signaling. Nucleic acids are not immune; the formation of 8-hydroxy-2′-deoxyguanosine (8-OHdG) from DNA damage is a recognized marker of OS and can itself interfere with methyltransferase binding, linking OS directly to epigenetic dysregulation (14). The fact that oxidative damage simultaneously compromises lipids, proteins and DNA suggests that rescue strategies aimed at a single class of macromolecules are unlikely to be sufficient, reinforcing the need for interventions that bolster global cellular resilience rather than targeting isolated downstream effects.

OS does not occur in isolation; it is intimately interconnected with other major pathological mechanisms of preterm brain injury. A central relationship exists between OS and neuroinflammation. Shin et al (30) showed that tumor necrosis factor-α (TNF-α) antagonism attenuates lipopolysaccharide (LPS)-induced white matter injury, underscoring the role of inflammatory cytokines. The interaction of OS with excitotoxicity is also well-established. Overactivation of N-methyl-D-aspartate receptors, as occurs in hypoxia-ischemia, leads to calcium influx and mitochondrial dysfunction, further promoting ROS production (31,32). Mitochondrial dysfunction itself is a critical hub, as highlighted by Odorcyk et al (33), who demonstrated age-dependent mitochondrial dysfunction and apoptosis following neonatal hypoxia-ischemia. This bioenergetic failure not only generates more ROS but also compromises cellular repair mechanisms.

Furthermore, OS is a key driver of ferroptosis, an iron-dependent form of regulated cell death, which Chang et al (34) reviewed in the context of hypoxia-associated brain injury. The interplay between these pathways creates a vicious cycle where OS amplifies inflammation and excitotoxicity, which in turn exacerbate OS, culminating in cell death and long-term neurodevelopmental impairment (35,36).

In summary, there is broad consensus that mitochondria, NADPH oxidase, xanthine oxidase and iron-mediated Fenton reactions are major sources of OS in the preterm brain and that the developmental vulnerability of OPCs and immature antioxidant defenses are key determinants of injury. However, significant controversies remain. The relative contribution of each OS source likely varies with gestational age, injury phase and clinical context, yet no study has systematically compared these sources across the same cohort. Furthermore, the extent to which OS acts as a primary driver vs. a secondary amplifier of injury, particularly in relation to neuroinflammation and excitotoxicity, remains unresolved. Key unanswered questions include the following: Which OS source is the most tractable therapeutic target; whether antioxidant strategies should be initiated before, during, or after the oxidative burst; and how OS interacts with genetic and epigenetic factors to determine individual susceptibility. Addressing these questions will require integrated multi-omics approaches and longitudinal cohort studies with serial biosampling.

The clinical evidence for oxidative biomarkers (such as F2-isoprostanes, 8-OHdG) and their association with brain injury comes primarily from prospective cohort and case-control studies in preterm infants (27–29), whereas the mechanistic sources of ROS and the developmental vulnerability of OPCs are largely derived from preclinical models (14–16,20,23,26).

Epigenetic mechanisms, including DNA methylation, histone modifications and non-coding (nc)RNAs, are fundamental regulators of gene expression that operate without altering the underlying DNA sequence. During brain development, these processes orchestrate critical events such as neural stem cell proliferation, neuronal and glial differentiation, synaptogenesis and the establishment of complex neural circuits (37,38). DNA methylation, the addition of a methyl group to cytosine residues in CpG dinucleotides, is typically associated with transcriptional repression when it occurs in promoter regions and is dynamically regulated by DNA methyltransferases and ten-eleven translocation (TET) demethylases (39). Histone modifications, such as acetylation, methylation and phosphorylation, alter chromatin structure and gene accessibility, with writers, erasers and readers of these marks determining the transcriptional output (40,41). microRNAs (miRNAs) provide an additional layer of post-transcriptional control by degrading or inhibiting the translation of target mRNAs (42,43). The exquisite temporal and spatial control of these processes render the developing brain particularly susceptible to environmental perturbations that can disrupt the normal epigenetic trajectory.

OS can directly and indirectly influence the establishment and maintenance of epigenetic marks. One primary mechanism involves the modification of enzymes that regulate the epigenome. For instance, the activity of TET enzymes, which catalyze DNA demethylation, is dependent on iron and α-ketoglutarate, making them vulnerable to redox imbalance (44,45). ROS can also cause DNA damage and the repair process itself can lead to the recruitment of DNA methyltransferases to the site of damage, potentially establishing aberrant methylation patterns (11,46). Furthermore, ROS can oxidize guanine bases in DNA to form 8-oxo-7,8-dihydroguanine, which can inhibit the binding of methyltransferases to adjacent CpG sites, leading to localized hypomethylation (47). Scarpato et al (11) provided an overview of how oxidative genome damage can be a determinant of pathological conditions in the newborn by influencing DNA methylation. This direct chemical interplay between ROS and the DNA methylation machinery underscores how an acute oxidative burst can have lasting consequences for the epigenome.

A critical question that emerges from these observations is whether the OS-driven epigenetic changes represent a primary pathogenic mechanism or a secondary epiphenomenon. The evidence for causality is strongest for TET enzyme inhibition and 8-oxodG formation, where direct chemical mechanisms have been established in vitro and in animal models. For TET enzymes, their dioxygenase activity is iron- and α-ketoglutarate-dependent, rendering them sensitive to redox imbalance. OS has been shown to upregulate TET1 expression and induce active DNA demethylation in human cells (48), while TET2 counteracts aberrant DNA methylation during OS through acetylation-dependent mechanisms (49). For 8-oxodG, this oxidative DNA lesion accumulates in GC-rich promoters and regulatory regions, where it can inhibit DNMT binding to adjacent CpG sites, recruit OGG1 to initiate repair processes and interfere with cytosine methylation. 8-oxodG has also been shown to interplay with histone modifications and DNA methylation through complex regulatory networks (50). However, for most histone modifications and non-coding RNA changes, the causal chain from OS to epigenetic mark to functional outcome remains inferential rather than proven. For histone modifications, OS has been shown to modulate HDAC activity in a context-dependent manner; activated microglia can decrease histone acetylation in astrocytes via increased HDAC activity (51), while specific HDAC isoforms (HDAC1-11) are expressed in glial cells and regulate the neuroinflammatory response. For non-coding RNAs, emerging evidence has revealed complex regulatory networks involving long non-coding (lnc)RNAs, miRNAs and mRNAs that are altered in association with inflammatory and OS markers in neonatal brain injury models, although direct causal relationships remain to be fully elucidated. Specific examples include miR-34c-5p, which attenuates apoptosis and OS following hypoxic-ischemic brain injury in neonatal mice by targeting Arhgap26 (52); miR-210, which promotes microglial pro-inflammatory activation through reprogramming mitochondrial metabolism (53); the lncRNA EVADR, which mitigates neonatal hypoxic-ischemic injury via the miR-145/WNT/β-catenin axis (54); and lncRNA GAS5, which prevents mitochondrial apoptosis through the miR-128-3p/Bax/Akt/GSK-3β axis (55). Circular RNAs (circRNAs) have also been implicated, with distinct expression patterns observed in the brain tissues of neonatal rats with periventricular white matter damage, suggesting that circRNAs actively respond to hypoxic-ischemic insults (56). Moreover, the temporal relationship between OS exposure and epigenetic modification is rarely examined in human studies; cross-sectional designs cannot distinguish whether methylation changes precede or follow injury. This ambiguity has practical consequences: If epigenetic changes are predominantly secondary to established injury, then targeting them therapeutically may offer limited benefit; if they are primary drivers, they represent compelling intervention targets. Future longitudinal studies with serial sampling from the perinatal period are urgently needed to resolve this question.

A substantial body of research has now linked specific epigenetic changes to preterm brain injury. DNA methylation is the most extensively studied mark in this context. Lorente-Pozo et al (57) delivered a seminal finding by demonstrating that the oxygen load supplied during delivery room stabilization of preterm infants modifies the DNA methylation profile. This work provided direct clinical evidence that a major source of OS, resuscitation with high oxygen, can rapidly alter the neonatal epigenome. Similarly, Rzemieniec et al (44) showed that perinatal asphyxia in a rodent model led to hypermethylation of specific genes, an effect that was counteracted by a neuroprotective compound. More recently, Xu et al (58) demonstrated that maternal exposure to multiple metals, a known provocation for OS, is associated with an increased risk of spontaneous preterm birth, a relationship potentially mediated by alterations in DNA methylation. Balasubramanian et al (46) provided mechanistic depth by showing that traumatic brain injury leads to persistent oxidative damage in the hippocampus via DNA Methyltransferase 3b-mediated DNA methylation and silencing of the antioxidant enzyme superoxide dismutase2. This illustrates a vicious cycle whereby OS-induced methylation can impair endogenous antioxidant defenses, perpetuating oxidative injury.

Histone modifications are also emerging as critical players. The acetylation and deacetylation of histones, controlled by histone acetyltransferases and deacetylases (HDACs), respectively, are sensitive to cellular energy and redox states. Ye et al (41), in a comprehensive review, highlighted the potential of HDAC inhibitors as a therapeutic strategy in traumatic brain injury. Experimental studies support this concept. Liu et al (59) demonstrated that HDAC2 regulates nuclear factor erythroid 2-related factor 2 (Nrf2) acetylation, influencing neuronal ferroptosis in neonatal rats with hypoxic-ischemic brain injury. Furthermore, Xu et al (60) showed that targeting HDAC3 suppresses ferroptosis and demyelination in white matter injury by restoring pyruvate dehydrogenase kinase 4-mediated iron homeostasis. These studies point towards a crucial role for histone deacetylases in orchestrating the cellular response to OS and injury. Beyond acetylation, other histone marks are implicated. For instance, Xiang et al (61) found that exposure to fluoride, which can induce OS, exacerbates cognitive deficits in diabetic rats, potentially through mechanisms involving the poly (ADP-ribose) polymerase-1 pathway, a sensor of DNA damage that also modifies chromatin structure.

ncRNAs, particularly miRNAs, add another layer of complexity. These small RNAs can be dysregulated by OS and, in turn, regulate the expression of genes involved in antioxidant defense, inflammation and cell survival. For example, Zhou et al (43) demonstrated that suppression of the lncRNA metastasis-associated lung adenocarcinoma transcript 1 alleviated neuronal apoptosis and ROS production through the miR-499-5p/SRY-box transcription factor 6 axis in a model of subarachnoid hemorrhage. Similarly, Chen et al (42) reported that depleting SRY-box transcription factor 2 improved ischemic stroke outcomes via the lncRNA plasmacytoma variant translocation 1/microRNA-24-3p/signal transducer and activator of transcription 3 (STAT3) axis. The rapid responsiveness of miRNAs to cellular stress makes them attractive candidates as both early biomarkers and therapeutic targets (62).

It is crucial to recognize that epigenetic responses to OS are not uniform across all brain cell types. The specific vulnerability of OPCs, discussed in the previous section, likely has an epigenetic basis. Kulkarni et al (63) demonstrated that brain trauma induces DNA methylation-mediated regulation of the Mitofusin 2 gene, which is critical for mitochondrial dynamics, in the brain, but the cell-specific nature of this regulation requires further elucidation. Astrocytes also undergo significant epigenetic reprogramming in response to injury. Li et al (64) showed that alcohol exposure, which induces OS, activates the Nrf2 pathway and influences the expression of genes related to glutathione metabolism in astrocytes, a process with potential epigenetic underpinnings. Microglia, the brain's resident immune cells, are particularly prone to epigenetically mediated priming, where an initial insult leads to a sensitized state and an exaggerated response to a subsequent challenge. Chen et al (65) recently reported that HDAC1 dysregulation promotes pro-inflammatory microglial activation and aggravates post-stroke neuroinflammation. This suggests that OS-induced epigenetic changes in microglia could perpetuate a chronic neuroinflammatory state, contributing to ongoing white matter dysmaturation. While much work remains to be conducted, it is evident that the epigenetic landscape of each cell type dictates its unique response to the OS caused by preterm birth. The convergence of evidence from clinical cohorts demonstrating oxygen-induced methylation changes (57) and preclinical models elucidating the pathways of HDAC-mediated ferroptosis (59,60) strongly supports the concept that OS serves as a direct molecular bridge to enduring epigenetic dysregulation. This paradigm shifts from viewing OS as merely an acute insult to recognizing it as a trigger for long-term epigenetic reprogramming opens new avenues for biomarkers and therapies aimed at mitigating the lifelong neurodevelopmental consequences of preterm birth.

The direct clinical evidence linking oxygen exposure to DNA methylation changes in preterm infants is provided by a prospective cohort study (57), while most mechanistic insights into histone modifications and ncRNAs originate from preclinical models (41–44,59–61).

OPCs represent the most vulnerable cellular population in the preterm brain, with their susceptibility underpinning the predominant white matter injury phenotype. Srivastava et al (66) demonstrated that a Toll-like receptor (TLR)/protein kinase B (AKT)/forkhead box O3 immune tolerance-like pathway specifically disrupts the repair capacity of OPCs, providing mechanistic insight into why these cells fail to mature after injury. This maturational arrest is further exacerbated by inflammatory signals; Boccazzi et al (67) showed that TLR3 activation in oligodendrocytes triggers a pro-inflammatory response that compromises their survival in a murine model of preterm white matter injury.

The susceptibility extends beyond inflammatory mediators to excitotoxic mechanisms, as Al-Griw et al (68) demonstrated that inhibition of ionotropic glutamate receptor signaling preserves the oligodendrocyte lineage in an ex vivo model of ischemic injury. Importantly, the fate of OPCs is not uniformly determined. Chang et al (69) provided evidence that intrauterine growth restriction combined with postnatal hyperoxia induces white matter injury through mechanisms involving OPCs disruption, while Lin et al (70) characterized the dynamic changes in oligodendrogenesis following hypoxic-ischemic injury, revealing temporal windows of vulnerability and limited regenerative capacity. These findings collectively indicate that OPCs injury involves multiple converging pathways and the failure of endogenous repair mechanisms, rather than simply cell death, may be the critical determinant of long-term dysmyelination. The observation that OPCs can survive but fail to mature suggests that therapeutic strategies should target differentiation blockade rather than solely preventing acute cell loss.

Astrocytes exhibit context-dependent responses that can either contain injury or amplify damage in the preterm brain. Parfenova et al (71) demonstrated a protective role, showing that astrocyte-produced carbon monoxide protects against cerebrovascular dysfunction following neonatal asphyxia. This beneficial capacity is consistent with findings from Nobuta et al (72), who reported that STAT3-mediated astrogliosis protects myelin development by modulating microglial transforming growth factor beta-1 (TGFβ-1) expression, thereby preventing impaired oligodendrocyte maturation.

Conversely, astrocytes can adopt pathogenic phenotypes under certain conditions. Renz et al (73) demonstrated that neuroinflammatory reactive astrocyte formation correlates with adverse outcomes in perinatal white matter injury, suggesting that the nature of astrocytic reactivity determines clinical trajectory. The signaling pathways governing these divergent responses have been increasingly characterized. Fang et al (74) showed that fibroblast growth factor-10 protects the neurovascular unit, including astrocytes, in neonatal hypoxic-ischemic injury, while Xu et al (60) demonstrated that targeting HDAC3 suppresses ferroptosis and demyelination partly through restoring PDK4-mediated iron homeostasis. The duality of astrocyte function (protective under some conditions but detrimental under others) highlights the need for nuanced therapeutic approaches that preserve beneficial functions while limiting maladaptive responses. This balance may be particularly critical in the preterm brain, where astrocyte development is ongoing and thus highly susceptible to perturbation.

Microglial cells serve as central orchestrators of the neuroinflammatory response, with their activation state critically influencing injury progression and repair (75). Activated microglia produce large quantities of ROS via NADPH oxidase 2 (NOX2), which can in turn perpetuate inflammatory signaling, establishing a feed-forward loop between OS and neuroinflammation (8). Yang et al (76) demonstrated that caffeine treatment, initiated before injury, reduces hypoxic-ischemic white matter damage by regulating phenotypic microglial polarization, shifting cells toward a less inflammatory state. This concept of microglial plasticity is supported by Charriaut-Marlangue et al (77), who identified sex differences in microglial phenotypes following neonatal stroke, with females showing distinct responses to poly (ADP-ribose) polymerase inhibition compared with males.

The signaling pathways regulating microglial reactivity have been extensively investigated. Zaghloul et al (78) reported that prophylactic inhibition of NF-κB expression in microglia attenuates hypoxic-ischemic injury, while Mairesse et al (79) showed that oxytocin receptor agonist reduces perinatal brain damage by directly targeting microglia. Van Steenwinckel et al (80) provided mechanistic depth by demonstrating that decreased microglial Wnt/β-catenin signaling drives pro-inflammatory activation in the developing brain. The convergence of these studies suggests that microglial responses are not binary but represent a spectrum of activation states and the capacity to modulate these states, either pharmacologically or through endogenous pathways, offers therapeutic potential. Importantly, the work of Bernis et al (81), which temporally characterized microglia-associated gene expression in a sensitized hypoxia-ischemia model, emphasizes that the timing of microglial modulation may be as critical as the specific intervention.

Although white matter injury predominates in preterm encephalopathy, neuronal populations are also markedly affected, contributing to long-term cognitive and motor deficits. Sathyanesan et al (82) demonstrated that neonatal brain injury causes cerebellar learning deficits and Purkinje cell dysfunction, revealing that the cerebellum is a previously underappreciated target. This finding aligns with subsequent work from the same group showing that disruption of neonatal Purkinje cell function underlies injury-related learning deficits (83). Cortical neuronal populations are similarly vulnerable. Scheuer et al (26) provided evidence that neonatal OS impairs cortical synapse formation and GABA homeostasis in parvalbumin-expressing interneurons, a finding extended by Lacaille et al (84) who demonstrated impaired interneuron development in a novel model of neonatal brain injury. Sheikh et al (85) showed that neonatal hypoxia-ischemia causes functional circuit changes in subplate neurons, which serve as crucial relay stations during cortical development.

The cholinergic system also appears susceptible; Northington et al (86) reported that basal forebrain magnocellular cholinergic systems are damaged following neonatal hypoxia-ischemia in mice. These diverse neuronal vulnerabilities suggest that preterm brain injury represents a multi-system disorder affecting multiple neurotransmitter systems and brain regions. The recognition that interneurons, particularly parvalbumin-positive subtypes, are consistently affected provides a plausible cellular basis for the long-term neuropsychiatric sequelae observed in preterm survivors.

The concept of the neurovascular unit emphasizes that injury to any component (endothelium, pericytes, astrocytes, or neurons) affects the integrity of the whole. Lee et al (87) demonstrated that hypoxia-preconditioned human umbilical vein endothelial cells protect against neurovascular damage following hypoxic ischemia, suggesting that enhancing endothelial resilience confers broader neuroprotection. Chand et al (88) showed that neurovascular unit alterations in growth-restricted newborns are improved following ibuprofen treatment, linking hemodynamic modulation to structural protection. Fang et al (89) provided additional evidence that chloroquine protects against neonatal brain injury by mitigating blood-brain barrier disruption, while Wang et al (90) demonstrated that oligogenesis in the ‘oligovascular unit’ involves PI3K/AKT/mTOR signaling. The interdependence of these cellular components has important therapeutic implications: Interventions targeting a single cell type may fail if other components remain compromised, suggesting that multimodal approaches addressing the entire neurovascular unit may be necessary.

Emerging evidence implicates the gut microbiota as a significant modulator of brain injury in preterm infants. Seki et al (91) provided seminal findings by demonstrating that aberrant gut-microbiota-immune-brain axis development occurs in premature neonates with brain damage, with Klebsiella overgrowth highly predictive of injury and associated with a pro-inflammatory immunological tone. This clinical observation is supported by experimental work from Drobyshevsky et al (92), who showed that intestinal microbiota modulates neuroinflammatory response and brain injury following neonatal hypoxia-ischemia. Liu et al (93) extended these findings using multi-omics analyses, demonstrating aberrant microbiota-gut-brain axis in very low birth weight infants with white matter injury. Vaher et al (94) recently confirmed that the neonatal gut microbiota plays a role in the encephalopathy of prematurity, with specific microbial profiles associated with adverse neurodevelopmental outcomes. Seki et al (95) subsequently identified gut microbiota genome features associated with brain injury, suggesting that the mechanistic basis for these associations likely involves microbial metabolites and immune modulation. This evolving field opens novel therapeutic possibilities, including probiotic interventions and dietary modulation, although causal relationships require further validation. The recognition that the gut-brain axis operates during the neonatal period challenges traditional conceptualizations of brain injury as a purely CNS-limited process and suggests that systemic interventions may have central consequences.

The cellular landscape of preterm brain injury reveals a highly interconnected system in which the response of any single cell type influences and is influenced by others. This complexity probably explains why single-target interventions have shown limited clinical success and supports the need for approaches that consider the multicellular nature of injury and repair. Critically, the OS epigenetic axis provides the mechanistic thread that connects these disparate cellular responses: OS modifies the epigenome of each cell type in a context-specific manner and these epigenetic changes, in turn, dictate divergent outcomes ranging from OPCs maturational arrest to astrocyte reactivity and microglial priming. Thus, the cellular diversity of preterm brain injury is not a random assortment of pathologies but a spectrum of cell-type-specific manifestations of a common molecular derangement. However, the cell-type specificity of epigenetic responses also poses a major translational challenge: Interventions that modify the epigenome globally may have unintended consequences in cell types where the same epigenetic mark serves a protective rather than pathogenic function. For example, HDAC inhibition may benefit oligodendrocyte survival while simultaneously promoting astrocyte reactivity or microglial priming, depending on the cellular context and timing of intervention. The responses of astrocytes and microglia to HDAC inhibition are highly context-dependent, varying with the specific HDAC inhibitor used, the cell type and the inflammatory milieu. In pure microglial cultures, HDAC inhibitors such as trichostatin A (TSA) and suberoylanilide hydroxamic acid can potentiate LPS-induced pro-inflammatory cytokine and nitric oxide responses (96). Conversely, in astrocyte-rich cultures exposed to activated microglial factors, TSA and valproic acid restored histone acetylation, upregulated the Nrf2-inducible antioxidant defense and conferred protection against OS (51). In a neonatal rat model of white matter injury, the selective HDAC2/3 inhibitor MI192 administered post-insult markedly reduced oligodendrocyte loss and suppressed microglial activation (97). These findings collectively illustrate that the timing of HDAC inhibition and the specific cell types targeted critically determine therapeutic outcomes, whether the inhibitor is administered prophylactically, immediately after injury, or during the repair phase. This duality underscores the need for cell-type-specific delivery strategies or interventions targeted to the relevant epigenetic enzymes in a temporally restricted manner. It must be emphasized, however, that the majority of mechanistic insights discussed in this section derive from preclinical models (rodent and ex vivo systems). While these findings are invaluable for understanding pathogenesis, their translation to the human preterm brain requires confirmation in clinical cohorts.

Clinical evidence for the gut-brain axis in preterm brain injury comes from prospective cohorts and case-control studies (91,93–95), whereas the majority of cellular response data (oligodendrocyte, astrocyte, microglia, neuron) are derived from animal models. See Table I for full classification.

The pathological responses of individual cell types in preterm brain injury do not occur in isolation; rather, they are shaped by dynamic intercellular communication networks that collectively determine injury progression and repair capacity. Microglia and astrocytes engage in bidirectional crosstalk that profoundly influences the neuroinflammatory milieu. Activated microglia can decrease histone acetylation in astrocytes via increased HDAC activity, thereby impairing astrocytic antioxidant defenses and perpetuating OS (51). Conversely, astrocytes modulate microglial phenotype through STAT3-dependent mechanisms; Nobuta et al (72) demonstrated that STAT3-mediated astrogliosis protects myelin development by preventing pathological TGFβ-1 expression in microglia, which would otherwise inhibit oligodendrocyte progenitor maturation. This protective astrocyte-microglia interaction is critical, as TGFβ-1 directly and dose-dependently inhibits the maturation of purified oligodendrocyte progenitors (11).

Astrocytes also exert direct effects on oligodendrocyte lineage cells. Shiow et al (98) showed that reactive astrocytes in neonatal white matter injury develop ‘A2’ reactivity and produce COX2-derived prostaglandin E, which acts directly on oligodendrocyte progenitor cells via the EP1 receptor to block their maturation. In vivo inhibition of COX2 rescues hypomyelination and behavioral impairment, demonstrating that astrocyte-derived inflammatory mediators are functionally significant drivers of OPC maturation arrest (98).

Microglia-oligodendrocyte interactions are equally critical. Inflammatory activation of microglia releases cytokines and ROS that directly impair oligodendrocyte survival and maturation (8,30). Conversely, protective microglial phenotypes can support oligodendrocyte development; M2 microglia-derived exosomal miR-144-5p has been shown to promote OPC growth and reduce white matter injury by regulating the PTEN/AKT pathway (99). The work of Bernis et al (81) further emphasizes that the temporal dynamics of microglial activation, shifting from pro-inflammatory to anti-inflammatory gene expression over time, determine whether microglia exert harmful or beneficial effects on oligodendrocyte lineage cells.

The neurovascular unit represents another critical interface for intercellular crosstalk. Endothelial cells, pericytes, astrocytes and neurons communicate through reciprocal signaling that maintains blood-brain barrier integrity and regulates cerebral blood flow. Lee et al (87) demonstrated that hypoxia-preconditioned endothelial cells confer broader neuroprotection, suggesting that enhancing the resilience of one cellular component benefits the entire unit. Wang et al (90) showed that oligogenesis in the ‘oligovascular unit’ involves PI3K/AKT/mTOR signaling, illustrating how vascular and glial compartments coordinately regulate white matter repair. Collectively, these findings establish that preterm brain injury is not a cell-autonomous process but a network disorder in which the responses of microglia, astrocytes, oligodendrocytes and vascular cells are intimately interconnected through reciprocal signaling pathways. This interconnectedness provides a mechanistic explanation for why single-cell-type interventions have shown limited success and supports the rationale for multimodal therapeutic strategies that target the broader cellular network.

Melatonin has been extensively investigated as an antioxidant neuroprotectant. Carloni et al (100) characterized melatonin pharmacokinetics following oral administration in preterm neonates, demonstrating significant interindividual variability that complicates dosing optimization. Domínguez Rubio et al (101) reported that maternal melatonin administration exerted both short- and long-term neuroprotective effects in offspring from lipopolysaccharide-treated mice. Garofoli et al (102) recently confirmed that orally administered melatonin reaches systemic circulation and reduces OS markers in preterm newborns, although clinical impact on neurodevelopmental outcomes awaits definitive trials. Lee et al (103) demonstrated in a mouse model that melatonin administration prevented preterm birth and associated fetal brain injury. However, the systematic review by Ahmed et al (104) concluded that while safety appears favorable, evidence for neuroprotection remains insufficient due to small sample sizes and heterogeneity in dosing and timing. A more recent review by Häusler et al (105) systematically examined the pathomechanisms underlying preterm brain injury and correlated them with melatonin's neuroprotective potential, while emphasizing significant pharmacokinetic and pharmacodynamic uncertainties that must be addressed before clinical translation can be realized. Favrais et al (106) showed only partial protective effects of melatonin in a rat model of chorioamnionitis, suggesting that monotherapy may be inadequate against multifactorial injury. Other antioxidants, including N-acetylcysteine (107), vitamin C (108) and uridine (109), have shown preclinical promise but limited clinical translation. Chen et al (110) recently demonstrated that targeting mitochondrial OS through Nrf2 induction protects against preterm birth and fetal brain injury in experimental models, although clinical validation is awaited. Beyond these agents, emerging preclinical evidence continues to expand the repertoire of potential antioxidant neuroprotectants. Gunduz et al (111) recently demonstrated that polydatin, a natural polyphenolic compound, exerts dose-dependent protective effects against sodium fluoride-induced oxidative and inflammatory damage in rats, as evidenced by restored antioxidant enzyme activities (glutathione, catalase, superoxide dismutase), reduced lipid peroxidation (malondialdehyde) and suppressed pro-inflammatory gene expression (TNF-α, NF-κB, IFN-γ). Although this study was conducted in a fluoride-toxicity model rather than a perinatal brain injury context, it reinforces the broader principle that pharmacological agents capable of restoring redox balance and modulating inflammatory pathways hold therapeutic potential for conditions characterized by OS. Importantly, polydatin has also been shown to exert neuroprotective effects in neonatal rat models of hypoxic-ischemic brain injury, improving learning and memory impairments through upregulation of brain-derived neurotrophic factor (112). These findings suggest that polydatin warrants further investigation as a candidate neuroprotective agent in the preterm population, although rigorous preclinical optimization and clinical translation studies are required before any clinical recommendations can be made.

Erythropoietin (EPO) represents the most extensively studied pathway-targeted agent. Fischer et al (113) performed a meta-analysis concluding that while EPO reduces transfusions, evidence for improved neurodevelopmental outcomes was inconclusive. The landmark Preterm Erythropoietin Neuroprotection Trial (PENUT) trial by Juul et al (114) randomized 941 extremely preterm infants to high-dose EPO or placebo and found no significant difference in the primary outcome of mortality or severe neurodevelopmental impairment at 2 years between the two groups. Ehrenreich et al (115) emphasized that negative findings highlight the importance of patient selection and dosing rather than refuting EPO's potential. By contrast, Song et al (116) reported that EPO improved outcomes specifically in preterm infants with intraventricular hemorrhage, suggesting subgroup-specific benefits. Jakab et al (117) used network-based statistics to reveal trophic effects of early high-dose EPO on brain connectivity in very preterm infants. Wassink et al (118) showed that prolonged EPO infusion provided partial white and grey matter protection after asphyxia in preterm fetal sheep. Ma and Shi (119) comprehensively reviewed cellular mechanisms of EPO neuroprotection, concluding that clinical translation requires improved patient stratification. The discrepancy between the neutral PENUT trial and positive preclinical and subgroup studies likely reflects several factors: The PENUT trial administered EPO at 24 h of life, whereas preclinical data suggest earlier administration (within 6 h) may be more effective; the extremely high baseline rate of favourable outcomes in the control group (79.6% survival without impairment) limited statistical power to detect a benefit; and the trial excluded infants with significant prior brain injury, potentially removing the population most likely to respond (114,119). These observations suggest that EPO may benefit specific subgroups, particularly infants with established injury or those treated earlier, rather than providing universal neuroprotection and that future trials should stratify by injury severity and timing of intervention. The conflicting results between consistently positive preclinical models and largely neutral clinical trials underscore the limitations of animal models and the critical importance of adequately powered trials.

Caffeine has emerged as a candidate neuroprotective agent. Endesfelder et al (120) demonstrated that caffeine provides neuroprotection against hyperoxia-induced neonatal brain injury in rats. Yang et al (76) showed that caffeine treatment reduced hypoxic-ischemic white matter damage through regulation of phenotypic microglia polarization. Di Martino et al (121) defined a therapeutic window for caffeine neuroprotection, showing that early administration is critical for efficacy. A recent network meta-analysis by Wang et al (122) comparing caffeine, erythropoietin, magnesium sulfate and thyroxine found that caffeine showed the most promising effect in reducing the risk of neurodevelopmental impairment [relative risk (RR): 0.43, 95% confidence interval (CI): 0.22–0.86] and cerebral palsy (RR: 0.55, 95% CI: 0.43–0.70) in preterm infants. These preclinical data, combined with caffeine's established safety profile, support its continued use, although prospective trials with neurodevelopmental primary outcomes are lacking. Emerging pathway-targeted strategies include HDAC modulation, with Liu et al (59) demonstrating that HDAC2 regulates Nrf2 acetylation and thereby influences neuronal ferroptosis. Peroxisome proliferator-activated receptor gamma (PPAR-γ) activation has also been investigated, with Fang et al (123) demonstrating inhibition of microglia-mediated neuroinflammation and Feng et al (124) showing promotion of oligodendrocyte precursor cell differentiation. All remain at preclinical stages.

Beyond direct antioxidant and pathway-targeted approaches, modulating the inflammatory cascade itself represents a distinct and complementary therapeutic avenue, given that neuroinflammation both amplifies oxidative injury and operates through partially independent molecular networks (125). Shin et al (30) demonstrated that TNF-α antagonist attenuated lipopolysaccharide-induced white matter injury in neonatal rats. Prasad et al (126) comprehensively reviewed anti-inflammatory therapies, examining agents including TNF-α antagonists (such as etanercept), minocycline, COX-2 inhibitors, melatonin and IL-1 receptor antagonists. While numerous agents show preclinical efficacy, clinical translation has been limited by the complexity of the inflammatory cascade and the difficulty of timing interventions appropriately. The inflammatory cascade involves sequential activation of multiple mediators: peripheral immune responses trigger central neuroinflammation characterized by proinflammatory cytokine release (TNF-α, IL-1β, IL-6), microglial activation and reactive gliosis, which in turn activate NADPH oxidase enzymes generating ROS, establishing a feed-forward loop between OS and inflammation (8,30). The temporal heterogeneity of microglial responses, shifting from pro-inflammatory to anti-inflammatory gene expression over time, further complicates therapeutic targeting (80). Different anti-inflammatory agents require distinct therapeutic windows: etanercept administered at three days after hypoxia-ischemia ameliorates white matter injury (30); early postnatal caffeine treatment confers improved neuroprotection than late treatment (76,120,121); minocycline requires early post-insult administration to target secondary neuroinflammation; and dexamethasone exhibits a developmentally sensitive therapeutic window (126). These temporal constraints, combined with the difficulty of predicting the onset and progression of inflammation in individual preterm infants, pose substantial barriers to clinical translation.

IL-1 has emerged as a promising target, with Takahashi et al (127) demonstrating that pharmacological IL-1 receptor blockade suppressed lipopolysaccharide-induced neuroinflammation in preterm fetal sheep. Van Steenwinckel et al (80) identified decreased microglial Wnt/β-catenin signaling as a driver of pro-inflammatory activation. Bokobza et al (128) demonstrated that miR-146b protects against microglia-induced hypomyelination, suggesting that microRNA-based therapies could modulate neuroinflammation more selectively. The temporal heterogeneity of microglial responses, emphasized by Bernis et al (81), implies that anti-inflammatory interventions must be precisely timed to avoid interfering with beneficial microglial functions.

Oxygen management is paramount. Rantakari et al (129) demonstrated that both low and high oxygen saturations are associated with adverse outcomes. The targeted oxygen for resuscitation of preterm infants and their developmental outcomes (TORPIDO) 30/60 trial randomized 1,469 preterm infants and found no significant association between initial FiO2 concentration (0.6 vs. 0.3) for resuscitation and death or brain injury at 36 weeks' corrected gestational age (130). Engur et al (131) provided mechanistic insight by demonstrating that supplemental oxygen alters the pentose phosphate pathway through sirtuin signaling.

Developmental care practices have been codified into neuroprotective care bundles. Milette et al (132) published guidelines for institutional implementation of developmental neuroprotective care. Murthy et al (133) evaluated a neuroprotection care bundle, reporting reduced intraventricular hemorrhage rates following implementation. Travis et al (134) demonstrated that skin-to-skin holding is associated with improved white matter microstructure. Nutritional strategies have also been investigated. Belfort and Inder (135) reviewed the relationship between human milk and brain development, concluding that breastfeeding is associated with improved outcomes. Brandt et al (136) demonstrated that human milk oligosaccharides improve white matter and interneuron development in a rat model. These strategies offer low cost and wide availability, but definitive evidence of neuroprotective efficacy remains limited.

Multimodal neuromonitoring has become a cornerstone of neurocritical care in preterm infants, enabling early detection of physiological instability and evolving brain injury before irreversible damage occurs. El-Dib et al (137) provided a comprehensive framework for integrating multiple monitoring modalities, including aEEG, near-infrared spectroscopy (NIRS) and magnetic resonance imaging (MRI), to detect cerebral dysfunction across different physiological domains. Among these, aEEG has emerged as a particularly valuable bedside tool. Variane et al (138) demonstrated that early aEEG monitoring effectively identifies neonates at high risk for brain injury, facilitating timely intervention. Griesmaier et al (139) further validated the prognostic utility of aEEG by demonstrating significant associations between aEEG patterns and MRI defined injury severity at term equivalent age, establishing a correlation between functional and structural assessments that enhances clinical interpretation. Beyond electrophysiological assessment, Lin et al (140) integrated aEEG with General Movements assessment, showing that combining neurophysiological and motor function evaluations provides a more comprehensive prediction of neural recovery than either modality alone. Despite these advances, the clinical value of multimodal monitoring ultimately depends on the availability of effective, evidence-based interventions to act upon the detected abnormalities. The integration of real time neuromonitoring data with AI-driven decision support systems and biomarker panels represents a promising avenue for translating early detection into improved outcomes, although prospective validation in large cohorts is still needed.

Cell-based therapies represent an emerging frontier. Drommelschmidt et al (141) demonstrated that mesenchymal stem cell-derived extracellular vesicles ameliorate inflammation-induced preterm brain injury. Thomi et al (142) showed that intranasally administered exosomes reduce microglia-mediated neuroinflammation. Ahn et al (143) conducted a phase I dose-escalation clinical trial of mesenchymal stem cells for severe intraventricular hemorrhage, demonstrating feasibility and safety. However, the Cochrane systematic review by Romantsik et al (144) sought to include randomized controlled trials comparing stem cell-based interventions vs. control for the prevention (in infants <28 weeks' gestation, ≤24 h of age, without ultrasound diagnosis) or treatment (in infants <37 weeks' gestation, with ultrasound-confirmed germinal matrix-intraventricular hemorrhage or encephalopathy of prematurity) of these conditions. However, no completed studies met the inclusion criteria; only three ongoing trials were identified. The authors concluded that no evidence is currently available to evaluate the benefits and harms of stem cell-based interventions for treatment or prevention of these conditions in preterm infants (144). Tscherrig et al (145) identified specific miRNAs from mesenchymal stromal cell extracellular vesicles that rescue white matter injury. Vaes et al (146) demonstrated that modifying the secretome prolongs the regenerative treatment window. Despite extensive preclinical literature, the gap between experimental promise and clinical evidence remains substantial, necessitating rigorous trial design before cell-based therapies enter routine practice. Viewed through the OS epigenetic lens, the therapeutic landscape becomes coherent: Antioxidants and oxygen management aim to reduce the upstream trigger; HDAC inhibitors and miRNA-based approaches target the epigenetic machinery that translates oxidative insults into persistent cellular dysfunction; and cell-based therapies seek to modulate the downstream cellular responses that determine repair vs. injury. This conceptual taxonomy underscores that effective neuroprotection will likely require multimodal strategies that address multiple nodes of the integrated OS epigenetic cellular pathway, timed to the developmental window of vulnerability.

In synthesizing the neuroprotective evidence, several points of consensus emerge: Melatonin and erythropoietin have been the most extensively studied agents, both showing favourable safety profiles but inconsistent efficacy; caffeine has a strong preclinical rationale and established clinical safety; and cell-based therapies hold substantial promise but remain experimental. However, major controversies and unresolved questions persist. The most critical is the translational gap between consistently positive preclinical findings and largely neutral or inconclusive clinical trials, a discrepancy that may reflect differences in timing, dosing, patient selection, or the inherent limitations of animal models. There is also no consensus on whether combination therapy (such as antioxidants plus anti-inflammatory agents) is superior to monotherapy, or whether interventions should be targeted to specific injury subtypes. Emerging preclinical evidence has begun to explore multimodal combination strategies. For instance, combination therapy with sulforaphane (an Nrf2 activator with NF-κB inhibitory properties) and deferiprone (an iron chelator) has been shown to mitigate M1 microglial infiltration, suppress inflammation and reduce ferroptosis in a neonatal rabbit model of intraventricular hemorrhage (147). Combined omega-3 polyunsaturated fatty acid and folic acid supplementation has demonstrated enhanced neuroprotection against neonatal hypoxic-ischemic brain injury through synergistic anti-inflammatory and anti-apoptotic mechanisms (148). The combination of erythropoietin and melatonin has shown sustained repair of functional deficits in a rat model of cerebral palsy following preterm brain injury (149), with ongoing clinical trials evaluating the safety of this combinatorial approach in very preterm infants with intraventricular hemorrhage or white matter injury. N-acetylcysteine combined with vitamin D has been shown to decrease OS in plasma and central nervous system and is associated with favorable developmental outcomes in neonatal hypoxic-ischemic encephalopathy models (150). Furthermore, allopurinol (a xanthine oxidase inhibitor) is currently being evaluated in the phase III ALBINO trial as an adjunct to therapeutic hypothermia for hypoxic-ischemic brain injury (17). These emerging combination approaches, although primarily at preclinical or early clinical stages, highlight the potential for multimodal strategies targeting both oxidative and inflammatory pathways to achieve superior neuroprotection compared with monotherapy. Furthermore, the optimal window for intervention, before, during, or after the oxidative insult, has not been established for most agents. Future research must prioritise adequately powered, biomarker-stratified trials with long-term neurodevelopmental follow-up to resolve these uncertainties and translate preclinical promise into clinical reality.

The clinical evidence base for neuroprotective agents is strongest for caffeine [network meta-analysis of randomized controlled trials (RCTs)], erythropoietin (PENUT RCT and cohort studies) and oxygen titration (TORPIDO RCT), whereas melatonin, HDAC inhibitors, PPAR-γ agonists and cell-based therapies remain largely at preclinical or early-phase clinical stages (see Table II for detailed evidence levels and model systems). Several systematic reviews and meta-analyses have amalgamated the evidence for neuroprotective agents in preterm infants (105,122,144). A Cochrane review on stem cell-based interventions concluded that no completed RCTs met inclusion criteria, highlighting the substantial gap between preclinical promise and clinical evidence (144). Network meta-analysis data suggest caffeine may be the most promising agent for reducing neurodevelopmental impairment and cerebral palsy risk (122), although this finding derives from indirect comparisons and requires confirmation in prospective trials. For melatonin, a systematic review consistently identified small sample sizes and dosing heterogeneity as major limitations precluding definitive conclusions (105). Formal guidelines from international bodies provide specific recommendations for neuroprotective interventions. The American Academy of Pediatrics recommends therapeutic hypothermia for neonates with moderate-to-severe hypoxic-ischemic encephalopathy born at ≥36 weeks' gestation, initiated within 6 h of birth and continued for 72 h (151). UK national guidelines recommend antenatal magnesium sulfate for mothers in preterm labour (<30 weeks' gestation) to reduce the risk of cerebral palsy (152). The European Resuscitation Council 2025 guidelines provide evidence-based recommendations for newborn life support, including neuroprotective strategies during transition at birth (153). Additionally, structured neuroprotection care bundles have been shown to markedly reduce death or severe brain injury in extremely preterm infants (133). Despite these established guidelines, formal consensus from the American Academy of Pediatrics or European bodies specifically addressing pharmacological neuroprotection for preterm brain injury beyond therapeutic hypothermia and magnesium sulfate remains limited.

AI has substantially enhanced neuroimaging interpretation accuracy and efficiency. Ahmad et al (154) demonstrated that deep learning models could classify cranial ultrasound images for brain injury detection in very preterm infants, achieving performance comparable to expert clinicians. Chen et al (155) developed an automated neonatal nnU-Net brain MRI extractor trained on a large multi-institutional dataset, showing robust generalizability across different clinical settings. For subtle pathology, Estermann et al (156) introduced CACTUS, a multiview classifier for detecting punctate white matter lesions in cranial ultrasound volumes, addressing a diagnostically challenging task with significant neurodevelopmental implications. These advances suggest AI-based image analysis has substantial potential to enhance neuroimaging interpretation, although protocol variability across institutions, limited external validation and the absence of prospective clinical trials remain important barriers to clinical deployment.

Continuous neuromonitoring generates vast physiological data ideally suited for AI-based analysis. Wang et al (157) conducted a landmark cohort study demonstrating that quantitative features from early amplitude-integrated electroencephalography predicted long-term neurodevelopmental outcomes in extremely preterm infants, with automated feature extraction performing equivalently to combined qualitative and quantitative assessments. Abbasi et al (158) developed deep learning models for seizure detection using wavelet-scalogram approaches in preclinical models, subsequently refined by Roozbehi et al (159) using transformer-based architectures. Ashoori et al (160) applied machine learning to NIRS signals, demonstrating that data-driven definitions of prolonged desaturation outperformed traditional threshold-based approaches for detecting intraventricular hemorrhage. Hibner et al (161) extended this concept by developing a multimodal approach combining echocardiography, NIRS and electrical cardiometry, showing integrated signal analysis provides complementary information unavailable from any single modality.

AI-based predictive models have shown considerable promise for early risk stratification. He et al (162) developed PBIPred, an explainable machine learning model using seven clinical features that achieved an area under the curve of 0.8229 for predicting preterm brain injury, with Shapley Additive Explanations rendering predictions interpretable at individual patient level. For intraventricular hemorrhage specifically, Yang et al (163) conducted a nationwide multicenter study demonstrating robust prediction using machine learning. Han et al (164) incorporated time-series analysis of physiological data, showing that temporal dynamics improve prediction compared with static variables alone. For white matter injury, Song et al (165) developed two risk assessment models in extremely preterm infants, demonstrating that ensemble methods may outperform single algorithms. Shu et al (166) applied machine learning to predict mortality and multiple morbidities simultaneously, showing comprehensive outcome prediction is feasible using routinely collected data.

Long-term neurodevelopmental prognostication represents the most clinically challenging AI application. Li et al (167) demonstrated that functional connectivity measures predict outcomes in preterm infants without severe brain injury, suggesting subtle network alterations captured by AI analysis identify risk even with normal conventional imaging. Liu et al (168) used graph convolutional neural networks to predict brain age, showing deviations from normative maturation trajectories explain outcomes more accurately than static injury assessments. For cerebral palsy, Marinelli et al (169) compared predictions at ages 2 and 10 years, revealing that early predictions require refinement as children mature. Yuan et al (170) developed a nomogram for predicting intellectual disability, providing a practical tool for early intervention targeting. Despite these advances, two recent systematic reviews identify critical limitations in AI applications for neonatal care. Ortega-Leon et al (171) reported that >70% of studies lacked a validation procedure, highlighting a critical gap in methodological rigor in their systematic review of machine learning techniques for predicting neurodevelopmental impairments in premature infants. Similarly, Martinez-Millana et al (172) found that among AI studies in the neonatal intensive care unit setting, the majority lacked external validation, limiting generalisability to new clinical settings. These limitations suggest that while AI applications have progressed from proof-of-concept to clinical validation, generalisability remains the principal barrier to widespread adoption.

Despite the proliferation of AI studies, several fundamental limitations impede clinical translation. First, external validation, the critical step for demonstrating generalizability, is performed in <30% of published models and performance consistently declines in cross-institutional settings (154,173). Second, heterogeneity in imaging protocols, EEG configurations and clinical data definitions across centers creates substantial challenges for model generalization (155,157). Third, the ‘black-box’ nature of numerous deep learning algorithms, combined with poor clinical interpretability, undermines clinician trust and hinders bedside adoption (162). Fourth, the absence of prospective impact studies, trials demonstrating that AI-assisted decision-making actually improves patient outcomes rather than merely achieving statistical accuracy, represents the most critical evidence gap (173). These limitations suggest that while AI applications have progressed from proof-of-concept to clinical validation, generalizability and implementation remain the principal barriers to widespread adoption.