Machine Learning Links Maternal Overprotection to Premenstrual Disorders in Exploratory Study

In a newly published exploratory investigation, researchers applying interpretable machine-learning techniques found that perceived parental bonding characteristics—most notably maternal overprotection—ranked among the strongest predictive factors associated with the presence of premenstrual disorder. Published in BioPsychoSocial Medicine by a team from Kyoto University, the cross-sectional study evaluated whether psychosocial and physiological variables not conventionally viewed as classic risk factors could help classify premenstrual disorder. Across an automated predictive model, maternal relationship characteristics accounted for six of the top twenty most influential predictive features, with four specifically reflecting maternal overprotection.

Premenstrual disorder encompasses both premenstrual syndrome and premenstrual dysphoric disorder, conditions characterized by recurring somatic and psychological distress during the luteal phase of the menstrual cycle. Premenstrual syndrome occurs three to ten days before menstruation and affects approximately half of women of reproductive age, presenting with physical and emotional symptoms such as breast tenderness, bloating, headaches, mood swings, depression, anxiety, anger, and irritability. Premenstrual dysphoric disorder represents a severe manifestation dominated by psychiatric symptoms, first introduced as a research criterion in the Diagnostic and Statistical Manual of Mental Disorders Fourth Edition and subsequently classified as a depressive disorder in the DSM-5. Women affected by these disorders encounter substantial disruption in social functioning, daily activities, and interpersonal relationships.

Although premenstrual disorder has documented associations with complex neuroendocrine mechanisms involving central neurotransmitters, ovarian steroids, and neuroactive steroids, existing treatments remain unevenly effective across patients. Common therapeutic strategies include pharmacological interventions such as selective serotonin reuptake inhibitors alongside non-pharmacological modalities like cognitive behavioral therapy and routine exercise, yet individualized management remains challenging. Prior epidemiological research has identified several conventional risk factors, including obesity, smoking, alcohol use, and nutritional deficiencies, as well as psychosocial determinants such as childhood trauma and sexual abuse. Nevertheless, fluctuations in emotional cognition, memory, and subtle regulatory traits across the menstrual cycle suggest that other unexplored factors also contribute to clinical vulnerability.

To explore these complex, multifactorial relationships without confining the analysis to established risk factors, the researchers implemented machine-learning algorithms capable of evaluating high-dimensional biological and psychological data. The study analyzed data collected between December 2018 and April 2021 from 60 women recruited at Kyoto University through posters and flyers, comprising female undergraduate students, graduate students, and university employees. Participants were between 18 and 44 years of age and had regular menstrual cycles. Individuals were excluded if they experienced secondary amenorrhea lasting more than three months, oligomenorrhea with cycles exceeding 39 days, or used hormonal medications such as oral contraceptives, or had magnetic resonance imaging contraindications.

Enrolled participants attended clinical assessments during two distinct phases of their menstrual cycle: once during the follicular phase, defined as five to ten days after the onset of menstrual bleeding, and once during the late-luteal phase, defined as within five days prior to the subsequent menstruation. The timing of the late-luteal phase was verified by prospectively tracking the actual onset date of the subsequent menses. Across the cohort, 121 menstrual phase observations were gathered, counting each examined phase as an individual data point. After excluding two participants who lacked completed screening assessments, the final analytical dataset comprised 58 participants providing 117 valid data points.

Classification of premenstrual disorder status was determined using the Japanese version of the Premenstrual Symptom Screening Tool, which operationalizes DSM-IV diagnostic criteria. Participants meeting criteria for moderate-to-severe premenstrual syndrome or premenstrual dysphoric disorder were designated into the premenstrual disorder cohort, which accounted for 17 participants and 34 data points, while those with milder or absent symptoms were assigned to the non-premenstrual disorder group, comprising 41 participants and 83 data points. Explanatory features were constructed from thirteen comprehensive questionnaire instruments administered during study visits, combined with spectral analyses of heartrate variability captured during sitting-to-standing orthostatic testing to monitor sympathetic and parasympathetic autonomic nervous activity.

The evaluation generated an extensive profile of 443 explanatory features spanning autonomic tone, sleep quality, self-esteem, mood, spontaneity, quality of life, and early parental bonding recollections assessed through Parker’s 25-item Parental Bonding Instrument, which measures recalled parenting behaviors during the first sixteen years of life. To identify patterns predictive of premenstrual disorder, the investigators deployed DataRobot software, selecting a Random Forest Classifier as the optimal algorithm. The feature set was subsequently refined using Recursive Feature Elimination, which identified 31 key variables that maximized discriminative capability, and the resulting architecture was validated through stratified five-fold cross-validation.

The final predictive model demonstrated robust diagnostic discrimination on test data, achieving an area under the receiver operating characteristic curve of 0.90, with a 95 percent confidence interval spanning from 0.82 to 0.98. To interpret the model’s inner mechanics and quantify the influence of individual features, the researchers calculated Shapley Additive Explanations values, an approach rooted in cooperative game theory that assigns each variable an importance score reflecting its contribution to the classification outcome. The analysis revealed that early relational dynamics heavily weighted the predictive output, with parental bonding measures occupying a prominent share of the most decisive variables.

Among the twenty features exhibiting the highest Shapley Additive Explanations scores, eight corresponded to items from the Parental Bonding Instrument, including six pertaining to maternal bonding and two to paternal bonding. Four of the six influential maternal features were specifically tied to overprotection rather than parental care. Statistical comparisons between cohorts highlighted distinct item-level differences: participants with premenstrual disorder reported significantly higher scores on the maternal item measuring perceived overprotection, with a median score of 2 compared to 1 among controls, and on maternal invasion of privacy, with a median of 1 compared to 1 but differing interquartile distributions, alongside lower scores on whether the mother provided needed assistance.

While classical statistical hypothesis tests did not detect significant differences between the two groups on broad aggregate scores for overall parental care or general overprotection, machine-learning feature importance isolated specific parent-child interaction items as salient markers. The authors noted that prior research has established associations between parental bonding styles and conditions such as mood disorders, chronic pain, and inflammatory bowel disease, with maternal bonding frequently exhibiting stronger associations in women. However, this study represents the first reported observation connecting recalled parental bonding styles to premenstrual disorder, contrasting previous research that focused primarily on overt childhood maltreatment and abuse.

In interpreting the unexpected predictive prominence of maternal overprotection, the researchers proposed a framework centered on latent affective vulnerability and heightened biological sensitivity. Women predisposed to premenstrual disorders may possess an innate neurobiological sensitivity to normal physiological shifts in estrogen and progesterone during the luteal phase, coupled with lower thresholds for perceiving somatic discomfort. Such affective vulnerability, arising from complex gene-environment interactions, may present challenges in flexible coping early in life, which could elicit more protective behaviors from mothers. Additionally, maternal figures may share inherited affective traits or anxiety tendencies that manifest as overprotective parenting, making it impossible to infer a direct causal pathway from parenting styles to disorder onset.

The authors emphasized important limitations, noting that the study was exploratory and cross-sectional in design, preventing conclusions regarding causality. Key established risk factors, including direct histories of physical trauma, abuse, and socioeconomic status, were not evaluated in the survey, precluding comparative assessments against parental bonding variables. Furthermore, the sample was small, recruited through non-random convenience sampling from a single academic institution in Japan, and may not generalize across broader populations. Finally, premenstrual disorder was evaluated exclusively through the Japanese version of the Premenstrual Symptom Screening Tool, highlighting the necessity for larger, prospective investigations combining psychosocial contexts with biological monitoring.

Subject of Research: Machine learning analysis of maternal bonding and premenstrual disorders

Article Title: Association between maternal overprotection and premenstrual disorder: a machine learning based exploratory study

Article References: Tsuyuki, K., Egawa, M., Ohsuga, T., Ueda, A., Shimada, K., Ueno, T., Hiyoshi, K., Ueda, K., & Mandai, M. (2025). Association between maternal overprotection and premenstrual disorder: a machine learning based exploratory study. BioPsychoSocial Medicine, 19(1), Article 4. https://doi.org/10.1186/s13030-025-00326-y

Image Credits: AI Generated

DOI: 10.1186/s13030-025-00326-y

Keywords: premenstrual disorder, premenstrual syndrome, premenstrual dysphoric disorder, maternal overprotection, parental bonding instrument, machine learning, Random Forest Classifier, Shapley Additive exPlanations, SHAP values, heartrate variability, follicular phase, luteal phase, Premenstrual Symptom Screening Tool, Kyoto University, autonomic nervous system, emotional cognition, recursive feature elimination