Application of generative artificial intelligence (AI) to support pain neuroscience research in persons who are pregnant

The brain mechanisms responsible for both pain perception and pain modulation in persons who are pregnant remain to be fully explored. Generative artificial intelligence methods have the capacity to interface with brain imaging technologies to dramatically improve our understanding of pain neuroscience in maternal health.

Holmes et al. discuss using generative artificial intelligence applied to brain imaging data to improve maternal health. A better understanding of pain perception and modulation could improve pain management during and after pregnancy.

Generative artificial intelligence (AI) and other AI domains have the potential to either improve or present problems when incorporated into studies of pain neuroscience in women’s health. Pain research has historically overlooked the impact of sex or female transitional phases on pain perception due to the “complexities” of hormonal fluctuations¹. Even fewer investigations have addressed the neurological components and antecedents of pain. However, as we discuss in this Comment, pregnancy is an area where our understanding of pain perception and pain neuroscience could be significantly improved through the appropriate application of machine learning/AI, particularly as it relates to understanding the brain mechanisms underlying pain perception.

Pain during and after pregnancy

Pregnancy is a highly dynamic period associated with pain and discomfort. Pain has been considered a difficult topic during pregnancy, as there are concerns that there may be drug interactions with the fetus, and there is poor knowledge surrounding the physiological systems in the mother. However, pain is common. For example, prior research has reported that 45% of all pregnant women and 25% of postpartum women suffer from pregnancy-related pelvic girdle pain (PGP)^2,3. Indeed, PGP can negatively impact activities of daily life up to a decade after pregnancy⁴, suggesting pain experienced during pregnancy has long-term implications. Moreover, around 25% of all pregnant women and around 5% of all women in the postpartum period suffer from lumbopelvic pain sufficient to require some form of medical assistance⁵. For those who have multiple pregnancies, pain reporting during early⁶ or in previous pregnancies⁷ are predictors of pain during later and subsequent pregnancies. These findings underscore two critical points: (1) pain resulting from pregnancy is not isolated to the time period during which a person is pregnant, and (2) pregnancy-related pain appears to have a sensitization component. It is clear that more knowledge is required to fully understand the dynamics of pain and pain behaviors during pregnancy.

Pain neuroscience provides a mechanistic understanding of pain perception and integrates the concept of modulation through which pain signals can be amplified or attenuated. For example, ascending pain pathways where noxious stimuli are transferred from the periphery through the spinal cord towards the brain integrate activity in higher order regions such as the dorsolateral prefrontal cortex (cognitive) and amygdala (emotional), as well as secondary regions such as the hypothalamus⁸. Alternatively, descending pain modulation occurs across networks that may include the periaqueductal gray and ventral medulla⁹ to modify the pain-nociceptive relationship. The concept of pain modulation can be achieved through activities such as movement¹⁰ and cognitive control¹¹, making physical activity a critical component towards healthy pain processing in the general population. In the context of pregnancy, lower levels of physical activity are associated with an increase in back pain and pelvic cavity pain⁶. Data suggest that physical activity can lead to a decrease in the severity of different forms of pain during pregnancy¹² and postpartum¹³. Importantly, this and other work to date have focused on pain perception and not the underlying neurological mechanisms. Although there is published evidence of altered brain structure and function^14,15, no research to date has evaluated brain representations of pain during pregnancy, or the impact of physical activity on brain structure or brain function during pain exposure. This is a notable limitation as pain modulation processes can create disparity between the extent of a nociceptive stimulus and active pain perception, which can leave women vulnerable to chronic pain conditions through sensitization.

Utilizing neuroimaging

There has been growing interest and capability to evaluate brain health using neuroimaging in persons who are pregnant. Recent findings have shown, for example, that during pregnancy, there may be a gradual loss of gray matter in the brain with a slight return (though not to pre-pregnancy levels) after birth. Analyses focused on gray matter using MRI on 179 pregnant women throughout pregnancy found a 4.9% GM volume decrease during pregnancy that was followed by a 3.4% increase from late pregnancy to 6 months postpartum¹⁶. A similar finding was observed in a more recent publication showing changes to gray and white matter, as well as specific changes in pain-related networks (i.e., salience network) in the brain¹⁵. Changes in gray matter volume have been correlated with maternal mental health¹⁶ highlighting a relationship during pregnancy that links brain changes to externalized behavior. Available research using fMRI has shown changes to the default mode network, which the authors proposed could relate to maternal behaviors¹⁴. Research using techniques such as transcranial Doppler in persons who are pregnant has found changes in blood flow¹⁷, which may suggest that functional brain changes could be mediated by a form of sensitization and that perhaps traditional neurovascular coupling relationships are perturbed in this population. No research to date has directly evaluated brain structure or function in persons who are pregnant as it relates to pain. These structural and functional observations demonstrate the need for focused research on pain and women’s health that evaluates women-specific conditions such as pregnancy. Efforts aimed at building databases of brain imaging scans, such as that by Dr. Emily Jacobs, with the Maternal Brain Project (https://jacobs.psych.ucsb.edu/research/maternal-brain-project), are a great start and require much more participation.

Applications of generative AI to brain imaging data

Generative AI enables the ability to create novel products from learned attributes of input datasets. To date, these efforts have been popularized in terms of text and image generation with programs such as GROK and CHAT-GPT; however, efforts in the field of neuroscience have shown that such tools can be applied towards brain imaging. Early efforts have shown that such technologies can generate novel synthetic sets of MRI images (e.g., T1-weighted and T2-weighted)¹⁸, and CT scans¹⁹ and can be applied to generate whole brain images as well as novel images of tumors²⁰. These synthetic datasets can be applied to better develop functions such as tumor segmentation or to create novel images to complement missing data, such as to create a T2 image from a T1 image. With sufficient data, applications can infer missing data and create models for proposed brain states. Synthetic images can be created from large databases of the same modality (e.g, T1-weighted to T1-weighted images or from different modalities (e.g., T1-weighted to T2-weighted images). The Pix2Pix architecture has been applied towards generating synthetic images to support tumor segmentation (via the DeepMedic algorithm), where they found that the Pix2Pix architecture was a viable means of supporting brain tumor segmentation efforts when one input image (e.g., T1) is missing²¹. As such, these tools could, in theory, help create model datasets of brain images of pregnant persons or create cross-modal datasets for persons unable to complete multiple image sequences. Indeed, using techniques such as generative adversarial networks (GANs), authors have created sample datasets that have been contested against authentic datasets, showing reliability²². A review of the application of GAN-synthesized images highlights their application in roles such as data augmentation, image translation, registration, super-resolution, denoising, motion correction, segmentation, reconstruction, and contrast enhancement²³. However, in women's health, these datasets cannot reliably be used directly as they may contain mixed populations, which often consist of men. These populations (male) cannot experience issues such as endometrial pain, menstrual pain, or nerve compression injuries resulting from pregnancy. Brain imaging supports the impact of different types of pain on brain region representation²⁴, and therefore, we cannot consider all pain forms equivalent. Despite the progress of AI, there remains a dire need for basic science and clinical research in women’s pain neuroscience to fuel AI and machine learning models generating such novel image sets.

Concluding remarks and recommended next steps

Pregnancy is an ideal application for generative AI initiatives to accelerate pain neuroscience research. The strengths of generative AI are speed and the generation of novel images; however, they rely on existing datasets to make novel predictions. Previous applications of AI and machine learning in pregnancy have focused on predicting conditions such as pre-eclampsia, fetal movement and congenital anomalies, as well as pregnancy complications²⁵. However, there have been no investigations into brain health relating to pain neuroscience. To effectively apply generative AI in women’s pain neuroscience and improve Women’s health research, we must focus on three directions. The first is to continue to build out data on women during and after pregnancy that includes information relating to nervous system (peripheral and central) health. Non-synthetic data obtained from human participants will be the most informative for health initiatives, and so highly collaborative environments (i.e., Maternal Brain Project) will be essential to align and synergize data collection efforts. The second direction is to undertake specific research on pain perception and nociceptive processing in women in the stages leading up to, during, and after pregnancy, to identify specific mechanisms of pain and pain modulation in these cohorts. Finally, machine learning and AI methods should be used that leverage existing images to translate cross-modal images, and datasets to create novel synthetic images where no equivalents were available before. By following these three directions, we can most effectively use the data at hand to support existing needs, and catalyze efforts to bring pain neuroscience research in womens health and pregnancy to where it should be.

Author contributions

The original idea for this manuscript was put together by S.H., and the manuscript was written by S.H., M.D., C.R., and A.H. All authors contributed to this manuscript.

Peer review

Peer review information

Communications Medicine thanks Lewis S. Crawford for their contribution to the peer review of this work. A peer review file is available.

Funding

This work was completed with the assistance of funding from the Department of Anesthesia, Critical Care, and Pain Medicine at Boston Children’s Hospital.

Data availability

No primary data went into the production of this manuscript; however, all information and sources will be made available upon request.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s43856-026-01773-6.