Multimodal large language models for women's reproductive mental health

This perspective paper highlights the transformative potential of multimodal large language models in advancing reproductive psychiatry. By integrating diverse data sources, including text, images, audio, and video, MLLMs provide a comprehensive framework for diagnosing, predicting, and managing mental health conditions across various reproductive stages. These models can significantly enhance personalized care by analyzing complex multimodal data relevant to reproductive transitions such as pregnancy, postpartum, menopause, and gynecological conditions. However, challenges like data consistency, bias mitigation, and privacy concerns must be addressed to realize their full impact. Our exploration emphasizes the critical need for initiatives to establish standardized multimodal datasets and robust ethical frameworks specific to women’s reproductive mental health. It is important to recognize that MLLMs are not a one-size-fits-all solution. In some cases, traditional ML methods or unimodal LLMs may be more appropriate, depending on data availability, task complexity, and resource limitations. Integrating MLLMs into reproductive psychiatry can help bridge existing gaps in research and clinical practice, paving the way toward more precise, personalized, and comprehensive mental health care for women.

In recent decades, women’s mental health, particularly in relation to reproductive transitions, has gained considerable attention. There is growing recognition and concern that women face a disproportionately high burden of mental health conditions across the reproductive lifespan (Kornstein and Clayton 2023 ; Kedare et al. 2024 ). For example, approximately 10–20% of women experience postpartum depression (Glasheen et al. 2015 ), 64% of women with polycystic ovary syndrome (PCOS) report depressive symptoms (Chaudhari et al. 2018 ; Xing et al. 2022 ), and 20–30% of menopausal women report significant mood disturbances (Maki et al. 2019 ). The prevalence of premenstrual dysphoric disorder (PMDD) is around 8% (Reilly et al. 2024 ), and anxiety and depression rates are elevated in women undergoing infertility treatments (Cousineau and Domar 2007 ). However, despite the scale of these challenges, mental health conditions in reproductive contexts are often underdiagnosed and understudied, with gaps in research, technology, awareness, and resources that hinder the development of effective interventions and support systems (Redshaw and Wynter 2022 ). Reproductive psychiatry is an important and emerging field within general psychiatry that focuses on the mental health issues linked to reproductive transitions and hormonal changes in women, including: infertility and assisted reproduction, menstruation, pregnancy, childbirth, the postpartum period, perimenopause and menopause, and chronic conditions affecting reproductive health (Howard et al. 2024 ; Epperson et al. 2024 ). Despite its importance, reproductive psychiatry often remains overlooked in broader psychiatric practice, even though approximately half of the population experiences mental health issues related to these reproductive transitions (Hutner, et al. 2022 ). Evaluations in reproductive psychiatry are tailored to specific stages, such as pre-pregnancy assessments for individuals with existing psychiatric diagnoses, postpartum evaluations for mental health following childbirth, and perimenopausal assessments for mood and cognitive changes (Hutner, et al. 2022 ). These evaluations require a multi-dimensional approach that integrates data from diverse sources, including electronic health records (EHRs), genetics (e.g., pharmacogenomic profiles and family history), hormonal profiles, imaging studies, patient-reported outcomes, and wearable devices (Fig.  1 ). Thecomplex interplay of biological, environmental, and lifestyle factors further underscores the need for comprehensive tools to analyze these big data sources, yet current clinical tools often lack this integration, limiting insights and the effectiveness of care in women’s reproductive mental health. Fig. 1 Multifactorial Influences and Multimodal Data in Women's Reproductive Mental Health. This illustration demonstrates the complex interplay of factors affecting women’s reproductive mental health and highlights the multimodal data approaches used in reproductive psychiatry assessments. On the left, various biological, environmental, and lifestyle factors are shown to influence mental well-being. On the right, diverse data modalities—including medical imaging, genetics, wearable devices, and electronic health records—represent illustrative examples of data sources with potential application across all domains. These modalities are not necessarily in current use for each condition but reflect emerging opportunities for MLLM integration. Together, these factors contribute to advancing our understanding of mental health during key reproductive phases and conditions such as pregnancy, postpartum, menopause, infertility, endometriosis, and polycystic ovary syndrome Multifactorial Influences and Multimodal Data in Women's Reproductive Mental Health. This illustration demonstrates the complex interplay of factors affecting women’s reproductive mental health and highlights the multimodal data approaches used in reproductive psychiatry assessments. On the left, various biological, environmental, and lifestyle factors are shown to influence mental well-being. On the right, diverse data modalities—including medical imaging, genetics, wearable devices, and electronic health records—represent illustrative examples of data sources with potential application across all domains. These modalities are not necessarily in current use for each condition but reflect emerging opportunities for MLLM integration. Together, these factors contribute to advancing our understanding of mental health during key reproductive phases and conditions such as pregnancy, postpartum, menopause, infertility, endometriosis, and polycystic ovary syndrome Despite significant advancements in artificial intelligence (AI) and machine learning (ML) models, their use in reproductive psychiatry remains limited in both scope and integration (Thakkar et al. 2024 ). These models often rely on single data types (Olawade et al. 2024 ), typically structured data or text, which restricts their ability to capture the multifaceted nature of reproductive mental health. While AI-based systems have improved predictive capabilities for specific conditions, such as postpartum depression (Nakamura et al. 2024 ), they often lack the ability to comprehensively synthesize diverse data inputs, such as hormonal fluctuations, genetic profiles, patient-reported outcomes, and wearable data, simultaneously. Additionally, conventional models frequently struggle to interpret contextual or temporal factors unique to each stage of reproductive health, resulting in fragmented insights. For example, postpartum depression often goes undiagnosed due to underreporting and lack of early screening tools in primary care settings (Gaynes et al. 2005 ).These gaps underscore the need for more robust, multimodal, and integrative AI systems capable of integrating and analyzing complex datasets to achieve a holistic understanding of mental health across the reproductive lifespan. Traditional machine learning models typically rely on structured and engineered features and are often limited to single data modalities such as numerical clinical measurements or questionnaire data. While these models can, in principle, be adapted to incorporate multimodal inputs, doing so generally requires extensive preprocessing, feature engineering, and parallel pipelines, which may hinder scalability and real-time application. Large language models (LLMs), which are trained on vast textual corpora, advance ML capabilities by enabling a nuanced understanding of unstructured text, yet remain confined to language-based inputs. Multimodal large language models (MLLMs) represent a significant advancement by unifying these capabilities in a single architecture capable of simultaneously processing diverse data types, including text, images, audio, video, and physiological signals (AlSaad et al. 2024a ). This allows for more dynamic, context-aware integration of heterogeneous data streams central to reproductive psychiatry and women’s mental health. The reproductive psychiatry subfield presents unique challenges that make it particularly well suited for MLLM-based solutions. First, reproductive mental health conditions often emerge from complex interactions between physiological, hormonal, psychological, and social factors—each potentially represented in distinct data modalities such as clinical notes, biosignals, audio recordings, and hormonal profiles. Second, the temporal nature of reproductive life stages (e.g., pregnancy, postpartum, menopause) demands models that can process longitudinal data to detect subtle changes over time. MLLMs are uniquely capable of integrating and reasoning across such heterogeneous, temporally-structured inputs, making them ideal for advancing precision mental healthcare in this domain. However, it is important to acknowledge that MLLMs currently demand substantial initial development effort, including advanced data harmonization, preprocessing, and alignment of modalities, due to the complexity of integrating heterogeneous inputs. As pre-trained MLLMs become more widely available, their deployment may eventually streamline downstream predictive and diagnostic tasks, but these efficiencies are not immediate and depend on substantial upfront model and data preparation. By incorporating MLLMs into reproductive psychiatry, we can ultimately better address unresolved clinical questions and fill knowledge gaps left by limited research. This paper explores seven key applications of MLLMs in reproductive psychiatry, highlighting their potential role in diagnosing, predicting, and managing psychiatric conditions associated with premenstrual mood syndromes, pregnancy, postpartum, menopause, infertility, and gynecological conditions such as endometriosis and polycystic ovary syndrome, as well as providing personalized educational resources for patients. Finally, we discuss the challenges and ethical considerations of leveraging MLLMs to improve women’s reproductive mental health. In this section, we examine seven key applications of MLLMs in reproductive psychiatry, highlighting their potential use in diagnosing, predicting, and managing psychiatric conditions related to reproductive health, including transitions between key reproductive phases. We also discuss how MLLMs can deliver tailored educational resources to patients. To ensure a structured and focused analysis, each application is presented in a consistent format that addresses three guiding questions: (1) What are the existing gaps in clinical knowledge and practice? (2) Why are current AI methods insufficient in addressing these gaps? and (3) What unique advantages do MLLMs offer in this context, including specific examples of relevant data types and use cases? Premenstrual mood syndromes, including premenstrual syndrome (PMS) and premenstrual dysphoric disorder (PMDD), are conditions characterized by significant mood disturbances during the luteal phase of the menstrual cycle. PMS affects 48% of menstruating women worldwide (A, D.M.,, et al. 2014 ), while PMDD impacts approximately 8% (Reilly et al. 2024 ). PMDD manifests with intense emotional symptoms, such as severe anxiety, depression, irritability, mood swings, and even suicidal thoughts, making it more serious and debilitating than typical PMS (Gao et al. 2022 ; Akyuz Cim and Cim 2023 ). Despite the significant impact of PMDD, there are no specific diagnostic tests, and the exact cause of hormonal sensitivity remains unclear (Hantsoo and Payne 2023 ). Diagnosis typically relies on mood logs from the past two menstrual cycles to confirm the pattern of symptoms, though many patients struggle to consistently track their symptoms, often leading to misdiagnosis (Salvatore et al. 2024 ; Ko et al. 2023 ). This challenge, along with a lack of women's health expertise among some physicians, results in PMDD being mistaken for typical PMS. Reproductive psychiatry plays a crucial role in managing PMDD through a combination of medication and therapy, significantly improving the quality of life for affected women. Despite the significant impact of PMDD, there has been limited research on using AI algorithms for its diagnosis or treatment. Current AI approaches to PMDD diagnosis are highly dependent on patient-reported symptoms and retrospective mood tracking, which are often incomplete, inconsistently recorded, or subject to recall bias (Dubol et al. 2022 ; Sosnowski, et al. 2022 ). There is a lack of integration between subjective symptom patterns and objective biomarkers such as hormonal fluctuations, sleep patterns, or stress markers. Moreover, psychiatric comorbidities such as major depression or anxiety disorders often go undetected in PMDD populations due to siloed mental and reproductive health data. MLLMs offer a powerful approach to managing PMDD by integrating diverse data sources—including hormone level trajectories, EHRs, patient-reported outcomes, structural MRI (sMRI), wearable signals, speech tone, and genetic information—to uncover complex, multimodal patterns that traditional models may overlook. They can detect temporally aligned symptom onset by synthesizing behavioral, physiological, and self-reported data with high resolution, enabling early identification of mood deterioration. This allows for personalized, timely interventions such as adjusting medication schedules, initiating cognitive behavioral therapy (CBT) modules, or prompting hormonal assessments. Moreover, MLLMs can evaluate the effectiveness of serotonin reuptake inhibitors (SSRIs) in PMDD, identify predictors of treatment response, and analyze psychiatric comorbidities such as depression and anxiety by revealing shared pathophysiological mechanisms. By tracking longitudinal data across reproductive and mental health domains, MLLMs can also assess long-term risks such as suicidality or chronic mood instability, offering a more complete and proactive model for PMDD care. During pregnancy, hormonal fluctuations are associated with various mental health conditions in the mother, including depression, anxiety, eating disorders, and post-traumatic stress disorder (PTSD), especially following a pregnancy loss. Among those with pre-existing mental health conditions relapse is high. For example, evidence strongly supports that women with a history of eating disorders are at an increased risk of relapse during pregnancy, with 67% experiencing a recurrence (Makino et al. 2020 ). Maternal stress during pregnancy also has been shown to influence fetal brain development and increase infant’s vulnerability to mental health issues like anxiety or depression later in life (Wu et al. 2024 ; Bergh et al. 2020 ). In addition, mental health issues during different gestational periods may differentially affect specific brain regions and circuits, potentially resulting in distinct neurodevelopmental and psychiatric outcomes (Collins et al. 2024 ). While hormonal fluctuations are a key contributor to perinatal mental health, they represent just one aspect of a much more complex picture. Social determinants such as financial stress, intimate partner violence, limited social support, poor access to prenatal care, food insecurity, and co-occurring physical illnesses are also strongly associated with elevated risk for anxiety and depression during pregnancy (Barat et al. 2024 ; Ariasih et al. 2024 ). Reproductive psychiatrists often manage stress and anxiety during pregnancy by combining pharmacologic treatments, such as selective SSRIs, with non-pharmacologic approaches like CBT and social support, depending on the severity of the presentation (Bertozzi-Villa et al. 2025 ). They may recommend continuation of psychiatric medication after conducting a thorough risk analysis, considering both the mother's mental health and the consequences of not treating the underlying mental health condition, and the potential effects of maternal–fetal medication exposure (Gumusoglu and Santillan 2022 ). Untreated depression or anxiety during pregnancy can contribute to adverse maternal and fetal outcomes, such as cesarean delivery, pre-eclampsia, preterm birth, low birth weight and small for gestational age. However, uninformed primary care providers, obstetricians, or general psychiatrists might inappropriately discontinue psychiatric medication during pregnancy due to concerns about fetal exposure alone (Gumusoglu and Santillan 2022 ). Unfortunately, medication discontinuation is strongly associated with a risk of mental health relapse during pregnancy (Bayrampour, et al. 2020 ). Although several AI models have been developed to support mental health care during pregnancy (Garbazza et al. 2024 ; Krishnamurti et al. 2024 ), many of these systems rely on a single data modality—such as text from EHR clinical notes or standardized patient questionnaires—and are often trained on relatively small or homogenous cohorts. This narrow focus limits their ability to generalize across diverse populations and fails to reflect the complexity of real-world perinatal care. In contrast, MLLMs offer a more holistic approach by integrating heterogeneous data sources that are particularly relevant to mental health during pregnancy. These can include clinician notes capturing mood fluctuations or sleep disturbances, physiological signals from wearable devices that track heart rate variability or sleep quality, patient-reported outcomes on mood and stress, and psychosocial screening data on support systems, financial strain, or trauma history. This integrative approach enables MLLMs to identify subtle and context-specific patterns that may be missed by traditional ML models, supporting more personalized and precise clinical insights. For example, a pregnant patient exhibiting depressive symptoms following a change in psychiatric medication could be identified by an MLLM that integrates pharmacogenomic data, longitudinal treatment history, and clinician documentation. Based on this multimodal input, the model may recommend medication adjustments, flag potential adverse interactions, or generate alerts for timely psychiatric follow-up. Alternatively, if the MLLM detects that psychological distress is more closely associated with contextual factors such as food insecurity or housing instability, extracted from both structured social determinants data and free-text clinical notes, it could prioritize an alternative care pathway. This may include referral to social services, culturally tailored cognitive behavioral therapy, or connection to community-based mental health resources. By synthesizing these diverse data sources within a biopsychosocial framework, MLLMs can support personalized, context-aware interventions for perinatal mental health that go beyond one-size-fits-all diagnostic or treatment recommendations. MLLMs can analyze larger datasets from pregnancy cohorts, enabling more robust pattern recognition. This enhanced capacity allows MLLMs to improve predictive analytics by combining EHRs and physiological data, such as heart rate variability and sleep patterns from wearable devices, to identify women at high risk for mental health conditions like depression and anxiety (AlSaad et al. 2024a ),. Additionally, MLLMs support real-time monitoring and intervention through the analysis of speech, neuroimaging, and fetal conditions, allowing for continuous mental health assessments and tailored therapeutic recommendations. Furthermore, MLLMs aid in personalized medication management by assessing individual patient data, recommending safe alternatives to minimize the risk of psychiatric relapses while ensuring maternal and fetal safety. Overall, MLLMs provide a comprehensive and dynamic approach to addressing the complex mental health needs of pregnant women. Spontaneous and elective pregnancy loss—including miscarriage, recurrent pregnancy loss (RPL), and abortion—are emotionally complex experiences that can result in profound psychological consequences, including grief, anxiety, guilt, and depression. Studies show that nearly 20% of known pregnancies end in miscarriage, with up to 2–3% of women experiencing RPL (Strumpf et al. 2021 ). Abortion can also be accompanied by social stigma and emotional distress depending on the individual’s context and support system (Ludermir et al. 2011 ; Makleff et al. 2019 ). Importantly, mental health outcomes following these events are influenced by various factors including hormonal shifts, prior psychiatric history, social support, and cultural environment. Despite the frequency and psychological impact of these experiences, they are underrepresented in both clinical mental health evaluations and AI research. Existing AI models rely heavily on structured data (e.g., diagnostic codes, medication history) or static self-report scales, which fail to capture the emotional complexity and contextual variability of pregnancy loss experiences. These models are rarely trained on datasets that reflect the unique physiological and psychological responses following abortion or miscarriage, including grief trajectories, cultural nuances, or trauma-specific reactions. Additionally, the lack of integrated data across obstetric, psychiatric, and behavioral domains limits their utility for personalized risk prediction. Few algorithms incorporate hormonal data or longitudinal distress markers, further reducing sensitivity to delayed or evolving mental health outcomes (Kwok et al. 2024 ). As a result, current AI systems are poorly equipped to identify or respond to the nuanced mental health needs of individuals affected by pregnancy loss. MLLMs offer a novel and timely approach to supporting women in these scenarios by integrating heterogeneous data streams—such as hormonal profiles (e.g., cortisol, estradiol levels indicating stress or mood vulnerability), EHR notes documenting prior mood disorders or sleep disturbances, structured clinical data on obstetric history, and wearable sensor inputs (e.g., sleep disruption, reduced activity levels, heart rate variability as a proxy for autonomic dysregulation)—to identify individuals at elevated risk of developing acute or prolonged psychiatric symptoms following pregnancy loss. Additionally, MLLMs can analyze free-text patient narratives and digital journaling entries, extracting linguistic markers of distress, which may be early indicators of depressive or anxiety disorders. These capabilities enable the development of emotionally intelligent digital health tools that can monitor symptom progression longitudinally, deliver personalized psychoeducation, and initiate empathetic, context-aware conversations in real time. For example, an MLLM-powered chatbot could offer evidence-based cognitive restructuring techniques to someone expressing guilt after a miscarriage or guide users through mindfulness exercises tailored to grief after abortion. MLLMs have the potential to facilitate confidential screening for post-abortion trauma, miscarriage-related distress, or complicated grief in individuals who are reluctant to seek in-person care due to stigma, emotional sensitivity, or privacy concerns. From the clinical side, these models could assist practitioners by flagging high-risk patients based on subtle multimodal cues—such as combined low HRV, insomnia patterns, and obstetric and psychiatry history—enabling earlier referrals to reproductive mental health specialists. MLLMs have the potential to bridge longstanding gaps in diagnosis, monitoring, and treatment within this highly stigmatized and underserved domain of reproductive mental health. The postpartum period involves significant hormonal changes, with approximately 85% of women experiencing some form of mood disturbance during this time (Massachusetts General Hospital Center for Women's Mental Health 2023 ). While most cases are mild and temporary, 10 to 20% of women encounter more severe symptoms of postpartum depression (PPD) (Glasheen et al. 2015 ). Although less recognized, postpartum anxiety disorders are similarly estimated to impact 10–20% of women during the postpartum period (Esquivel Lauzurique et al. 2022 ). Additionally, postpartum PTSD occurs in 1–6% of new mothers (Grekin and O'Hara 2014 ), while postpartum obsessive–compulsive disorder (OCD) is reported in approximately 2–3% of women after childbirth (Miller et al. 2013 ). Postpartum mental disorders are frequently under-diagnosed and inadequately treated, which can lead to lasting impacts on the mother, child, and entire family unit (Dimcea et al. 2024 ). Moreover, emerging evidence suggests that genetic and epigenetic factors contribute to the etiology and severity of perinatal mental health disorders, including postpartum depression and anxiety (Guintivano et al. 2018 ). High-throughput sequencing technologies, such as whole-genome sequencing, transcriptomics, and DNA methylation profiling, enable fine-grained characterization of individual biological risk profiles. While currently underutilized in mental health prediction tasks, the integration of such biological data into MLLMs, in conjunction with clinical, behavioral, and contextual information, holds substantial promise. Current machine learning models for postpartum psychiatric disorders primarily rely on EHR data and patient self-reports to predict conditions like postpartum depression and anxiety (Nakamura et al. 2024 ; Liu et al. 2023a ; AlSaad et al. 2025b ). However, their effectiveness is limited by narrow observational windows—often restricted to the pregnancy or immediate postpartum period—and by fragmented datasets that exclude critical maternal-infant linkages. In many healthcare systems, the inability to connect mother and infant records further hinders comprehensive analysis. As a result, postpartum mental health is frequently evaluated in isolation, overlooking the extended and evolving nature of these conditions and their interaction with infant outcomes. MLLMs offer a potential paradigm shift by enabling continuous, multimodal, and longitudinal postpartum mental health assessment. Unlike traditional models, MLLMs could ingest and contextualize heterogeneous data types, including facial expressions and vocal tone from video consultations, physiological stress signals from wearable devices (e.g., sleep disturbance, heart rate variability), and even digital biomarkers from smartphone usage. They can also link maternal mental health with infant-related variables, such as feeding schedules, sleep–wake patterns, and neonatal complications, allowing for a more holistic understanding of risk. Moreover, MLLMs could support real-time, adaptive mental health support systems. For example, they can power virtual companions that deliver tailored check-ins, peer support access, or escalate to clinical care based on dynamic distress signals across modalities. This model not only enhances detection but also extends supportive care to those who might otherwise be unreachable due to stigma, access issues, or time constraints (Ferrara 2024 ; Sadeghi et al. 2024 ). It is important to note that maternal mental health challenges may also extend into the broader parenting period; however, the present perspective focuses specifically on the postpartum phase as part of reproductive mental health. The menopausal transition, including perimenopause and post-menopause, poses several mental health challenges such as mood swings, depression, anxiety, irritability, sleep disturbances, forgetfulness, and difficulties with concentration and decision-making (O'Reilly et al. 2024 ; Wen et al. 2024 ). Approximately 20–30% of women experience depressive symptoms during menopause (Maki et al. 2019 ), and 25% of women aged 60 and over in the United States reported using antidepressants (Brody and Gu 2020 ). Forgetfulness affects 31–44% of women, contributing to increased levels of distress and anxiety (Gold et al. 2004 ). Around 30% of menopausal women experience sleep disturbances, primarily due to vasomotor symptoms (hot flashes and night sweats), which are strongly linked to anxiety and depression (Shieu et al. 2023 ). Menopausal hormonal fluctuations, particularly reduced estrogen levels, can disrupt neurotransmitter systems like serotonin and dopamine, triggering a relapse or alteration in mental health conditions (Naguy et al. 2021 ). Although female mental health during menopause is poorly understood, reproductive psychiatry is crucial in managing the mental health challenges women face during menopause as well as cognitive challenges, identity and self-esteem issues, and sexual health changes, and giving a voice to women’s concerns during this key period of their lives which has historically been overlooked (Brown et al. 2024 ). AI, including LLMs, has been underutilized in studying menopausal mental health issues that impact long-term health and quality of life (Roa Diaz et al. 2019 ). MLLMs may become valuable tools for reproductive psychiatrists to create personalized treatment plans that guide women through the menopausal transition with greater ease and confidence. First, menopause exhibits distinct molecular signatures that regulate sleep and activity (Zambotti et al. 2015 ; Liu et al. 2023b ). MLLMs can analyze complex genetic and transcriptomic datasets to identify molecular pathways involved in menopause-related sleep disturbances. Moreover, MLLMs may eventually incorporate longitudinal audio data from in-home smart speakers or wearable biosensors to analyze sleep patterns and nighttime arousals. Specialized wearable devices, such as wristbands and skin sensors that detect changes in skin conductance or temperature, can accurately track hot flashes and nocturnal sweating episodes, providing high-resolution physiological data related to vasomotor instability. When further combined with hormonal profiles and mood tracking, such models can detect early deviations in sleep quality that precede clinical symptoms of mood disorders (D'Ambrosio et al. 2005 ; Auranen 2022 ; Kravitz et al. 2003 ). Unlike unimodal models, MLLMs can contextually weight these multimodal signals, enabling more precise, real-time monitoring and interventions tailored to menopausal mental health. Second, diet plays a crucial role in managing mental health symptoms during menopause (Vega-Rivera et al. 2024 ; Bodnaruc et al. 2023 ; Grigolon et al. 2023 ). While traditional models might only analyze dietary intake or biomarkers in isolation, MLLMs can integrate diverse menopause-specific inputs, such as age-related metabolic changes, dietary preferences, micronutrient deficiencies, comorbid conditions (e.g., osteoporosis or cardiovascular risk), and symptom trajectories, to generate adaptive, individualized nutritional plans. This integrative approach helps optimize dietary strategies for alleviating mood disturbances, sleep disruption, and cognitive complaints associated with menopause. Third, MLLM-powered chatbots offer significant advantages in managing menopausal mental health by analyzing user sentiment, voice tone, physiological signals, and verbal disclosures about menopause-related symptoms. Unlike unimodal tools, these chatbots can detect subtle affective shifts (such as irritability, fatigue, or emotional blunting) linked to perimenopausal mood changes, and deliver contextually relevant, empathic responses. This continuous and personalized support can serve as a scalable adjunct to care, significantly enhancing mental well-being throughout the menopausal transition (Denche-Zamorano et al. 2024 ). Infertility can be a highly distressing experience, significantly affecting a woman’s emotional and psychological well-being during treatments such as in vitro fertilization (IVF), intrauterine insemination (IUI), and fertility medication regimens (Lacey et al. 2021 ; Vikstrom et al. 2015 ; Stewart et al. 2015 ). These treatments often involve intense hormonal therapies, like gonadotropins and letrozole, which may have profound effects on mental health, contributing to stress, anxiety, and mood disturbances that are frequently underrecognized and inadequately addressed (Cousineau and Domar 2007 ; Rahimi et al. 2021 ). The onset of depression related to infertility varies, but one study (Luk and Loke 2015 ) suggests that it typically develops after 2–3 years of dealing with the condition. The mental health of women undergoing IVF plays a crucial role in treatment outcomes, as higher stress and anxiety levels are associated with lower success rates, increased risk of treatment discontinuation, and adverse maternal outcomes (Miller et al. 2019 ; Dayan et al. 2022 ). Additionally, eating disorders can disrupt ovulation by affecting hormonal regulation and menstrual cycles. Severe weight loss from anorexia nervosa may lead to amenorrhea, while bulimia nervosa often causes menstrual irregularities; both of these conditions may impair the hypothalamic-pituitary-ovarian (HPO) axis and further complicate infertility (Bruneau et al. 2017 ). Despite growing interest in AI applications in reproductive medicine, most existing models focus narrowly on predicting clinical outcomes, such as embryo quality or fertilization success, without incorporating the patient’s emotional or psychological state (Orovou et al. 2025 ; AlSaad et al. 2024b ; AlSaad et al. 2025a ). Current systems rarely integrate mental health data, and when they do, it's typically limited to static variables such as baseline anxiety scores or medication use. Emotional distress associated with infertility is dynamic and evolves across treatment cycles, shaped by factors such as hormonal fluctuations, cycle stage, interpersonal stressors, and prior psychological trauma—variables that are frequently underrepresented or entirely absent in conventional fertility-related algorithmic models. Moreover, there is little interoperability between reproductive endocrinology datasets and psychiatric records, making it difficult to build a holistic picture of patient well-being (Harder et al. 2024 ; Gao et al. 2020 ; Estes et al. 2021 ). This results in limited capacity to detect at-risk patients early or to tailor psychosocial interventions in real time. MLLMs offer a transformative approach to addressing the mental health dimensions of infertility by integrating diverse, temporally aligned data sources that go beyond the scope of traditional AI models. These include patient-reported outcomes, structured psychological assessments, hormone levels, and multimodal signals such as voice tone, facial affect, and sleep patterns collected via wearable devices. Unlike conventional models that rely on static or isolated inputs, MLLMs can continuously monitor and contextualize emotional and physiological responses across different phases of the treatment cycle. For instance, a surge in anxiety or sleep disruption during ovarian stimulation could trigger early delivery of personalized interventions, such as CBT modules, mindfulness-based stress reduction (MBSR), or digital psychoeducation tailored to the patient’s evolving needs. MLLMs also enhance patient–provider communication by analyzing language and sentiment in real time, allowing clinicians to adjust tone, frequency, or content of feedback during emotionally taxing treatment phases like embryo transfer or post-treatment waiting periods. This enables more empathetic, responsive care that fosters trust and reduces dropout rates. By offering proactive, personalized, and multimodal mental health support, MLLMs empower fertility specialists and reproductive psychiatrists to co-manage the psychological burden of infertility—ultimately improving adherence, satisfaction, and overall treatment outcomes (Borghi et al. 2021 ). Polycystic Ovary Syndrome (PCOS) is the most common endocrinopathy, affecting 5% to 18% of women of reproductive age (Bozdag et al. 2016 ). It is a chronic condition that persists through reproductive years and can continue to have health impacts beyond menopause (Millan-de-Meer et al. 2023 ). The condition is stigmatizing and affects a woman’s identity and quality of life. PCOS can lead to significant mental health challenges, with 64% and 39% of women with PCOS experience depression and anxiety, respectively (Chaudhari et al. 2018 ; Xing et al. 2022 ). Additionally, women with PCOS show a higher prevalence of less common psychiatric disorders such as bipolar disorder, schizophrenia, eating disorders, and personality disorders, further complicating their mental health (Doretto et al. 2020 ). Despite this high psychiatric burden, the mental health disorders associated with PCOS remain under-recognized and are often overlooked in both clinical practice and research. Most existing AI models for PCOS are focused on improving clinical diagnosis using physical or biochemical markers, while largely overlooking psychiatric symptoms (Verma et al. 2024 ; Barrera et al. 2023 ). These models rarely incorporate mood instability, body image distress, or sleep disruption—key psychological components of PCOS-related mental health burden. Furthermore, they tend to operate on siloed datasets that separate reproductive endocrinology from psychiatric data, failing to capture the interplay between hormonal imbalance and emotional dysregulation (Ahmed et al. 2023 ). Static prediction models also struggle to reflect the chronic and fluctuating nature of PCOS, limiting their ability to inform mental health risk over time or personalize psychological care strategies. The potential of MLLMs in treating and managing mental health conditions linked to PCOS is significant. Unlike conventional models, MLLMs can integrate multimodal data—including hormonal profiles, wearable biosignals, neurocognitive assessments, and patient-reported mood and stress levels—to capture a holistic view of mental health in PCOS. MLLMs can analyze these diverse inputs to identify specific biomarkers and psychosocial risk factors for psychiatric comorbidities, as well as predict the likelihood of developing these conditions, enabling early and personalized interventions. Moreover, given that PCOS is a chronic health condition, MLLMs are well-suited for longitudinal modeling that tracks symptom fluctuations over time, monitors treatment response, and adapts intervention strategies accordingly. This approach supports more precise, scalable, and proactive mental health care for women living with PCOS. Endometriosis is a chronic condition where endometrial tissue grows outside the uterus, causing pain due to inflammatory mediators released during menstruation (Hutner, et al. 2022 ). Endometriosis affects approximately 7% to 15% of women of reproductive age, with an estimated 200 million women affected globally (Swift et al. 2024 ; Harder et al. 2024 ). Women with endometriosis have an increased risk of various psychiatric disorders, including anxiety (Gao et al. 2020 ), depression (Estes et al. 2021 ), stress (Gao et al. 2020 ), bipolar disorder (Chen et al. 2020 ), and eating disorders (Gao et al. 2020 ). However, it remains unclear whether this increased risk is due to chronic pelvic pain associated with endometriosis, the etiological factors of endometriosis, or other factors such as chronic inflammation (Barneveld et al. 2022 ). Neuroimaging studies (Maulitz et al. 2022 ) suggest that brain regions involved in pain processing, emotion, cognition, self-regulation, and reward are affected in women with endometriosis. Despite advances in AI applications for physical diagnosis of endometriosis, existing models are largely limited to identifying disease presence via imaging, symptom reports, or surgical confirmation (Dungate et al. 2024 ; Sivajohan et al. 2022 ). These approaches rarely incorporate dynamic psychological indicators such as mood volatility, trauma exposure, or emotional exhaustion, which are common in women living with chronic pelvic pain. Furthermore, AI tools currently lack integration between gynecological, psychiatric, and neurological data streams, hindering their ability to identify the complex bidirectional relationship between chronic pain and mental health. As a result, these models remain insufficiently equipped to predict or manage psychiatric outcomes or to guide interdisciplinary interventions. MLLMs offer a powerful framework for identifying and managing psychiatric comorbidities in endometriosis by synthesizing multimodal data—including pain scores, hormonal and inflammatory biomarkers, wearable sensor outputs, neuroimaging, and patient-reported outcomes. These models can uncover nuanced correlations between chronic pelvic pain and psychological symptoms such as depression and anxiety, while also tracking symptom trajectories in relation to cycle stage, treatment phase, and external stressors. By incorporating neurobiological and behavioral signals alongside contextual data, MLLMs could enable early risk detection, personalized intervention planning, and ongoing mental health monitoring—providing a scalable, data-driven approach to interdisciplinary care for women with endometriosis. Stigma surrounding mental health conditions often prevent patients from seeking the help they need. This challenge is particularly acute in reproductive psychiatry, where emotional distress may be dismissed as a “normal” part of hormonal changes, leading to under recognition and undertreatment. Educational resources for reproductive mental health are often generic, static, and poorly tailored to individual needs—failing to account for personal, cultural, or situational differences in how women experience and understand their symptoms. There is a pressing need for scalable, personalized, and stigma-sensitive educational strategies that empower women to better understand and manage their mental health. Current AI and digital education tools often fall short in reproductive psychiatry because they lack the ability to personalize content to reflect the user’s reproductive stage, comorbidities, cultural context, or emotional readiness. Existing platforms tend to deliver one-size-fits-all information, often focused on general mental health or generic symptom management. Furthermore, these systems rarely integrate real-time feedback, natural language queries, or cross-cultural adaptation, which limits their relevance and usability. As a result, patients may disengage or mistrust educational content that feels impersonal, overly clinical, or misaligned with their lived experience. MLLMs could play a crucial role in overcoming these limitations by providing accessible, adaptive, and personalized educational content tailored to each patient’s reproductive and mental health journey. By leveraging individual-level data and user preferences, healthcare providers can use MLLMs to generate interactive resources that align with different learning styles, integrate culturally sensitive language, and present content in multimodal formats—such as text, audio, and video (Aydin et al. 2024 ; Aghamaliyev et al. 2025 ). This personalized approach ensures that women receive relevant and comprehensible information, empowering them to make informed decisions about their care and improving treatment adherence. MLLMs can also incorporate real-time feedback and emerging clinical evidence to keep educational content current and aligned with best practices. Additionally, these models can facilitate virtual counseling and peer support forums, creating safe, stigma-free environments for women to explore mental health topics. At the societal level, MLLMs can also generate targeted educational content for partners, family members, friends, and employers to improve support systems around women facing reproductive mental health challenges. By tailoring this content to the unique emotional, cultural, and clinical context of each patient, MLLMs enable more compassionate, informed engagement from those in a woman’s social network—broadening the reach and impact of reproductive mental health education. While the general challenges of employing AI, LLMs, and MLLMs in healthcare have been extensively discussed in the literature (AlSaad et al. 2024a ; Zhou et al. 2024 ; Solaiman et al. 2023 ; Ong et al. 2024 ; Starke et al. 2022 ), this section focuses on the specific challenges associated with training and applying MLLMs in women’s reproductive psychiatry. Figure  2 provides an overview of the five key challenges discussed in this section. Fig. 2 Key challenges in applying Multimodal Large Language Models (MLLMs) to women’s reproductive psychiatry Key challenges in applying Multimodal Large Language Models (MLLMs) to women’s reproductive psychiatry Longitudinal data is key for studying women’s reproductive mental health, as conditions, like hormonal fluctuations, fertility issues, and menopause, require long-term monitoring to understand their progression and impact on overall well-being. Additionally, there is a shared risk between the different hormonally-mediated reproductive health conditions, such that an individual with one reproductive mental health condition such as PMDD is also more likely to experience postpartum depression or perimenopausal depression as well. However, maintaining the quality and consistency of such data over time is challenging. Changes in healthcare providers, varying data collection methods, lack of a universal EHR, and inconsistent follow-up can result in fragmented and incomplete datasets. These challenges are further exacerbated by system-level barriers such as lack of reimbursement for longitudinal mental health monitoring, limited integration between obstetric and psychiatric records, and transitions between reproductive life stages that fall outside of continuous care pathways. Data ownership and interoperability issues, particularly between commercial apps, hospital systems, and personal health devices, also hinder the consolidation of longitudinal streams into formats usable by MLLMs. Although commercial technologies such as next-generation smartwatches, AI-enabled home assistants with visual and audio sensing, and mobile platforms for mental health tracking offer promising opportunities for real-world, longitudinal, and multimodal data collection, they also present significant ethical and legal challenges—particularly when commercial entities serve as intermediaries or primary collectors of user data that may later be used in MLLM training or inference. These companies may retain ownership over the data, obscure its downstream uses, or engage in commercial resale, thereby undermining users’ autonomy and trust. Future work in MLLM development for reproductive psychiatry should prioritize establishing high-quality, standardized longitudinal data to enhance the reliability and effectiveness of these models in managing women’s reproductive mental health. Technological innovations offer promising avenues for improving longitudinal data capture. For example, wearable biosensors can continuously track physiological signals such as sleep patterns, heart rate variability, and temperature changes, which are relevant to hormonal fluctuations and mood regulation. Mobile health applications can prompt real-time symptom logging, reducing recall bias and improving consistency. Emerging digital biomarkers and point-of-care hormone monitors (e.g., salivary cortisol, LH, or estradiol sensors) provide opportunities for real-time hormonal profiling across reproductive transitions. When integrated into structured platforms, these tools can fill gaps in clinical data and enable MLLMs to learn from continuous, multimodal streams over time. To support these technologies at scale, we recommend integrating them into coordinated care models—such as digital platforms that span obstetric, psychiatric, and primary care—and promoting interoperability standards that facilitate secure data sharing. Grant-funded registries, institutional data lakes, and federated learning networks may also enable aggregation of longitudinal, multimodal data across populations and institutions while preserving patient privacy. When integrated into structured platforms, these tools can fill gaps in clinical research data and enable MLLMs to learn from continuous, multimodal streams over time. Bias in training data related to women’s health can significantly hinder the development of reliable and fair MLLM models. One major issue is the scarcity of high-quality, representative data on women’s reproductive and mental health. Historically, research has often underrepresented diverse populations, such as women of color, while disproportionately focusing on negative aspects of reproductive health, such as premenstrual syndrome (Van and Meleis 2003 ; Boyden et al. 2014 ). In the United States, it was not until 1993 that legislation was enacted to mandate the inclusion of women in research studies. Prior to this, research often focused on white men (Kornstein and Clayton 2023 ). Furthermore, ethical limitations make randomized controlled trials often unfeasible in pregnant populations. Socio-cultural factors, such as the stigmatization surrounding infertility or mental health conditions, often lead to underreporting, resulting in biased datasets (Ballone and Richards 2023 ). Disparities in healthcare access further exacerbate this issue, with women from marginalized communities are frequently underrepresented due to limited access to healthcare services. Moreover, biases in the data collection tools—such as surveys or medical devices— are often designed with a bias toward male physiology, further distorting assessments of women’s health. Human bias in psychiatric diagnosis and reliance on self-reported outcomes is a well-known challenge in mental health (Prakash et al. 2023 ). As a result, the lack of fair and representative health data may hamper the ability to leverage advances in MLLMs to enable holistic reproductive and psychiatric care in women. In MLLMs, these biases can be further amplified due to the multimodal nature of inputs. For example, if wearable biosensor data is primarily collected from higher-income, tech-literate populations, the model may underperform or misclassify cases from underserved groups. Errors in one modality (e.g., mislabeling in clinical notes) can compound across others (e.g., facial affect or voice tone), leading to biased predictions or interventions. To mitigate these issues within the context of MLLMs for women’s reproductive psychiatry, targeted strategies must be adopted at both the data and model development levels. First, federated learning frameworks can enable training on sensitive reproductive and psychiatric data across diverse clinical sites—such as fertility clinics, OB-GYN departments, and psychiatric practices—without requiring raw data to be centrally stored, thus preserving privacy while improving population diversity. Second, MLLM evaluation pipelines should incorporate bias-aware diagnostics that disaggregate performance by reproductive stage (e.g., pregnancy, menopause), race/ethnicity, and socioeconomic status to detect subgroup disparities in predictions or outputs. Third, when data diversity is limited—such as underrepresentation of postpartum mental health in minority populations—data augmentation or oversampling techniques can help mitigate skewed learning. Finally, involving women and community advocates in the design of symptom-tracking tools, wearable sensors, and mental health assessments can ensure that multimodal data inputs reflect culturally and contextually relevant indicators of distress, rather than relying on narrowly defined constructs. Standardizing health definitions for women’s reproductive mental health is particularly challenging due to the complex, dynamic, and individualized nature of conditions such as postpartum depression, premenstrual dysphoric disorder, and menopause-related mood disturbances. What constitutes a “normal” menstrual cycle, mood fluctuation, or menopausal transition can vary widely across medical disciplines, cultures, and patient experiences. For example, emotional changes in the postpartum period may be interpreted as typical “baby blues” in some settings but meet criteria for clinical depression in others. These definitional inconsistencies extend to terminology used in health records—for instance, “mood instability” in one clinical setting may be documented as “emotional lability” in another—creating challenges for data harmonization. Such variability poses significant obstacles for MLLMs, which rely on consistent semantic labels and time-aligned clinical data to learn patterns across multimodal sources. Without standardized definitions of symptom onset, severity, and reproductive stage, it becomes difficult to integrate data from structured fields, clinical notes, wearable sensors, and patient-reported outcomes. These discrepancies introduce noise during training and reduce both the accuracy and generalizability of MLLMs across healthcare systems and populations. To address these limitations, we propose establishing consensus-driven, stage-specific definitions for reproductive mental health conditions through collaborative efforts among psychiatrists, OB-GYNs, endocrinologists, and patient advocates. For instance, defining what constitutes clinically significant postpartum anxiety should include structured symptom thresholds, standardized timeframes, and culturally adaptive descriptors. Incorporating international coding standards (e.g., SNOMED-CT, ICD) alongside validated instruments (e.g., EPDS, PMDD Diagnostic Tools) into EHR metadata and training data pipelines can enhance semantic consistency across sites. In MLLMs, these standardized labels would improve supervised learning and facilitate better transferability across healthcare systems, languages, and populations. There is a notable lack of multimodal datasets specific to women's reproductive mental health, largely due to the challenges of data acquisition, integration, and the decentralization of data collection across specialties. Existing datasets are often siloed—psychiatric EHRs, OB-GYN notes, hormone assays, and wearable data are rarely linked—limiting the ability of MLLMs to capture the complex interplay between biological, psychological, and social dimensions of reproductive mental health. This lack of integration hampers the development of personalized MLLM-based interventions. For instance, in managing PCOS, an MLLM may struggle to link metabolic data (e.g., insulin resistance), psychiatric symptoms (e.g., anxiety or disordered eating), and hormonal profiles if these data streams are stored separately or inconsistently labeled. Without centralized, time-aligned data across modalities, MLLMs cannot generate accurate or generalizable predictions, limiting their ability to support individualized care across diverse patient populations. To address this, we recommend the development of longitudinal, multimodal registries for reproductive mental health, modeled after successful disease-specific consortia. These should include structured data elements across hormonal, psychiatric, and patient-reported domains. Integration pipelines that synchronize wearable and EHR data by reproductive stage are critical for training usable models. Cross-disciplinary collaboration—among psychiatrists, obstetricians, endocrinologists, and data scientists—should guide both dataset design and ethical governance. Community engagement is essential to ensure cultural relevance and equitable representation. Additionally, initiatives should incorporate secure data-sharing protocols and transparent consent models to support scalable, privacy-conscious deployment of MLLMs. Without such efforts, the full potential of MLLMs to deliver individualized, real-time support for reproductive mental health will remain unrealized. Women’s reproductive mental health data is inherently sensitive, touching on deeply personal experiences and long-standing social stigmas (Appel 2024 ). In the context of MLLMs, privacy concerns are further intensified by the integration of multiple data modalities, which may collectively expose intricate details of a woman’s reproductive and psychiatric history. These concerns are not hypothetical. In the wake of the Dobbs decision in the U.S., women of reproductive age face very real legal and social repercussions for sharing health data related to pregnancy, abortion, or contraceptive use (Kapadia et al. 2024 ; Tanne 2025 ). There is growing distrust in digital health systems and significant pushback against using fertility-tracking apps, wearables, or telehealth services due to fears of surveillance, criminalization, and data misuse (Schreiber et al. 2023 ; Madden et al. 2024 ; Mosley et al. 2025 ). MLLM developers must acknowledge that reluctance to share data is a rational response to perceived and actual threats, and any responsible deployment strategy must be designed accordingly. Technical safeguards alone, such as data de-identification, encryption, and access control, are no longer sufficient. Privacy-preserving machine learning methods, including federated learning and differential privacy, should be integrated into MLLM pipelines by default Transparency in the consent process must extend beyond the scope of data usage to explicitly address risks of data breaches and misuse. Participants should be clearly informed—using accessible, non-technical language—about the potential for unauthorized access, model misuse, and secondary data inferences. Consent protocols should incorporate dynamic and tiered consent frameworks, where users can specify levels of data sharing, revoke access at any time, and receive updates when risk levels change (e.g., following a breach). Moreover, transparency requires ongoing communication with users, including breach notification policies and audit logs detailing how their data has been accessed and used. Future research must advocate for transparent multimodal longitudinal data governance frameworks, legal safeguards against unauthorized data monetization, and clear separation between health research use and commercial exploitation. Platforms should also offer real-time consent dashboards and audit trails that enable users to revoke data access at any time. In compliance with ethical standards for digital mental health research, revoking data access should also entail the secure deletion of previously collected data from the user upon request. This includes implementing robust data governance protocols that ensure timely and verifiable destruction of data from all active storage systems, backups, and caches, where feasible, to respect participants' digital autonomy and right to be excluded. Additionally, privacy risks related to ambient data sources, such as longitudinal audio recordings, must be critically addressed. These data sources could inadvertently capture private conversations or sensitive interactions involving non-consenting individuals, raising ethical concerns around surveillance, autonomy, and contextual integrity. Any integration of such data into MLLMs must be governed by strict opt-in frameworks, real-time consent prompts, household-level consent policies, and the development of on-device processing pipelines that minimize data sharing and storage. Where legal risk cannot be mitigated through design, we believe it is ethically necessary to explicitly communicate that certain applications, such as ovulation tracking, pregnancy monitoring, or abortion decision support, may not be feasible or safe in specific contexts. In such cases, the limitations of MLLM deployment must be acknowledged transparently as part of informed consent processes and research disclosures. Finally, any future development of MLLMs for reproductive psychiatry must involve not only IRBs and data scientists, but also legal experts, community advocates, and end-users. Building trust in these models requires participatory governance, explainability of outputs, and shared decision-making on how, when, and whether personal data is used.