Identification of features associated with dysmenorrhea in women aged 20-50 years using logistic regression and synthetic minority oversampling technique.

Observational OA: gold
⚙ AI-generated deep summary by qwen3.7-flash, 2026-11-18 · read from full text ⓘ

This study analyzed a cohort of 1,038 women aged 20 to 50 years from the Taiwan MJ health screening database to identify features associated with self-reported dysmenorrhea. Using logistic regression combined with synthetic minority over-sampling technique to address class imbalance, the researchers evaluated demographic, biochemical, and lifestyle variables for predictive value. The model incorporating SMOTE demonstrated improved sensitivity in detecting dysmenorrhea cases compared to standard logistic regression, identifying factors such as age, menstrual flow characteristics, and parity as significant predictors. Relevance to endometriosis: secondary dysmenorrhea is explicitly noted as being associated with pelvic pathologies including endometriosis, though the study did not distinguish between primary and secondary causes due to reliance on self-reported data.

Read from the paper's body, not the abstract. Not a substitute for reading the paper. No clinical advice. How this works

Abstract

BackgroundDysmenorrhea (Dys) is a prevalent gynecological condition that significantly impairs women's quality of life. This study investigates the key predictive features associated with Dys in Taiwanese women aged 20-50, using logistic regression (LR) with and without the Synthetic Minority Oversampling Technique (SMOTE) to address class imbalance.MethodsData were obtained from the MJ Health Database, involving 1,038 women. Dys was defined via self-report. A total of 41 clinical, biochemical, and lifestyle features were analyzed. Performance metrics-including accuracy, sensitivity, specificity, precision, F1 score, and area under the ROC and PR curves-were computed via 5-fold cross-validation. Logistic regression coefficients and odds ratios were calculated to evaluate feature associations.ResultsWithout SMOTE, the LR model showed high specificity (97.9%) and accuracy (93.2%) but low sensitivity (51.9%). With SMOTE, sensitivity improved significantly (83.2%), although precision and specificity declined, demonstrating a trade-off between detecting positive cases and generating false positives. Both models maintained high AUC scores (ROC AUC ~ 0.92; PR AUC ~ 0.68). Variable importance analysis, based on standardized coefficients, highlighted six key predictive features: childbirth history, age, glutamic pyruvic transaminase (GPT), glutamic oxaloacetic transaminase (GOT), menstrual flow volume, and diastolic blood pressure (DBP).ConclusionSMOTE-enhanced logistic regression offers a highly sensitive model for detecting Dys in imbalanced datasets. The identified predictive features could inform targeted screening interventions in primary care settings, though external validation is required.
Full text 29,520 characters · extracted from pmc-nxml · 5 sections · click to expand

Methods

Data were sourced from the Taiwan MJ cohort, an ongoing prospective health screening cohort of health examinations conducted by the MJ Health Screening Centers in Taiwan. These examinations include more than 1000 biological indicators, such as anthropometric measurements, blood tests, and imaging examinations. Each participant completed a self-administered questionnaire to provide personal and family medical history, current health status, lifestyle behaviors, physical activity, sleep patterns, and dietary habits. The data used in this study were obtained from the MJ clinic. At the time of health examination, all participants provided general consent for future anonymous research use. Informed consent was obtained from all participants for the use of their data. The database is maintained by the MJ Health Research Foundation. All or part of the data used in this research were authorized and obtained from the MJ Health Research Foundation (Authorization Code: MJHRF2025022A). Any interpretations or conclusions derived from these data do not represent the views of MJ Health Research Foundation. The present study was approved by the Institutional Review Board of the Tri-Service General Hospital (IRB No.: A202405083). As a secondary data analysis, it underwent an expedited review process. Dysmenorrhea was defined based on a self-reported response to a standardized questionnaire item asking whether the participant experienced painful menstruation (binary: yes/no). Due to the self-reported nature of the data, we could not distinguish between primary and secondary dysmenorrhea, nor assess pain severity. Initially, 1,574,271 subjects were enrolled. After applying inclusion criteria (female, premenopausal, aged 20–50 years, no cancer diagnosis) and excluding records with missing data via listwise deletion, a final cohort of 1,038 women was included in the analysis (Fig.  1 ). Fig. 1 Flowchart of sample selection from the study cohort Flowchart of sample selection from the study cohort Blood samples were obtained after a 10-hour fasting period. Within one hour of collection, plasma was separated from the blood and stored at -30 degrees Celsius until lipid profile analysis. Total cholesterol and triglyceride (TG) levels were measured using a dry, multilayer analytical slide method, employing the Fuji Dri-Chem 3000 analyzer (Fuji Photo Film, Tokyo, Japan). Serum high-density lipoprotein cholesterol (HDL-C) and low-density lipoprotein cholesterol (LDL-C) concentrations were determined through an enzymatic cholesterol assay following dextran sulfate precipitation. Continuous variables were tested for normality using the Shapiro-Wilk test; independent t-tests or Mann-Whitney U tests were used accordingly. Categorical variables were analyzed using chi-square tests. To address the class imbalance (minority class comprising ~ 10% of the dataset), SMOTE was selected over alternative techniques (e.g., ADASYN, undersampling) as it is the most established and widely validated method for generating synthetic minority samples, thereby avoiding the simple duplication of data seen in basic oversampling while preserving information from the majority class [ 18 , 19 ]. Logistic regression was trained using 5-fold stratified cross-validation to ensure robust performance estimates [ 20 ]. Hyperparameters for SMOTE included k_neighbors = 5. Variable importance was determined using the absolute values of the standardized regression coefficients from the logistic regression model. These values were normalized to sum to 100% to represent the “percentage of importance” for each feature, allowing for direct comparison of the relative contribution of each predictor. The overall workflow of the proposed machine learning framework is illustrated in Fig.  2 . Fig. 2 Proposed machine learning prediction scheme Proposed machine learning prediction scheme

Results

The dataset exhibited a class imbalance ratio of approximately 0.11 (Fig.  3 ). The age distribution demonstrated that the prevalence of dysmenorrhea gradually decreased with increasing age, with the highest proportion observed among women aged 20–24 years (Fig.  4 ). Demographic and clinical characteristics of the 1,038 participants are presented in Tables  2 and 3 . Table  4 summarizes the averaged predictive performance of the logistic regression models with and without SMOTE across five-fold stratified cross-validation. Fig. 3 Class distribution of the study dataset before oversampling. Class 0 denotes participants without dysmenorrhea, and class 1 denotes participants with dysmenorrhea. The marked predominance of class 0 illustrates the class imbalance addressed in the subsequent SMOTE analysis. Abbreviation: SMOTE, Synthetic Minority Over-sampling Technique Class distribution of the study dataset before oversampling. Class 0 denotes participants without dysmenorrhea, and class 1 denotes participants with dysmenorrhea. The marked predominance of class 0 illustrates the class imbalance addressed in the subsequent SMOTE analysis. Abbreviation: SMOTE, Synthetic Minority Over-sampling Technique Fig. 4 Age-specific percentage distribution of participants with and without dysmenorrhea. Stacked bars show the proportion within each 5-year age group; blue represents participants with dysmenorrhea and white represents those without dysmenorrhea. Percentages within each age group sum to 100% Age-specific percentage distribution of participants with and without dysmenorrhea. Stacked bars show the proportion within each 5-year age group; blue represents participants with dysmenorrhea and white represents those without dysmenorrhea. Percentages within each age group sum to 100% Table 2 Participant demographics and the testing conditions for dysmenorrhea (t test) Variables Unit Mean ± SD p -value Dysmenorrhea Non-Dysmenorrhea Age years 34.44 ± 6.74 38.58 ± 6.73 < 0.001*** C-reactive protein (CRP) mg/dL 0.20 ± 0.31 0.18 ± 0.16 0.116 glutamic oxaloacetic transaminase (GOT) IU/L 19.61 ± 7.10 19.51 ± 6.12 0.833 Diastolic blood pressure (DBP) mmHg 68.77 ± 9.33 70.70 ± 9.87 0.007** Albumin (Alb) mg/dL 4.39 ± 0.21 4.34 ± 0.21 < 0.001*** Creatinine (Cr) mg/dL 0.81 ± 0.10 0.80 ± 0.09 0.087 Waist-to-hip ratio (WHR) ratio 0.75 ± 0.05 0.77 ± 0.05 < 0.001*** Total bilirubin (TBIL) mg/dL 0.91 ± 0.36 0.88 ± 0.36 0.294 glutamic pyruvic transaminase (GPT) IU/L 18.82 ± 12.36 20.28 ± 12.66 0.115 Estimated glomerular filtration rate (eGFR) mL/min/1.73 m² 88.23 ± 12.79 87.61 ± 13.11 0.517 Estradiol (E2) pg/mL 109.22 ± 85.47 107.75 ± 95.97 0.830 Sex hormone-binding globulin (SHBG) nmol/L 65.99 ± 47.92 70.66 ± 43.59 0.160 Uric acid (UA) mg/dL 4.73 ± 1.02 4.65 ± 1.04 0.328 Gamma-glutamyl transferase (GGT) IU/L 19.10 ± 16.75 20.70 ± 17.04 0.200 Fasting plasma glucose (FPG) mg/dL 94.75 ± 9.93 96.69 ± 11.49 0.017* Globulin (Glo) g/dL 3.11 ± 0.32 3.13 ± 0.33 0.397 Systolic blood pressure (SBP) mmHg 105.91 ± 13.04 108.12 ± 14.46 0.033* Follicle-stimulating hormone (FSH) mIU/mL 5.34 ± 6.73 7.71 ± 11.59 0.002** Platelets (Plt) *10^3/µL 250.99 ± 55.90 249.52 ± 55.66 0.721 Leukocyte (WBC) *10^3/µL 6.14 ± 1.55 6.06 ± 1.59 0.469 Triglycerides (TG) mg/dL 76.76 ± 40.95 81.53 ± 51.63 0.182 Drink area - 1.88 ± 8.35 2.17 ± 8.51 0.642 Sport area - 3.96 ± 5.75 4.86 ± 6.46 0.050* * : p  < 0.05 ** : p  < 0.01 *** : p  < 0.001 Participant demographics and the testing conditions for dysmenorrhea (t test) * : p  < 0.05 ** : p  < 0.01 *** : p  < 0.001 Table 3 Participant demographics and the testing conditions for dysmenorrhea (Chi-squared test) Variables N (%) p -value Dysmenorrhea Non-Dysmenorrhea Menstrual flow Normal 503 (88.1%) 254 (93.4%) 0.018* Heavy 68 (11.9%) 18 (6.6%) Menstrual cycle phase Follicular phase 193 (33.8%) 78 (28.7%) 0.233 Ovulation phase 19 (3.3%) 13 (4.8%) Luteal phase 359 (62.9%) 181 (66.5%) Irregular menstrual cycles No 407 (71.3%) 208 (76.5%) 0.113 Yes 164 (28.7%) 64 (23.5%) Given birth No 386 (67.6%) 112 (41.2%) < 0.001*** Yes 185 (32.4%) 160 (58.8%) Period duration Less than 3 days 32 (5.6%) 22 (8.1%) 0.137 3–7 days 458 (80.2%) 222 (81.6%) More than 7 days 81 (14.2%) 28 (10.3%) * : p  < 0.05 ** : p  < 0.01 *** : p  < 0.001 Participant demographics and the testing conditions for dysmenorrhea (Chi-squared test) * : p  < 0.05 ** : p  < 0.01 *** : p  < 0.001 Table 4 Overview of the averaged performance metrics for both the logistic regression model with and without synthetic minority oversampling technique, across the 5-fold stratified cross-validation repetitions Without SMOTE With SMOTE mean std mean std Accuracy 0.9322 0.014367 0.8914 0.017708 Precision 0.732647 0.088031 0.484255 0.051777 Sensitivity 0.519143 0.125929 0.832095 0.077207 Specificity 0.978638 0.009047 0.898112 0.023001 fpr 0.021362 0.009047 0.101888 0.023001 f1_score 0.60088 0.1056 0.608899 0.040745 roc_auc 0.924407 0.030905 0.92162 0.031888 pr_auc 0.685626 0.079681 0.680724 0.077909 SMOTE Synthetic Minority Over-sampling Technique, fpr False Positive Rate, roc_auc Receiver Operating Characteristic – Area Under the Curve, pr_auc Precision-Recall – Area Under the Curve Overview of the averaged performance metrics for both the logistic regression model with and without synthetic minority oversampling technique, across the 5-fold stratified cross-validation repetitions SMOTE Synthetic Minority Over-sampling Technique, fpr False Positive Rate, roc_auc Receiver Operating Characteristic – Area Under the Curve, pr_auc Precision-Recall – Area Under the Curve Without SMOTE, the LR model achieved high accuracy (93.2%) and specificity (97.9%) but relatively low sensitivity (51.9%). After applying SMOTE, sensitivity improved substantially to 83.2%, whereas precision (48.4%) and specificity (89.8%) decreased, reflecting the expected trade-off associated with addressing class imbalance. Despite these changes, both models maintained strong discriminative performance (ROC AUC ~ 0.92; PR AUC ~ 0.68), as illustrated by the representative ROC and precision–recall curves (Figs.  5 and 6 ). Representative precision–recall curves for the models with and without SMOTE are presented in Fig.  5 , demonstrating similar precision–recall performance across the repeated cross-validation folds. Similarly, the representative ROC curves demonstrated consistently high discriminative ability for both models across all cross-validation folds (Fig.  6 ). ROC AUC distributions were comparable between the SMOTE and non-SMOTE models across repeated cross-validation runs, indicating that SMOTE improved class balance without compromising the stability of model discrimination (Fig.  7 ). Fig. 5 Representative precision–recall curves from the first repetition of the stratified 5-fold cross-validation. Panel ( a ) shows the model trained without SMOTE, and panel ( b ) shows the model trained with SMOTE. Colored curves represent the individual validation folds, with fold-specific PR AUC values reported in the legends. Recall is equivalent to sensitivity, and precision is the positive predictive value. Abbreviations: PR, precision–recall; AUC, area under the curve; SMOTE, Synthetic Minority Over-sampling Technique Representative precision–recall curves from the first repetition of the stratified 5-fold cross-validation. Panel ( a ) shows the model trained without SMOTE, and panel ( b ) shows the model trained with SMOTE. Colored curves represent the individual validation folds, with fold-specific PR AUC values reported in the legends. Recall is equivalent to sensitivity, and precision is the positive predictive value. Abbreviations: PR, precision–recall; AUC, area under the curve; SMOTE, Synthetic Minority Over-sampling Technique Fig. 6 Representative receiver operating characteristic curves from the first repetition of the stratified 5-fold cross-validation. Panel ( a ) shows the model trained without SMOTE, and panel ( b ) shows the model trained with SMOTE. Colored curves represent the individual validation folds, with fold-specific ROC AUC values reported in the legends; the diagonal dashed line denotes discrimination expected by chance. Abbreviations: ROC, receiver operating characteristic; AUC, area under the curve; SMOTE, Synthetic Minority Over-sampling Technique Representative receiver operating characteristic curves from the first repetition of the stratified 5-fold cross-validation. Panel ( a ) shows the model trained without SMOTE, and panel ( b ) shows the model trained with SMOTE. Colored curves represent the individual validation folds, with fold-specific ROC AUC values reported in the legends; the diagonal dashed line denotes discrimination expected by chance. Abbreviations: ROC, receiver operating characteristic; AUC, area under the curve; SMOTE, Synthetic Minority Over-sampling Technique Fig. 7 Distribution of ROC AUC values for logistic regression models trained with and without SMOTE across five repetitions of stratified 5-fold cross-validation (25 validation-fold estimates per model). The center line indicates the median, the box represents the interquartile range, whiskers extend to 1.5 times the interquartile range, and points beyond the whiskers denote outliers. Abbreviations: ROC, receiver operating characteristic; AUC, area under the curve; SMOTE, Synthetic Minority Over-sampling Technique Distribution of ROC AUC values for logistic regression models trained with and without SMOTE across five repetitions of stratified 5-fold cross-validation (25 validation-fold estimates per model). The center line indicates the median, the box represents the interquartile range, whiskers extend to 1.5 times the interquartile range, and points beyond the whiskers denote outliers. Abbreviations: ROC, receiver operating characteristic; AUC, area under the curve; SMOTE, Synthetic Minority Over-sampling Technique Table  5 details the logistic regression coefficients, odds ratios (OR), and 95% confidence intervals for the final model. Table 5 Logistic regression coefficients, odds ratios, and 95% confidence intervals for predictors of dysmenorrhea (Excluding Betel Nut Use) Variable β Coefficient SE Wald Z p value Odds Ratio CI Lower CI Upper (Intercept) 2.6177 3.5156 0.7446 0.4565 13.7045 0.0139 13468.82 Age* -0.0402 0.0201 -1.9975 0.0458 0.9606 0.9234 0.9992 marriage 0.1175 0.2356 0.4988 0.6179 1.1247 0.7088 1.7846 education -0.0168 0.1047 -0.1604 0.8725 0.9833 0.8009 1.2073 Income** -0.2027 0.0714 -2.8401 0.0045 0.8166 0.71 0.9391 WHR -2.3116 2.1244 -1.0881 0.2765 0.0991 0.0015 6.3733 WBC -0.0494 0.0752 -0.657 0.5112 0.9518 0.8214 1.1029 Hb 0.0719 0.0828 0.8679 0.3855 1.0745 0.9135 1.2639 Platelet 0.0034 0.0019 1.7661 0.0774 1.0034 0.9996 1.0071 SBP 6.00E-04 0.0106 0.0538 0.9571 1.0006 0.9801 1.0215 DBP -0.0172 0.0149 -1.1569 0.2473 0.9829 0.9546 1.012 FPG -0.003 0.0111 -0.2705 0.7868 0.997 0.9756 1.0189 T Bili 0.3963 0.2935 1.3503 0.1769 1.4864 0.8361 2.6423 Alb 0.5975 0.4756 1.2563 0.209 1.8175 0.7156 4.6165 Glo -0.4061 0.3075 -1.3206 0.1866 0.6662 0.3646 1.2173 ALP -3.00E-04 0.0071 -0.036 0.9713 0.9997 0.9859 1.0138 GOT* 0.0783 0.0307 2.5487 0.0108 1.0815 1.0183 1.1486 GPT** -0.0467 0.0175 -2.6737 0.0075 0.9544 0.9223 0.9876 r-GT 0.0036 0.0072 0.5068 0.6123 1.0036 0.9896 1.0179 LDH -0.0052 0.0044 -1.1832 0.2367 0.9949 0.9864 1.0034 eGFR -0.0077 0.0078 -0.9838 0.3252 0.9924 0.9773 1.0076 UA 0.164 0.106 1.5476 0.1217 1.1782 0.9572 1.4503 TG 6.00E-04 0.0027 0.2352 0.8141 1.0006 0.9954 1.0059 HDLC -3.00E-04 0.0073 -0.0407 0.9676 0.9997 0.9855 1.0141 LDLC 0.0026 0.0035 0.744 0.4569 1.0026 0.9958 1.0095 CRP 0.696 0.4858 1.4327 0.1519 2.0057 0.7741 5.197 PRD -0.1124 0.1053 -1.0673 0.2859 0.8937 0.727 1.0986 FSH -0.0225 0.0153 -1.4731 0.1407 0.9778 0.949 1.0075 LH 0.0081 0.0141 0.5775 0.5636 1.0082 0.9807 1.0364 TTT 1.0509 0.9806 1.0718 0.2838 2.8603 0.4186 19.5461 SHBG -0.0018 0.0024 -0.7559 0.4497 0.9982 0.9936 1.0028 E2 -0.0011 0.0012 -0.9049 0.3655 0.9989 0.9965 1.0013 Prolactin 0.0019 0.0072 0.2653 0.7907 1.0019 0.9878 1.0163 Menstrual flow** 0.9591 0.3596 2.6674 0.0076 2.6093 1.2896 5.2792 Period duration 0.264 0.2706 0.9758 0.3292 1.3022 0.7662 2.2131 Irregular menstrual cycles 0.015 0.2391 0.0626 0.9501 1.0151 0.6353 1.6218 Given birth** -0.7856 0.2582 -3.043 0.0023 0.4559 0.2748 0.7561 Contraception -0.1602 0.2034 -0.7876 0.4309 0.852 0.5719 1.2693 Drinking -0.0029 0.0134 -0.2132 0.8311 0.9972 0.9714 1.0236 Drink area 0.0086 0.0147 0.5865 0.5575 1.0086 0.9801 1.0381 Sport area* -0.0383 0.0164 -2.3417 0.0192 0.9624 0.932 0.9938 Sleep -0.0185 0.1035 -0.179 0.8579 0.9816 0.8014 1.2024 The dependent outcome variable is the self-reported presence of dysmenorrhea (1 = Yes, 0 = No). The variable tracking betel nut usage (Betel.net) was excluded from this multivariate analysis due to extreme data sparsity causing complete separation and multi-collinearity boundaries, stabilizing overall matrix convergence β regression coefficient, SE standard error, Wald Z Wald statistic, p-value statistical significance level, OR odds ratio, CI confidence interval Comprehensive abbreviations utilized in this table include: Alb Albumin, ALP Alkaline Phosphatase, B Standardized Regression Coefficient, CI Confidence Interval, CRP C-Reactive Protein, Days MC Days of Menstrual Cycle, DBP Diastolic Blood Pressure, E2 Estradiol, eGFR Estimated Glomerular Filtration Rate, FPG Fasting Plasma Glucose, FSH Follicle-Stimulating Hormone, Glo Globulin, GOT Glutamic Oxaloacetic Transaminase / Aspartate Aminotransferase, GPT Glutamic Pyruvic Transaminase / Alanine Aminotransferase, Hb Hemoglobin, HDLC High-Density Lipoprotein Cholesterol, LDH Lactate Dehydrogenase, LDLC Low-Density Lipoprotein Cholesterol, LH Luteinizing Hormone, MC irregular Irregular Menstrual Cycle, OR Odds Ratio, PRD Period Duration, r-GT Gamma-Glutamyl Transferase, SBP Systolic Blood Pressure, SE Standard Error, SHBG Sex Hormone-Binding Globulin, T Bili Total Bilirubin, TG Triglycerides, TTT Total Testosterone, UA Uric Acid, WBC White Blood Cell / Leukocyte, and WHR Waist-to-Hip Ratio Statistical significance is indicated by * : p  < 0.05, * *: p  < 0.01, and *** : p  < 0.001 Logistic regression coefficients, odds ratios, and 95% confidence intervals for predictors of dysmenorrhea (Excluding Betel Nut Use) The dependent outcome variable is the self-reported presence of dysmenorrhea (1 = Yes, 0 = No). The variable tracking betel nut usage (Betel.net) was excluded from this multivariate analysis due to extreme data sparsity causing complete separation and multi-collinearity boundaries, stabilizing overall matrix convergence β regression coefficient, SE standard error, Wald Z Wald statistic, p-value statistical significance level, OR odds ratio, CI confidence interval Comprehensive abbreviations utilized in this table include: Alb Albumin, ALP Alkaline Phosphatase, B Standardized Regression Coefficient, CI Confidence Interval, CRP C-Reactive Protein, Days MC Days of Menstrual Cycle, DBP Diastolic Blood Pressure, E2 Estradiol, eGFR Estimated Glomerular Filtration Rate, FPG Fasting Plasma Glucose, FSH Follicle-Stimulating Hormone, Glo Globulin, GOT Glutamic Oxaloacetic Transaminase / Aspartate Aminotransferase, GPT Glutamic Pyruvic Transaminase / Alanine Aminotransferase, Hb Hemoglobin, HDLC High-Density Lipoprotein Cholesterol, LDH Lactate Dehydrogenase, LDLC Low-Density Lipoprotein Cholesterol, LH Luteinizing Hormone, MC irregular Irregular Menstrual Cycle, OR Odds Ratio, PRD Period Duration, r-GT Gamma-Glutamyl Transferase, SBP Systolic Blood Pressure, SE Standard Error, SHBG Sex Hormone-Binding Globulin, T Bili Total Bilirubin, TG Triglycerides, TTT Total Testosterone, UA Uric Acid, WBC White Blood Cell / Leukocyte, and WHR Waist-to-Hip Ratio Statistical significance is indicated by * : p  < 0.05, * *: p  < 0.01, and *** : p  < 0.001 Based on the standardized coefficients, the top six predictive features for Dys, in descending order of importance, were: childbirth history (negative association), age, GPT, GOT, menstrual flow volume, and DBP (Figs.  8 and 9 ). Features with negative standardized coefficients indicated a protective or inverse association with the presence of Dys. The variable importance rankings were highly consistent between the models trained with and without SMOTE, with childbirth history and age remaining the two strongest predictors in both models (Figs.  8 and 9 ). Although the relative importance values differed slightly after SMOTE, the overall ranking of the major predictive features remained largely unchanged, indicating that SMOTE improved class balance without substantially altering the clinical interpretation of the model. Fig. 8 Variable importance based on standardized logistic regression coefficients (with SMOTE). Note: Variable importance was calculated using the absolute values of the standardized logistic regression coefficients from the model trained with SMOTE. The coefficients were normalized to sum to 100%, allowing comparison of the relative contribution of each predictor Variable importance based on standardized logistic regression coefficients (with SMOTE). Note: Variable importance was calculated using the absolute values of the standardized logistic regression coefficients from the model trained with SMOTE. The coefficients were normalized to sum to 100%, allowing comparison of the relative contribution of each predictor Fig. 9 Variable importance based on standardized logistic regression coefficients (without SMOTE). Note: Variable importance was calculated using the absolute values of the standardized logistic regression coefficients from the model trained with SMOTE. The coefficients were normalized to sum to 100%, allowing comparison of the relative contribution of each predictor Variable importance based on standardized logistic regression coefficients (without SMOTE). Note: Variable importance was calculated using the absolute values of the standardized logistic regression coefficients from the model trained with SMOTE. The coefficients were normalized to sum to 100%, allowing comparison of the relative contribution of each predictor

Conclusion

In a cohort of 1,038 Taiwanese women aged 20–50, SMOTE-enhanced logistic regression significantly improved the sensitivity of dysmenorrhea detection. Childbirth history, age, GPT, GOT, menstrual flow, and DBP were the top predictive features. These findings highlight the utility of resampling techniques in imbalanced medical data and suggest potential variables for clinical screening, though prospective external validation is required.

Discussion

This study demonstrates that SMOTE-enhanced logistic regression effectively improves the detection of dysmenorrhea in imbalanced clinical datasets. The top predictive features identified—childbirth history, age, liver enzymes, menstrual flow, and DBP—provide potential targets for preliminary screening in primary care settings. Childbirth history emerged as the strongest predictive feature, consistent with literature indicating that parity, particularly vaginal delivery, is associated with a reduction in primary dysmenorrhea severity due to physiological changes in uterine contractility [ 21 , 22 ]. Age was the second most important predictor, aligning with evidence that dysmenorrhea prevalence and severity often peak in adolescence and early adulthood before declining [ 22 , 23 ]. Interestingly, GPT and GOT were identified as significant predictive features. Although GPT and GOT did not demonstrate significant differences in univariate analysis (Table 2 ), they emerged as significant predictors in the multivariate logistic regression model (Table 5 ), suggesting that their predictive capacity is revealed when adjusting for confounding metabolic and demographic clinical features. While direct causal mechanisms linking liver enzymes to dysmenorrhea remain unestablished, we hypothesize that these markers may reflect underlying systemic metabolic or inflammatory states. Some population-based studies suggest associations between liver enzyme levels and menstrual irregularities, potentially mediated by gonadal steroid metabolism [ 24 ]. Similarly, the predictive value of DBP may hint at subtle autonomic nervous system alterations or vascular resistance changes associated with the pain and stress of dysmenorrhea [ 25 ]. However, these biological links are highly speculative and require targeted mechanistic studies to validate. It is worth noting that while DBP did not achieve independent statistical significance in the final multivariate regression matrix ( p = 0.2473), it retained high variable importance based on its absolute standardized coefficient. In machine learning frameworks handling highly imbalanced classes, certain features serve as vital structural boundary contributors—acting in concert with other clinical markers to optimize the overall decision boundary—even if they do not exhibit isolated statistical significance. The SMOTE-enhanced model achieved a sensitivity of 83.2%, suggesting its potential utility as a preliminary screening tool in primary healthcare or community settings. By utilizing routinely collected biochemical and demographic data, clinicians could identify women at high risk for severe dysmenorrhea, prompting further gynecological evaluation or early lifestyle interventions. The trade-off of lower precision (higher false positives) is generally acceptable in a screening context, where the primary goal is to avoid missing true cases, which can then be ruled out via clinical examination. Several limitations must be acknowledged. First, the cross-sectional design precludes any causal inference. Second, dysmenorrhea was self-reported and binary; we could not distinguish between primary and secondary dysmenorrhea, nor account for pain severity. Third, we lacked data on potential confounding factors such as the use of oral contraceptives, NSAIDs, or a history of specific gynecological diseases (e.g., endometriosis). Fourth, the study relied on a specific health examination cohort, which may introduce selection bias. Fifth, the sample size of the positive class was relatively limited, and the model lacks external validation in an independent cohort. Finally, while LR provides excellent interpretability, it may not capture complex non-linear relationships as effectively as ensemble tree-based models; however, LR was chosen specifically for its clinical transparency and the ability to calculate odds ratios.

Introduction

Dysmenorrhea (Dys) refers to cramp-like pain in the lower abdomen occurring before or during menstruation. Beyond physical pain, Dys negatively affects psychological, social, and occupational functioning, and is linked to long-term risks of anxiety and depression [ 1 , 2 ]. Dys is one of the most common gynecological complaints globally, with prevalence rates among menstruating individuals ranging from 45 to 95% [ 3 , 4 ]. In Taiwan, studies indicate a high prevalence among young and working-age women [ 5 , 6 ]. Primary Dys is driven by excessive uterine prostaglandin production, while secondary Dys is associated with pelvic pathologies such as endometriosis [ 7 ]. Machine learning (ML) is rapidly transforming healthcare by enabling the analysis of complex datasets for early diagnosis and predictive modeling [ 8 – 10 ]. While several studies have applied ML to gynecological conditions [ 11 – 14 ] (Table 1 ), existing work has not comprehensively examined the combined roles of demographic, biochemical, and lifestyle factors in Dys prediction using interpretable models. Table 1 Summary of machine learning studies on dysmenorrhea Study/Area ML Technique Used Application Key Outcomes Neuroimaging in PD SVM classifier Brain activity analysis 96% accuracy in pain classification [ 11 ] Wearable Biometric Data (Planned ML use) Physiological signal tracking Feasibility for ML-based pain prediction [ 12 ] Endometriosis Prediction XGBoost, Random Forest App-based symptom data 89% accuracy in endometriosis classification [ 13 ] Clinical Decision Support Multiple ML classifiers Preoperative diagnosis Improved prediction of endometriosis [ 14 ] Wearable Technology AI/ML algorithms Personalized symptom tracking Adaptive and personalized pain insights [ 13 ] PD Primary Dysmenorrhea, SVM Support Vector Machine, ML Machine Learning, AI Artificial Intelligence, XGBoost Extreme Gradient Boosting Summary of machine learning studies on dysmenorrhea 96% accuracy in pain classification [ 11 ] Feasibility for ML-based pain prediction [ 12 ] 89% accuracy in endometriosis classification [ 13 ] Improved prediction of endometriosis [ 14 ] Adaptive and personalized pain insights [ 13 ] PD Primary Dysmenorrhea, SVM Support Vector Machine, ML Machine Learning, AI Artificial Intelligence, XGBoost Extreme Gradient Boosting In the present study, we utilized logistic regression (LR) combined with the synthetic minority over-sampling technique (SMOTE) in a cohort of women aged 20 to 50. Our goals were to: (1) compare the predictive performance of LR with and without SMOTE in handling class imbalance, and (2) identify the top predictive features associated with Dys to inform potential clinical screening tools. Recent advances in explainable artificial intelligence and machine learning for reproductive healthcare further highlight the importance of interpretable predictive frameworks [ 15 – 17 ].

Text is read by the "Ask this paper" AI Q&A widget below. Extraction quality varies by source — PMC NXML preserves structure cleanly, OA-HTML may include some navigation residue, and OA-PDF can have broken hyphenation. The publisher copy (via DOI) is the canonical version.

My notes (saved in your browser only)

⚙ Ask this paper AI returns verbatim quotes from the full text · source: pmc-nxml ⓘ

Answers must be backed by verbatim quotes from this paper's full text. Hallucinated quotes are dropped automatically; if no verbatim passage answers the question, we say so. How this works

Citation neighborhood (no data yet)

We don't have any in-corpus citations linked to this paper yet. This is a recent paper (2026) — citers typically take a year or two to land, and the OpenAlex reference graph may still be filling in.

Source provenance

europepmc
last seen: 2026-11-15T09:34:25.365398+00:00