The medial prefrontal cortex as an integrative hub in chronic pain: network mechanisms and the enabling role of artificial intelligence

Abstract

Chronic pain is recognized as a disorder of distributed brain networks rather than the consequence of persistent nociceptive input. Among these networks, the medial prefrontal cortex (mPFC) is a key integrative hub linking sensory processing with affective, cognitive, and stress-related dimensions of pain. Evidence from neuroimaging, neurochemical, and longitudinal studies indicates that mPFC dysfunction contributes to impaired top-down modulation, altered emotional regulation, and the persistence of pain states. Nevertheless, these findings should be interpreted within a system-level framework, as mPFC activity reflects network reorganization rather than serving as an isolated or validated clinical biomarker. Moreover, the generalizability of mPFC-centered models is limited across patient populations, including children and those with psychiatric comorbidities or cognitive impairment. This editorial critically examines the neurobiological basis of mPFC-centered network dysfunction in chronic pain and discusses its implications for translational research, with a key focus on artificial intelligence (AI). These technologies are framed not as a near-term clinical solution but as enabling and exploratory methods for integrating multimodal data and modeling complex brain–behavior relationships. Emerging generative AI approaches, agent-based models, and digital twins can also be implemented as conceptual tools for hypothesis generation and in silico exploration of individualized network dynamics, rather than as established clinical applications. Although AI-based approaches may accelerate hypothesis generation and the identification of latent network-level patterns, their clinical relevance is currently constrained by key methodological challenges, including limited generalizability, imperfect phenotypic classification, the absence of robust ground truth for pain, and the need for extensive external and longitudinal validation.

Trial registration Not applicable.

Background

The medial prefrontal cortex (mPFC) is a pivotal regulator of chronic pain with a central role in integrating sensory, affective, and cognitive dimensions of nociception [1]. Converging evidence from functional neuroimaging, neurochemical, and clinical studies suggests that the mPFC plays a crucial role in top-down modulation of pain perception and its emotional salience. In individuals who underwent transcutaneous electrical stimulation to induce secondary hyperalgesia, for instance, functional magnetic resonance imaging (fMRI) studies demonstrated that higher mPFC activation is associated with reduced hyperalgesic fields, suggesting a protective modulatory role [2, 3]. Importantly, mPFC activity was also inversely related to pharmacological responsiveness. Specifically, individuals exhibiting reduced descending pain modulation, presumably reflecting impaired mPFC-mediated top-down control, demonstrated a greater therapeutic response to systemic lidocaine [4]. On these bases, mPFC activity could be a candidate neurobiological correlate of analgesic responsiveness. Nevertheless, rather than being considered as an isolated biomarker, this brain area can be conceptualized as a system-level integrative hub linking affective, cognitive, and stress-related dimensions of chronic pain [5]. From this perspective, a range of modalities, such as biochemical, functional, and connectivity-based findings, converge on the mPFC in chronic pain states and can be investigated for a better understanding of pain mechanisms.

In this context, emerging computational approaches, particularly artificial intelligence (AI), can offer powerful tools for investigating within these complex networks and the multiple features that can be generated. Nonetheless, for pain research, their current role remains primarily exploratory, and, to date, AI should be regarded as an enabling methodology rather than a primary translational claim. Its output remains dependent on incompletely defined phenotypes, indirect surrogate labels, and limited external validation. Consequently, the central challenge lies in the formal definition of the modeling framework, the principled selection and representation of input variables (features), and the precise specification of clinically meaningful target tasks [6].

Accordingly, this editorial highlights the conceptual and methodological gap between mechanistic insights and clinical translation and advances a clear position: mPFC-centered network frameworks, supported by AI as an enabling methodology, are best understood as exploratory tools for hypothesis generation and system-level insight into chronic pain, not as clinically actionable instruments for near-term translation.

Searching for multimodal features

Collectively, biochemical, functional, and connectivity-based findings can provide multilevel descriptors useful to inform AI-based models of pain-related network dysfunction and support future translational research. Within this multilevel context and the AI-based multimodal strategy, biochemical features can be key substrates for integrative modeling. For example, chronic pain is associated with reduced glutamate levels in the mPFC, as determined by magnetic resonance spectroscopy. These neurochemical changes align with elevated anxiety, depression, and avoidance behaviors and suggest that mPFC dysfunction not only undermines pain modulation but also contributes to the emotional comorbidities that sustain the chronic pain state [7]. Other investigations highlighted the critical role of dopaminergic modulation within mPFC-centered circuits. Dopamine (DA) inputs from the ventral tegmental area (VTA) to the mPFC regulate cortical plasticity, aversive learning, and motivational states, all of which are integral to chronic pain processing. In neuropathic pain models, phasic activation of mesocortical DA projections to the mPFC reduces mechanical hypersensitivity and induces context-dependent preference, indicating engagement of endogenous pain-modulatory mechanisms. At the circuit level, dopaminergic signaling enhances the activity of mPFC neurons projecting to the ventrolateral periaqueductal gray (PAG), a key node of the descending pain control system [8]. Consequently, mPFC dysfunction in chronic pain reflects not only impaired excitatory balance but also altered neuromodulatory control, with dopaminergic signaling representing another biologically meaningful feature for a multimodal, circuit-informed modeling approach. In this context, at the microcircuit level, this pain-associated disruption of the excitation–inhibition balance within mPFC is also mediated by an altered function of inhibitory GABAergic interneurons. In particular, fast-spiking parvalbumin-positive (PV+) interneurons, which play a critical role in coordinating prefrontal network oscillations and top-down control, exhibit region- and layer-specific maladaptive plasticity in chronic pain states. Experimental models show that enhanced feed-forward inhibition mediated by PV+ interneurons can suppress the activity of mPFC pyramidal neurons projecting to the PAG, thereby weakening descending inhibitory control [9]. Conversely, loss of inhibitory synaptic input in other mPFC subregions contributes to pyramidal hyperexcitability and affective dysregulation [2]. These findings demonstrate that mPFC dysfunction in chronic pain is not solely driven by altered excitatory transmission, but also by interneuron-specific GABAergic remodeling that shapes network-level output. Furthermore, the cortical control of pain through fine-tuning of inhibitory microcircuits that gate descending modulatory output also involves µ-opioid receptors (MORs) within prefrontal microcircuits. MORs located on GABAergic interneurons, rather than on glutamatergic pyramidal neurons, appear to mediate stress-induced analgesic effects by disinhibiting mPFC neurons projecting to the PAG [10].

Beyond neurometabolic signatures and pathways, functional neurophysiological measures derived from resting-state and task-based fMRI provide further multivariate descriptors of altered brain dynamics in chronic pain. In particular, resting-state fMRI studies have identified convergent alterations across complementary measures of intrinsic brain activity and connectivity, such as amplitude of low-frequency fluctuations, regional homogeneity, and large-scale functional connectivity, capturing both local signal dynamics and distributed network reorganization [11]. Longitudinal neuroimaging has revealed a profound reorganization during the transition from acute to chronic pain. Specifically, these investigation strategies indicated a relative attenuation of activity within primary sensory cortices, accompanied by increased engagement of prefrontal–limbic regions, including the mPFC and amygdala. This reorganization preferentially involves nodes of the Default Mode Network (DMN), mostly the mPFC and precuneus, together with altered coupling to salience- and limbic-related regions, reflecting a shift from stimulus-driven sensory processing toward internally oriented, affect-laden representations of pain [11]. Strikingly, these shifts occur even without measurable changes in pain intensity and underscore the pivotal role of affective and cognitive modulation in chronic pain perception [12]. From a modeling perspective, this critical dissociation between subjective pain intensity and large-scale network dynamics underscores the limitations of univariate clinical labels, although suggesting the use of multivariate, network-informed representations. In this area, AI-based multimodal approaches may be particularly informative at an exploratory level.

Without implying a fixed hierarchy, the reciprocal connectivity between the mPFC and amygdala forms a crucial regulatory loop. Anatomically and functionally bidirectional, the mPFC exerts inhibitory control over amygdala output via excitatory projections to intercalated cells, which suppress activity in the latero-capsular division of the central nucleus, a region implicated in affective pain processing. In turn, the basolateral amygdala sends glutamatergic projections to the mPFC, creating a negative feedback circuit. In chronic pain models, hyperactive amygdala output dampens mPFC activity, impairing cognitive control over nociception, enabling amplification of negative effects. This represents a maladaptive cycle that shares partially overlapping circuit-level features with post-traumatic stress disorder (PTSD) and major depression, albeit with disorder-specific dynamics, symptom expression, and boundary conditions [13]. This convergence between pain and affective disorders is not coincidental. However, this phenomenon should be interpreted as reflecting shared vulnerabilities at the level of large-scale network dysfunction rather than a common or interchangeable pathophysiology across these conditions. Acute nociception, anxiety, and depression share an adaptive evolutionary function, limiting exploratory behavior to enhance survival. Additionally, when persistent, these states shift from protective to maladaptive, maintained by dysfunctional neurobiological circuits [12]. In cancer-related chronic pain and other severe pain syndromes, functional disintegration between the mPFC and limbic structures, including the amygdala, insula, and hypothalamus, has been associated with diminished top-down regulation and the emergence of automatic, dysregulated emotional responses, further highlighting the role of network-level dysfunction over isolated regional abnormalities [14, 15].

The mPFC plays a central role in stress physiology via its influence on the hypothalamic–pituitary–adrenal (HPA) axis and the autonomic nervous system [16]. It modulates the termination — rather than the peak — of glucocorticoid release, aided by dense glucocorticoid receptor expression. This regulation is disrupted, for instance, in chronic migraine, where altered mPFC–hypothalamus connectivity within the DMN correlates with subjective pain intensity, potentially reflecting a loss of cognitive reappraisal capacity [17, 18].

Toward AI-based multimodal integration

Multimodal AI integration, combining machine learning and deep learning, offers the potential to decode these multifaceted neurobiological signatures of chronic pain. For example, these innovative strategies can decode high-dimensional brain–behavior data, enabling automated, objective pain assessment (APA) and adaptive therapy monitoring. Nevertheless, while research has shown that such approaches may offer significant results for APA [19, 20], most importantly, this multidisciplinary approach can be implemented for neurophysiological research and to inform targeted neuromodulatory interventions, such as those aimed at restoring mPFC–limbic–hypothalamic balance. Therefore, these methods have the potential to transcend pain research, illuminating the broader landscape of affective and cognitive dysregulation across psychiatric and neurological disorders.

Concerning methodology, AI-based frameworks can integrate heterogeneous data sources from neuroimaging, neurophysiological signals, behavioral measures, and clinical variables, which capture complementary aspects of chronic pain. This integration can be achieved through feature-level fusion, in which modality-specific representations are combined into a shared latent space, or through model-level and decision-level fusion strategies that preserve modality-specific structure while enabling joint inference [21]. Deep learning architectures, such as convolutional neural networks for imaging data, recurrent or temporal convolutional networks for time-series signals, and graph-based models for network connectivity, allow the extraction of hierarchical representations that capture both local and distributed patterns of brain activity [22]. Interestingly, in this context, representation learning and multimodal embedding techniques can be used to identify latent network-level signatures associated with pain-related dysregulation, while dimensionality-reduction and attention-based mechanisms help prioritize informative features across modalities. Additionally, these approaches can enable the modeling of complex, non-linear interactions between brain networks, physiological responses, and behavioral outcomes that are not readily accessible through conventional statistical analyses. When applied longitudinally, such models can also capture dynamic changes over time, supporting the study of pain chronicization, treatment response, and state-dependent fluctuations. In this framework, generative AI approaches, agent-based models, and digital twins may further support hypothesis generation by simulating individualized network dynamics, exploring counterfactual scenarios, and testing how perturbations of mPFC-centered circuits could influence pain-related trajectories over time [23].

Generalizability of mPFC-centered models

Despite the translational relevance of mPFC-centered network dysfunction and its potential interrogation through multimodal AI approaches, it is mandatory to define the boundary conditions under which such frameworks may be applied. The applicability limitations extend to different patient populations. They include, for instance, children, older adults, patients with psychiatric comorbidities, long-standing chronic pain, cancer-related pain, cognitive impairment, and individuals under chronic pharmacological treatment. In these populations, prefrontal functioning is variably affected by aging, comorbidity, neuroplasticity, and contextual factors, thereby limiting the generalizability of mPFC-centered models. In pediatric populations, for instance, interpretative challenges in mPFC–pain research must consider issues related to the developmental stage and neurobiological maturation. Indeed, most mPFC-centered models of chronic pain have been derived from adult populations, and their extension to pediatric pain cannot be assumed. Emerging evidence from large-scale developmental neuroimaging studies suggests that altered brain signatures may precede the onset of multisite pain in children, including trends toward reduced mPFC activity and disrupted connectivity within default mode, salience, and sensorimotor networks. However, these findings also highlight that pediatric pain is embedded within a dynamic developmental context characterized by ongoing prefrontal maturation, heightened network plasticity, and age-dependent affective regulation. In this view, pediatric pain should be regarded as a distinct neurobiological condition, currently supported by limited empirical evidence. Therefore, given the gaps of region-centric interpretations, in this setting, pain is a critical research gap rather than an established translational application of adult-derived mPFC frameworks [4]. It reinforces the need for the development of informed, network-level frameworks.

Methodological and applicative limits of AI-based pain research

Although mPFC-related mechanisms and AI methodologies hold substantial conceptual promise, available clinical evidence remains preliminary. Several critical methodological issues currently limit the translational relevance of AI applications in pain research. These include bias and confounding related to cohort composition, limited generalizability across clinical populations and acquisition settings, and the nontrivial risk of data leakage in multimodal analytical pipelines [24]. Moreover, the definition of a valid ground truth for pain remains intrinsically problematic, as pain is a multidimensional, subjective, and context-dependent experience that cannot be fully captured by single labels or proxy outcomes. Progress toward reproducible and clinically meaningful evidence will likely depend on refining pain phenotyping and adopting more precise and biologically informed classifications. Importantly, the development of robust and clinically meaningful AI models in this domain requires extensive and time-consuming validation processes that are still largely lacking. These include multi-center data collection, harmonization of acquisition and annotation protocols, external validation across independent cohorts, longitudinal analyses to assess temporal stability, and systematic evaluation of model interpretability and failure modes [25]. Without these mandatory steps, claims regarding clinical or translational utility remain premature.

Furthermore, beyond the methodological limitations of AI, it is crucial to delineate its practical applicability, scope of use, and intended clinical objectives within this highly complex clinical context. Therefore, within this framework, mPFC-informed features may add value not by replacing symptom-based assessments, but by complementing them through the identification of latent network-level patterns and sources of interindividual variability that are otherwise inaccessible to conventional approaches.

In conclusion, mPFC-centered network dysfunction offers a compelling system-level framework for understanding the persistence and complexity of chronic pain. However, its value currently lies in mechanistic insight rather than clinical determinism. Within this context, AI should be regarded as an enabling and exploratory methodology for integrating, for example, heterogeneous data and revealing latent network-level patterns. Nevertheless, these technologies are not a surrogate for clinical judgment or a source of actionable biomarkers. Bridging the gap between network neuroscience and clinical practice will require rigorously designed, longitudinal, and developmentally informed studies, alongside transparent validation frameworks. Therefore, until such conditions are met, mPFC-centered and AI-enabled approaches should be interpreted as conceptual tools to guide hypothesis generation and refine pain models, rather than as instruments for immediate clinical translation.

Abbreviations

AI

Artificial intelligence

APA

Pain assessment

DA

Dopamine

DMN

Default mode network

fMRI

Functional magnetic resonance imaging

HPA

Hypothalamic–pituitary–adrenal

MOR

µ-opioid receptor

mPFC

Medial prefrontal cortex (mPFC)

PAG

Periaqueductal gray

PTSD

Post-traumatic stress disorder

PV +

Parvalbumin-positive

VTA

Ventral tegmental area

Author contributions

Conceptualization, M.C., M.M., and A.V.; resources, A.V.; writing—original draft preparation, M.C., M.M., and A.V.; writing—review and editing, M.C., M.M., and A.V.; funding acquisition, A.V. All authors have read and agreed to the published version of the manuscript.

Funding

This work was supported by the Italian Ministry of Health with “Current Research funds”.

Data availability

Not applicable.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.