Integrating social cognitive theory with machine learning to predict MSM-women sexual behavior: a multicenter random forest model development study in China

Abstract

Understanding and predicting sexual behavior patterns through the lens of Social Cognitive Theory (SCT) is essential for identifying behavioral bridges that facilitate HIV transmission. MSM who engage in sexual activity with women (MSM-W) represent a critical yet underexplored subgroup in China. However, existing studies lack theory-driven predictive frameworks that explicitly integrate psychological mechanisms with robust analytics. This study aimed to develop and validate an SCT-guided machine-learning model to predict MSM-W behavior and to evaluate its discriminatory and calibration performance across population subgroups. We developed a theory-driven machine-learning framework integrating SCT with a random forest–guided logistic regression model. Data were collected from 2,403 MSM across six urban regions in China. Model development included: (1) SCT-guided candidate variable construction, (2) random-forest feature screening using mean decrease accuracy (MDA), (3) nested cross-validation and bootstrap optimism correction, (4) calibration assessment, and (5) threshold selection using the Youden index. Model performance was evaluated using ROC-AUC, PR-AUC, Brier score, calibration slope/intercept, subgroup analyses (age, education, marital status, migrant status), and comparison with a baseline demographic-only model. Among all participants, 406 (16.9%) reported sexual behavior with women in the past six months. The final SCT-guided model incorporating nine predictors (GAD-7, PHQ-9, self-esteem, age, education level, marital status, sexual orientation, recent group sex with men, and pre-sex drug use) demonstrated good discrimination (AUC = 0.80), moderate PR-AUC (0.53), and good overall accuracy (Brier score = 0.107). Calibration analysis showed a slope of 0.91 (95% CI: 0.76–1.09) and intercept of − 0.20 (95% CI: −0.53 to 0.15), indicating acceptable agreement between predicted and observed risks. The optimal probability threshold selected by the Youden index was 0.148 (sensitivity 0.72; specificity 0.77; NPV 0.93). Subgroup analyses demonstrated stable performance across age, education, marital, and migrant strata. Compared with the baseline demographic-only model (AUC = 0.76; Brier = 0.113), the full SCT-integrated model showed consistent improvement. Results were translated into an individualized nomogram for practical risk assessment. This study provides a theory-informed and statistically robust framework for predicting MSM-W behavior. By integrating SCT with machine learning, the model captures the joint influence of psychological distress, self-evaluation, and behavioral context on bisexual behavior patterns. The tool offers a non-stigmatizing, early-screening approach for identifying behavioral risk bridges in MSM populations and may inform targeted HIV prevention strategies in China.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-36202-0.

Introduction

Predicting human sexual behavior patterns within complex social networks represents a fundamental challenge in behavioral science, particularly when viewed through the lens of Social Cognitive Theory (SCT). This theoretical framework becomes especially relevant when examining the intricate behavioral dynamics of men who have sex with men (MSM)¹, where sexual decision-making often transcends traditional analytical paradigms. The global MSM population, estimated at 39.6 million², with 12.01 million in China³, exhibits complex behavioral patterns deeply influenced by sociocultural determinants and psychological factors. Of particular significance is the concurrent sexual engagement with women, a behavior pattern shaped by social pressures, cultural expectations, and individual cognitive processes central to SCT frameworks^4,5. In China, this phenomenon manifests distinctively, with over 25% of MSM engaging in heterosexual relationships⁶ and 38–70% entering traditional marriages⁷, creating unique bridge networks that challenge conventional behavioral predictions. This pattern is primarily driven by strong sociocultural pressures in China, including filial piety, expectations of marriage and offspring continuity, and pervasive stigma surrounding same-sex orientation. Many MSM conceal their sexual identity through heterosexual marriage to maintain family harmony and social acceptance. This dual-identity context creates complex psychological stress, increases hidden transmission risk, and complicates conventional HIV prevention strategies, especially for female partners who are often unaware of their partner’s same-sex behavior. The public health implications are particularly concerning, as studies consistently demonstrate higher rates of unprotected intercourse with female partners compared to male partners^8–10, with condom use reported in less than 25% of encounters with regular female partners⁶, highlighting the urgent need for theory-driven predictive approaches.

The intricate nature of these behavioral patterns demands advanced prediction methodologies informed by SCT frameworks. While traditional interventions such as counseling and behavioral education have shown efficacy¹¹, their implementation remains challenging, particularly in hidden populations¹². Social cognitive research has identified key determinants: psychological factors, evidenced by prevalent depressive symptoms in Chinese MSM¹³; demographic characteristics, specifically age and education level variations¹⁴; and contextual influences, such as substance use in sexual settings¹⁵. Specifically, within the SCT framework, depressive symptoms and anxiety reflect outcome expectations and emotional regulation; self-esteem represents self-efficacy; peer sexual norms and group sex behaviors reflect observational learning; and substance use before sex reflects situational facilitation. These theory-guided constructs directly inform the selection of candidate predictors for our machine-learning model, ensuring that variable inclusion is grounded in established behavioral mechanisms rather than purely data-driven selection.

Traditional analytical methods have shown limitations in capturing the complexity of MSM sexual behaviors, particularly in understanding the interplay of multiple determinants within specific sociocultural contexts¹⁶. By integrating SCT with machine learning techniques, our study offers an innovative methodological framework. SCT’s fundamental principles of reciprocal determinism, observational learning, and self-efficacy provide theoretical insights into the dynamic interactions among individual cognition, behavior patterns, and environmental influences¹⁷. These theoretical constructs guide our variable selection across multiple dimensions: personal factors (psychological states, self-efficacy beliefs), behavioral patterns (sexual practices, risk behaviors), and environmental influences (social support, cultural pressures)^18,19. The Random Forest algorithm, with its capability to handle complex variable relationships while maintaining interpretability, aligns well with these SCT constructs, enabling systematic evaluation of behavioral predictors within this multidimensional theoretical framework. The present study aims to develop a theoretically-driven predictive model for MSM-women sexual engagement by integrating SCT with machine learning approaches. Accordingly, this study addresses the following research questions:

1. Can an SCT-guided machine-learning model accurately predict MSM–women sexual behavior?

2. Do SCT-based psychological and behavioral constructs provide incremental predictive value beyond demographic variables alone?

3. Does the model demonstrate stable predictive performance across key demographic subgroups such as age, education level, marital status, and migration status?

Through systematic examination of personal factors, behavioral patterns, and environmental influences within the SCT framework, this research seeks to identify key determinants of MSM-women sexual behaviors and evaluate their predictive significance using Random Forest algorithms. This integration of theoretical framework and advanced analytics will not only enhance our understanding of complex behavioral patterns in hidden populations but also provide evidence-based insights for developing targeted interventions, thereby advancing both theoretical applications and intervention strategies in behavioral health research. Model robustness and clinical applicability were further evaluated using nested cross-validation, bootstrap optimism correction, calibration analysis, subgroup performance assessment, and comparisons with a demographics-only baseline model, ensuring methodological rigor and transportability.

Methods

Study design and participant recruitment

An anonymous cross-sectional survey was conducted from November 2023 to March 2024 across six regions in China: Shanghai, Guangzhou, Xinjiang, Shaanxi, Yunnan, and Shenyang. Given the hidden nature of the MSM population, we established partnerships with local non-governmental organizations (NGOs) that maintain trusted relationships with MSM communities. NGO staff disseminated study information through their networks and actively engaged potential participants, providing comprehensive details about the study objectives and significance. Given the sensitive nature of the study topics and the hidden characteristics of the target population, all key behavioral variables were self-reported by participants through anonymous electronic questionnaires to reduce social desirability bias as much as possible.

Eligible participants met the following criteria: (1) male, aged ≥ 18 years; (2) reported male-to-male sexual behavior within the previous six months; and (3) primary residence in one of the study regions during the past six months. We excluded responses based on: (1) completion time < 300 s; (2) incorrect answers to validation questions; and (3) IP addresses outside study regions.

To ensure participant privacy and confidentiality, data collection was conducted through the QuestionStar platform, accessible via personal mobile devices. All participants provided informed consent after receiving detailed information about the study’s purpose, procedures, rights, and potential risks. The data collection process strictly adhered to privacy protection regulations.

Community organization involvement

This study was conducted in collaboration with six non-governmental organizations serving the MSM community: Shanghai Qingai Health Center (Shanghai), Consultation Center of AIDS Aid and Health Service (Liaoning), Trans China (Yunnan), Xinjiang Rong’ai Health Consulting Co., Ltd. (Xinjiang), Xi’an Zhitong AIDS Prevention and Health Center (Shaanxi), and Lingnan Partners Community Support Center (Guangdong). These organizations, which provide comprehensive health counseling services and are dedicated to improving MSM physical and mental health, facilitated participant recruitment and community engagement throughout the study.

Sample size and quality control

The sample size calculation was based on an estimated 25% prevalence of heterosexual behavior among MSM⁶, with a precision of 2.5% at a 95% confidence level. Using PASS software for cross-sectional survey design, we determined a minimum required sample size of 1,200. To account for the use of convenience sampling, a design effect of 1.5 was applied. Additionally, considering an anticipated 20% non-response rate, the adjusted minimum required sample size was calculated to be 2,250. Of the 2,481 questionnaires collected, 2,403 met validity criteria (96.86%), exceeding the calculated minimum requirement and ensuring sufficient statistical power. The geographic distribution of valid responses demonstrated appropriate regional representation: Shanghai (n = 569), Guangzhou (n = 500), Shenyang (n = 502), Xinjiang (n = 320), Yunnan (n = 313), and Shaanxi (n = 199). The sampling distribution is illustrated in Fig. 1.

Fig. 1.

Research framework and model development flowchart for predicting MSM-Female sexual engagement.

The design of the survey questionnaire incorporated advice and suggestions from multiple experts. After completing the initial draft, a pre-survey was conducted with several MSM participants. Based on the pre-survey results, modifications were made to the questionnaire, including adjustments in vocabulary and expressions to better suit the MSM population, enhancing the survey’s acceptability. For survey implementation, the study primarily relied on local NGO platforms, with NGO staff serving as researchers. Uniform training was provided to ensure these researchers understood the survey’s objectives, content, procedures, methods, and precautions. A data quality monitoring team was established to regularly check and provide feedback on the data within 48 h during data collection. After data collection, steps for data cleaning and processing were undertaken, including checking for outliers, missing data, and logical errors, and making necessary corrections and adjustments. Participants were given a malpractice allowance for completing the questionnaire to increase compliance with the survey.

To ensure data quality, multiple control measures were implemented. The questionnaire was developed through expert consultation and refined based on pre-survey feedback from MSM participants, particularly focusing on cultural appropriateness and acceptability of vocabulary and expressions. NGO staff, serving as field researchers, received standardized training on survey protocols, including objectives, content, procedures, and quality control measures. During data collection, a dedicated monitoring team conducted 48-hour review cycles, providing timely feedback. Post-collection quality assurance included systematic screening for outliers, missing data, and logical inconsistencies, with necessary corrections applied. Participant compensation was provided to enhance survey completion rates while maintaining data integrity.

Study variables and measurements

Definition of primary outcome

MSM-women sexual engagement, the primary outcome of this study, was defined as any sexual behavior between a male participant and a female partner within the past six months. This included both penetrative vaginal intercourse and other forms of sexual contact. Participants were asked “Have you engaged in any sexual behavior with women in the past six months?” with response options of “yes” or “no”. Those who responded “yes” were further asked about the frequency of sexual encounters, number of female sexual partners, and condom use during these encounters.

Predictor variables

Based on SCT, we incorporated 28 variables across personal, behavioral, and environmental domains to develop a predictive model.

Personal factors consisted of both continuous and categorical variables. Continuous variables included age and scores from standardized psychological assessments: the GAD-7 for anxiety²⁰, PHQ-9 for depression²¹, Rosenberg Self-Esteem Scale²², and Personality Scale²³. Internal consistency of all psychometric instruments was evaluated using Cronbach’s α coefficients in the present MSM sample. Categorical variables included self-identified sexual orientation (heterosexual, homosexual, bisexual, or not sure), recent HIV test results (positive, negative, unknown), and STI diagnosis in the past six months (positive vs. negative). Chronic health conditions, including hypertension, diabetes, and hyperlipidemia, were coded as positive, negative, or unknown.

Behavioral patterns encompassed substance use and specific sexual behaviors. Substance use was assessed by smoking frequency in the past 30 days (not smoking, sometimes, often, everyday) and alcohol consumption (not drinking, sometimes, often). Sexual behaviors included the number of male anal sex partners in the past six months (1, 2, or ≥ 3), frequency of anal sex with men (ranging from < 1 time/month to > 4 times/week), and sexual role (bottom, top, or versatile). Condom use in anal sex was categorized as never, seldom, often, or every time. Additionally, we included binary variables (Yes / No) for recent voluntary HIV testing, group sex with men, alcohol use before sex, and drug use before sex.

Environmental influences included sociodemographic characteristics and psychosocial support. Educational level was stratified into junior high and below, senior high, or college and above. Marital status was categorized as unmarried, married, or divorced/widowed. Occupation was grouped into six categories (e.g., students, freelance, business services, etc.). Average monthly income was classified into five levels: unemployed, ≤ 415 USD, 415–830 USD, 830–1660 USD, and > 1660 USD. Length of residence was categorized as local residents, 6 months–1 year, 1–5 years, or ≥ 5 years. Finally, perceived social support and discrimination experiences were assessed using the Perceived Social Support Scale and Discrimination Scale, both treated as continuous variables based on their mean scores.

Statistical methods

Descriptive statistical analysis

The dataset was randomly partitioned into training and validation sets at a ratio of 7:3. Continuous variables were expressed as mean ± standard deviation, while categorical variables were presented as frequencies (percentages). Differences in subject characteristics between the training and validation sets were assessed using Student’s t-tests and chi-square tests.

Calculate variable importance using the random forest algorithm

Random Forest (RF) is a widely used machine learning method in medical research for variable selection^24–26. In this study, the contribution of 28 candidate variables was evaluated using two established metrics: Mean Decrease Gini (MDG) and Mean Decrease Accuracy (MDA).

Construction of predictive model sets

Variables were sequentially entered into the predictive models based on the descending order of importance derived from the RF analysis. Consequently, two sets of models (based on MDG and MDA rankings) were trained using the training dataset. Each model was subsequently evaluated on the validation set. To further assess reliability and generalizability, a 5-fold cross-validation was conducted on the full dataset. Performance was quantified using the Area Under the Curve (AUC), and the optimal feature subset was identified by plotting AUC against the number of included variables.

Selection of the optimal predictive model

The better-performing set of predictive models was determined by comparing the initial performance, simple cross-validation performance, and 5-fold cross-validation performance of the two sets of models on the training data set, test data set, and entire data set. Subsequently, we compared the AUC of each predictive model within the superior group, taking into account the models’ applicability and practicality, to determine the optimal model for predicting MSM engaging in sexual behavior with females.

Visualization of the optimal model

The predictive model’s discrimination, calibration, and clinical utility were visualized by plotting the ROC curve, calibration curve, and decision curve analysis (DCA). Additionally, a nomogram for predicting MSM and female sexual behavior was created.The model development and validation process is shown in Fig. 1.

To minimize optimism arising from performing variable ranking and model building on the same dataset used for performance estimation, we implemented a 5 × 5 nested cross-validation (CV) framework. In the inner 5-fold CV loop, Random Forest–based mean decrease accuracy was used for variable importance ranking and sequential model construction. The outer 5-fold CV loop was reserved exclusively for unbiased performance evaluation of the final models.

Baseline model comparison

To quantify the incremental predictive value of SCT-based psychological and behavioral variables, a baseline demographic model (comprising age, education, marital status, and sexual orientation) was constructed using logistic regression. Performance comparisons between the baseline and full models were conducted on identical data splits using three complementary metrics: AUC, Area Under the Precision-Recall Curve (PR-AUC), and the Brier score.

Probability threshold determination

To facilitate practical implementation of the prediction model in real-world screening contexts, we performed systematic probability threshold selection based on two complementary strategies. First, the Youden index was applied to identify the optimal cutoff that maximized overall diagnostic accuracy. Second, an F1-score–based optimization strategy was used to balance precision and recall. For each candidate threshold, corresponding confusion matrices were constructed and F1-scores computed. The threshold yielding the maximal F1-score was selected as an alternative decision threshold for high-risk screening scenarios.

Subgroup performance analysis

Model robustness and generalizability were evaluated across key demographic strata: age (quartiles), education level, marital status, and migration status (dichotomized as local resident vs. domestic migrant). ROC curves and AUC values were calculated for each subgroup using predicted probabilities from the final model to assess the stability of discrimination across heterogeneous populations.

Model performance evaluation metrics

Discrimination was assessed via AUC. Given the imbalanced outcome distribution, PR-AUC was calculated as a complementary metric. Overall probabilistic accuracy was measured using the Brier score. Model calibration was evaluated by regressing observed outcomes on predicted logits; the calibration intercept and slope, along with 95% confidence intervals derived from 2,000 bootstrap resamples, were reported to assess bias and overfitting.

Results

Demographic characteristics

The study enrolled 2,403 MSM participants (mean age: 30.59 years). Most participants were well-educated (78.6% with college degrees or above) and predominantly engaged in freelance work, with 13.3% being students. Internal migrants constituted over 60% of the sample. The self-reported HIV prevalence was 8.3%. Regarding sexual behaviors, 16.9% reported male-to-female sexual encounters in the past six months, approximately 25% had more than three male anal sex partners, and 30% identified as versatile in sexual role preferences. Consistent condom use during male-to-male anal sex was reported by 58.1%. No statistically significant differences were found between the training and validation datasets, as shown in Table 1.Furthermore, reliability analysis confirmed the internal consistency of the applied scales. The GAD-7, PHQ-9, and Perceived Social Support Scale demonstrated excellent reliability, with Cronbach’s alpha coefficients of 0.954, 0.943, and 0.958, respectively. The Rosenberg Self-Esteem Scale and Discrimination Scale also exhibited good internal consistency, with coefficients of 0.872 and 0.873, respectively.

Table 1.

Sociodemographic, Psychological, and behavioral characteristics of MSM Participants.

Variables	Description	Overall (n = 2403)	Training set (n = 1682)	Validation set (n = 721)	p
Personal Factors
Age	Mean (SD)	30.59 (8.02)	30.72 (8.18)	30.27 (7.64)	0.2
Self-identified sexual orientation	Heterosexual	57 (2.4%)	39 (2.3%)	18 (2.5%)	0.654
	Homosexual	1887 (78.5%)	1326 (78.8%)	561 (77.8%)
	Bisexual	382 (15.9%)	268 (15.9%)	114 (15.8%)
	Not sure	77 (3.2%)	49 (2.9%)	28 (3.9%)
Most recent HIV test result	Positive	199 (8.3%)	144 (8.6%)	55 (7.6%)	0.707
	Negative	1996 (83.1%)	1395 (82.9%)	601 (83.4%)
	Unknown	208 (8.7%)	143 (8.5%)	65 (9.0%)
Diagnosed with an STI in the past six months	Positive	226 (9.4%)	157 (9.3%)	69 (9.6%)	0.916
	Negative	2177 (90.6%)	1525 (90.7%)	652 (90.4%)
Hypertension	Positive	155 (6.5%)	111 (6.6%)	44 (6.1%)	0.616
	Negative	2110 (87.8%)	1470 (87.4%)	640 (88.8%)
	Unknown	138 (5.7%)	101 (6.0%)	37 (5.1%)
Diabetes	Positive	69 (2.9%)	55 (3.3%)	14 (1.9%)	0.188
	Negative	2229 (92.8%)	1552 (92.3%)	677 (93.9%)
	Unknown	105 (4.4%)	75 (4.5%)	30 (4.2%)
Hyperlipidemia	Positive	222 (9.2%)	163 (9.7%)	59 (8.2%)	0.469
	Negative	2009 (83.6%)	1397 (83.1%)	612 (84.9%)
	Unknown	172 (7.2%)	122 (7.3%)	50 (6.9%)
GAD-7 score	Mean (SD)	11.74 (5.39)	11.72 (5.37)	11.80 (5.43)	0.742
PHQ-9 score	Mean (SD)	15.21 (6.37)	15.20 (6.32)	15.23 (6.51)	0.907
Rosenberg Self-Esteem Scale score	Mean (SD)	25.59 (3.43)	25.64 (3.42)	25.45 (3.43)	0.214
Personality Scale score	Mean (SD)	13.32 (2.35)	13.29 (2.33)	13.39 (2.40)	0.344
Behavioral Patterns
Smoking frequency in the past 30 days	Not smoking	1464 (60.9%)	1032 (61.4%)	432 (59.9%)	0.829
	Sometimes	392 (16.3%)	268 (15.9%)	124 (17.2%)
	Often	224 (9.3%)	154 (9.2%)	70 (9.7%)
	Everyday	323 (13.4%)	228 (13.6%)	95 (13.2%)
Alcohol consumption in past 30 days	Not drinking	1122 (46.7%)	796 (47.3%)	326 (45.2%)	0.603
	Sometimes	1067 (44.4%)	740 (44.0%)	327 (45.4%)
	Often	214 (8.9%)	146 (8.7%)	68 (9.4%)
Number of male anal sex partners in past 6 months	1	1206 (50.2%)	839 (49.9%)	367 (50.9%)	0.531
	2	598 (24.9%)	413 (24.6%)	185 (25.7%)
	3 or more	599 (24.9%)	430 (25.6%)	169 (23.4%)
Frequency of anal sex with men in past 6 months	More than 4 times a week	91 (3.8%)	63 (3.7%)	28 (3.9%)	0.437
	1–3 times a week	419 (17.4%)	307 (18.3%)	112 (15.5%)
	1–3 times a month	982 (40.9%)	677 (40.2%)	305 (42.3%)
	Less than 1 time a month	911 (37.9%)	635 (37.8%)	276 (38.3%)
Role in anal sex with men in past 6 months	Bottom	704 (29.3%)	476 (28.3%)	228 (31.6%)	0.209
	Top	921 (38.3%)	660 (39.2%)	261 (36.2%)
	Versatile	778 (32.4%)	546 (32.5%)	232 (32.2%)
Condom use in anal sex with men in past 6 months	Never	126 (5.2%)	94 (5.6%)	32 (4.4%)	0.379
	Seldom	248 (10.3%)	172 (10.2%)	76 (10.5%)
	Often	634 (26.4%)	455 (27.1%)	179 (24.8%)
	Every time	1395 (58.1%)	961 (57.1%)	434 (60.2%)
Sex with women in past 6 months	Yes	407 (16.9%)	289 (17.2%)	118 (16.4%)	0.668
	No	1996 (83.1%)	1393 (82.8%)	603 (83.6%)
Voluntary HIV testing recently	Yes	1946 (81.0%)	1368 (81.3%)	578 (80.2%)	0.542
	No	457 (19.0%)	314 (18.7%)	143 (19.8%)
Group sex with men recently	Yes	676 (28.1%)	476 (28.3%)	200 (27.7%)	0.818
	No	1727 (71.9%)	1206 (71.7%)	521 (72.3%)
Alcohol use before sex	Yes	893 (37.2%)	632 (37.6%)	261 (36.2%)	0.553
	No	1510 (62.8%)	1050 (62.4%)	460 (63.8%)
Drug use before sex	Yes	69 (2.9%)	48 (2.9%)	21 (2.9%)	0.998
	No	2334 (97.1%)	1634 (97.1%)	700 (97.1%)
Environmental Influences
Educational level	Junior high school and below	157 (6.5%)	116 (6.9%)	41 (5.7%)	0.535
	Senior high school	358 (14.9%)	248 (14.7%)	110 (15.3%)
	College and above	1888 (78.6%)	1318 (78.4%)	570 (79.1%)
Marital status	Unmarried	2035 (84.7%)	1424 (84.7%)	611 (84.7%)	0.859
	Married	255 (10.6%)	181 (10.8%)	74 (10.3%)
	Divorced or widowed	113 (4.7%)	77 (4.6%)	36 (5.0%)
Occupation	Cadre staff or career staff	245 (10.2%)	172 (10.2%)	73 (10.1%)	0.258
	Health care or teachers	141 (5.9%)	106 (6.3%)	35 (4.9%)
	Migrant workers or laborers	311 (12.9%)	225 (13.4%)	86 (11.9%)
	Students	320 (13.3%)	226 (13.4%)	94 (13.0%)
	Freelance	692 (28.8%)	461 (27.4%)	231 (32.0%)
	Business Services	463 (19.3%)	334 (19.9%)	129 (17.9%)
	Unemployed	231 (9.6%)	158 (9.4%)	73 (10.1%)
Average monthly income	415dollars and below	502 (20.9%)	346 (20.6%)	156 (21.6%)	0.512
	415–830 dollars	851 (35.4%)	610 (36.3%)	241 (33.4%)
	830–1660 dollars	802 (33.4%)	550 (32.7%)	252 (35.0%)
	Above 1660 dollars	248 (10.3%)	176 (10.5%)	72 (10.0%)
Length of residence	Local Residents	936 (39.0%)	652 (38.8%)	284 (39.4%)	0.96
	6 months to 1 year	318 (13.2%)	221 (13.1%)	97 (13.5%)
	1 to 5 years	552 (23.0%)	386 (22.9%)	166 (23.0%)
	5 years and above	597 (24.8%)	423 (25.1%)	174 (24.1%)
Perceived Social Support Scale score	Mean (SD)	57.66 (14.96)	57.57 (15.16)	57.88 (14.50)	0.645
Discrimination Scale score	Mean (SD)	10.02 (2.36)	10.01 (2.40)	10.05 (2.29)	0.722

Variable importance analysis

Random Forest analysis utilizing two distinct metrics - MDA and MDG - revealed differential patterns of predictor importance (Fig. 2). MDA analysis identified demographic and behavioral characteristics as primary predictors, with marital status and self-identified sexual orientation demonstrating the strongest influence, followed by substance use patterns and mental health indicators (PHQ-9 and GAD-7 scores). Conversely, MDG analysis highlighted self-identified sexual orientation and age as principal contributors, with psychosocial measures (Rosenberg Self-Esteem Scale and perceived social support scores) and marital status showing subsequent importance.

Fig. 2.

Variable Importance Rankings in the SCT-Based Random Forest Model for Predicting MSM–Female Sexual Engagement: (a) Mean Decrease Accuracy and (b) Mean Decrease Gini. In simple terms, this figure shows which psychological, behavioral, and demographic factors contribute most to predicting whether MSM engage in sexual behavior with women.

Cross-method comparison revealed consistent significance of certain predictors, particularly self-identified sexual orientation and marital status, spanning diverse domains including mental health, social support, sexual behaviors, and health status. The methodological contrast manifested in MDA’s identification of distinct high-impact variables versus MDG’s more uniform distribution of predictor importance, offering complementary perspectives on variable selection.

Predictive model development

The predictive modeling process employed sequential variable incorporation based on MDA and MDG importance rankings, resulting in 56 distinct models. Analysis of AUC trajectories (Fig. 3) demonstrated the comparative advantage of MDA-based models over MDG-based models in both training and 5-fold cross-validation phases. This performance superiority was most pronounced when incorporating fewer than 10 variables, highlighting MDA’s particular strength in identifying key predictive factors.

Fig. 3.

AUC Performance of the SCT-Based Random Forest Model for Predicting MSM–Female Sexual Engagement: Training Set and 5-Fold Cross-Validation Results. (c) shows the changes in AUC on the training set. The orange solid line (with circular markers) represents the MDA method, while the green dashed line (with triangular markers) represents the MDG method. (d) illustrates the changes in AUC during 5-fold cross-validation on the entire dataset, with the orange solid line representing the MDA method and the green dashed line representing the MDG method. The x-axis denotes the number of added variables, and the y-axis indicates the AUC values. This figure visually demonstrates how model performance changes as more predictors are added, highlighting that a compact set of predictors yields the most stable and accurate performance.

Further validation using the complete dataset revealed MDA’s distinct methodological advantages. The MDA approach maintained notably stable AUC trajectories throughout 5-fold cross-validation (Figure d), indicating robust variable importance assessment across different data partitions. Although MDG exhibited rapid initial gains in the training set (Figure c), MDA demonstrated superior generalization properties during cross-validation (Figure d), consistently maintaining strong AUC values. These systematic performance characteristics establish MDA as the optimal method for variable importance evaluation in this study, particularly in its capacity to identify predictors that enhance model stability and generalizability.

Model performance analysis

Comprehensive evaluation of AUC variations across the 56 predictive models revealed distinct performance patterns between MDA and MDG approaches (Fig. 3). Systematic comparison demonstrated MDA’s superior discriminative capacity in both training and cross-validation phases, particularly evident in models with streamlined variable sets (< 10 variables). This finding underscores MDA’s precision in identifying variables with strong predictive potential.

Cross-validation analysis of the complete dataset further emphasized the methodological strengths of the MDA approach. While MDG exhibited initial performance acceleration in the training phase (Figure c), MDA consistently demonstrated superior stability and generalization properties throughout the cross-validation process (Figure d). The maintenance of robust and elevated AUC values across validation iterations substantiates MDA’s effectiveness in identifying predictors that contribute fundamentally to model stability and accuracy. These results provide compelling evidence for MDA’s reliability in variable selection for predictive modeling.

To further quantify the incremental value of SCT-based psychological and behavioral predictors, we compared a demographic-only baseline model (including age, education, marital status, and sexual orientation) with the full SCT-guided model. The baseline model achieved an AUC of 0.756, PR-AUC of 0.462, and Brier score of 0.113. In contrast, the full model achieved an AUC of 0.802, PR-AUC of 0.526, and Brier score of 0.107, demonstrating clear improvements in discrimination and calibration beyond demographic factors alone.

Nested cross-validation (5 × 5) demonstrated robust generalization performance, with a mean outer-fold AUC of 0.809. Bootstrap optimism-correction estimated an optimism of 0.0065, yielding an optimism-adjusted AUC of 0.803. These findings indicate that the model maintains stable predictive performance after rigorous resampling-based bias correction. In addition to AUC-based discrimination, probabilistic calibration and imbalanced classification performance were further evaluated. The Brier score in the independent validation set was 0.1067, indicating good overall probabilistic accuracy. Calibration analysis showed a calibration intercept of − 0.202 (95% CI: − 0.527 to 0.150) and a calibration slope of 0.915 (95% CI: 0.758 to 1.092), suggesting no significant systematic over- or under-estimation of risk and near-ideal calibration. Given the relatively low outcome prevalence, Precision–Recall AUC (PR-AUC) was additionally assessed, yielding a value of 0.554 (95% CI: 0.505–0.605), corresponding to moderate discriminative performance under class imbalance.

Optimal model selection

Model performance analysis through cross-validation (Fig. 3) revealed a critical threshold in variable inclusion. Models incorporating more than 10 variables exhibited marked performance instability, characterized by substantial AUC fluctuations during cross-validation, indicating potential overfitting and information redundancy. This instability compromised predictive consistency across data partitions, limiting generalization potential.

The parsimonious 9-variable model emerged as the optimal configuration, demonstrating robust performance stability across validation procedures (AUC > 0.8). This model achieved superior discriminative capacity, with AUC values of 0.81 (95% CI: 0.78–0.84) and 0.80 (95% CI: 0.75–0.84) in training and validation sets, respectively²⁷. The concordance between training and validation ROC curves (Fig. 4e) further substantiates the model’s generalization capability.

Fig. 4.

Evaluation of the SCT-Based Predictive Model for MSM–Female Sexual Engagement: Discrimination, Clinical Utility, and Calibration Performance. (e) presents the ROC curves of the predictive model on the training and validation sets. The x-axis represents 1-specificity, and the y-axis represents sensitivity. The red curve represents the training set, the blue curve represents the validation set, and the dashed line represents the benchmark of random guessing. (f) shows the decision curve analysis of the predictive model at different thresholds. The x-axis represents the threshold probability, and the y-axis represents the net benefit. The red curve represents the training set, and the blue curve represents the validation set. (g) illustrates the calibration curves of the predictive model on the training and validation sets. The x-axis represents the predicted probability, and the y-axis represents the observed probability. The solid red line represents the original calibration curve of the training set; the dashed red line represents the calibration curve of the training set after 2000 bootstrap corrections; the solid blue line represents the original calibration curve of the validation set; the dashed blue line represents the calibration curve of the validation set after 2000 bootstrap corrections; and the black dashed line represents the ideal perfect calibration line (where predicted probabilities equal observed probabilities).This figure summarizes how well the final model distinguishes high-risk individuals, how clinically useful it is, and how closely predicted risks match observed outcomes.

Comprehensive validation through decision curve analysis demonstrated consistent clinical utility across probability thresholds. The alignment of net benefit curves between training and validation sets (Fig. 4f) indicates robust clinical applicability across diverse populations²⁸. Additionally, calibration analysis revealed excellent agreement between predicted and observed probabilities (Fig. 4g), particularly post-bias correction²⁹, validating the model’s predictive precision across different datasets. For practical implementation, an optimal probability threshold was determined using the Youden Index. The selected threshold was 0.148, yielding a sensitivity of 0.720, specificity of 0.769, positive predictive value (PPV) of 0.379, and negative predictive value (NPV) of 0.934. This threshold prioritizes high NPV and is suitable for behavioral risk screening scenarios requiring effective exclusion of low-risk individuals.

Clinical implementation through nomogram

A predictive nomogram was developed using the complete dataset to evaluate the likelihood of MSM engaging in male-to-female sexual behavior (Fig. 5). The nomogram incorporates nine key predictors: mental health indicators (GAD-7, PHQ-9), psychosocial measures (Rosenberg Self-Esteem Scale), behavioral factors (recent group sex, drug use before sex), and demographic characteristics (educational level, age, marital status, sexual orientation).

Fig. 5.

Nomogram of the SCT-based predictive model for MSM-Female sexual engagement. This nomogram provides a user-friendly tool that allows clinicians or outreach workers to estimate individual risk by simply summing points assigned to each predictor.

The scoring system assigns points (range: 0-450) based on individual variable contributions, with higher total scores indicating elevated risk. The linear predictor (− 5 to 5) transforms cumulative points into a standardized scale, which then converts to specific risk probabilities (0.1–0.9). This structured approach enables systematic risk assessment and prediction of male-to-female sexual behavior among MSM.

Risk factor analysis revealed several significant associations. Mental health indicators showed positive correlations with risk probability, with elevated GAD-7 and PHQ-9 scores associated with increased likelihood of male-to-female sexual behavior. Demographic and psychosocial factors demonstrated notable patterns: younger age, lower educational attainment, and decreased self-esteem scores correlated with higher risk profiles. Behavioral factors, including recent group sex participation and substance use before sexual activity, similarly indicated elevated risk. Marital status and sexual orientation emerged as significant predictors, with married individuals and those identifying as homosexual or bisexual showing increased probability of male-to-female sexual behavior. Subgroup performance analysis demonstrated stable discrimination across key demographic strata. The AUC ranged from 0.79 to 0.85 across age groups, 0.79 to 0.83 across educational levels, and 0.71 to 0.79 across marital status categories. Model discrimination remained robust among both domestic migrants (AUC = 0.830) and local residents (AUC = 0.814), indicating good transportability across population subgroups. Subgroup performance analysis demonstrated stable discrimination across key demographic strata. The AUC ranged from 0.79 to 0.85 across age groups, 0.79 to 0.83 across educational levels, and 0.71 to 0.79 across marital status categories. Model discrimination remained robust among both domestic migrants (AUC = 0.830) and local residents (AUC = 0.814), indicating good transportability across population subgroups.

Discussion

This study advances our understanding of male-to-female sexual behavior among MSM, a critical bridge population in HIV/STI transmission networks. Our analysis revealed that 16.9% of MSM participants reported sexual engagement with females within the past six months, demonstrating the prevalence of this behavioral pattern. By applying Random Forest algorithms within the SCT framework, we developed a robust predictive model (AUC: 0.81, 95% CI: 0.78–0.84) that identified nine key predictors across psychological, behavioral, and demographic dimensions. The model revealed that mental health indicators (GAD-7, PHQ-9), psychosocial factors (self-esteem), behavioral characteristics (group sex, substance use), and demographic variables (age, education, marital status, sexual orientation) collectively shape male-to-female sexual engagement patterns. From the perspective of SCT, these predictors represent interacting components of reciprocal determinism. Emotional distress reflected by elevated GAD-7 and PHQ-9 scores may impair self-regulatory capacity and increase reliance on emotion-driven coping strategies, thereby facilitating risk-compensatory behaviors. Reduced self-esteem may further weaken perceived self-efficacy and increase susceptibility to social conformity and validation-seeking through sexual relationships. Behavioral factors such as group sex participation and substance use before sex function as both outcome expectations and environmental cues that reinforce disinhibited decision-making. Together, these mechanisms illustrate how psychological vulnerability, behavioral reinforcement, and sociocultural constraints dynamically co-shape heterosexual engagement among MSM. These findings extend traditional behavioral research by highlighting psychological vulnerability as a crucial determinant, offering a more nuanced understanding of cross-gender sexual behavior. This comprehensive prediction framework provides evidence-based guidance for developing targeted interventions that address the unique health needs of MSMW within broader HIV/STI prevention strategies.

Analysis of performance trajectories revealed distinct patterns in the model’s predictive accuracy across variable selection methods (Fig. 3). The AUC curves exhibited characteristic phases: rapid initial improvement followed by plateau formation, indicating diminishing returns in predictive power with additional variables. This pattern highlights the disproportionate contribution of high-ranking variables, identified through importance measures, to overall model performance. The MDA method demonstrated superior discriminative capacity compared to MDG, particularly in parsimonious model configurations (Fig. 3d). This performance advantage was consistently observed across both training and validation phases, with MDA achieving notably higher AUC values in models with fewer variables. These findings establish MDA as the preferred variable importance measure for predicting this specific behavioral outcome. Through critical threshold analysis, we identified that models incorporating more than 10 variables exhibited performance instability during cross-validation, suggesting overfitting tendencies. Conversely, more parsimonious models maintained consistent performance between training and validation phases. The nine-variable configuration emerged as optimal, achieving maximum AUC values while preserving model stability and generalizability across validation scenarios.Beyond discrimination, the model also demonstrated strong overall probabilistic accuracy and calibration, with a Brier score of 0.1067, indicating low average prediction error. Calibration analysis yielded an intercept of − 0.202 and a slope of 0.915, both close to ideal values, suggesting minimal systematic over- or underestimation of risk. The PR-AUC of 0.554 further indicates moderate yet clinically meaningful performance under outcome imbalance. Subgroup analyses showed stable discrimination across age, education level, marital status, and migration status, supporting model robustness across heterogeneous MSM subpopulations. Importantly, compared with a demographics-only model, the SCT-based full model achieved clear incremental predictive value, confirming the added contribution of psychological and behavioral constructs. It should be emphasized that the present study is primarily predictive rather than causal in nature. All associations identified in the model reflect probabilistic relationships rather than unidirectional effects. Within the SCT framework, psychological, behavioral, and environmental factors are understood to influence each other dynamically through reciprocal and non-linear processes. Therefore, the observed associations should not be interpreted as deterministic causal pathways.

The predictive model reveals intricate interactions among psychological, social, and behavioral determinants shaping male-to-female sexual behavior in MSM populations. Mental health indicators emerged as primary predictors, with elevated anxiety and depression scores suggesting psychological distress may be associated with compensatory sexual behaviors and emotional seeking patterns^30,31. Age-related vulnerability among younger MSM aligns with developmental theories of sexual identity formation, reflecting heightened susceptibility to sociocultural influences during critical periods of sexual exploration³².

The model demonstrates significant pathways through which psychosocial and socioeconomic factors influence sexual behavior patterns. Lower self-esteem scores correlate with increased risk behaviors, potentially reflecting attempts to manage identity-related conflicts and seek social validation. Educational attainment demonstrates protective effects, highlighting the role of health literacy and socioeconomic stability in sexual decision-making processes. These findings underscore how psychological vulnerability intersects with social determinants to shape sexual risk patterns.

Analysis of behavioral and contextual factors illuminates key risk mechanisms within the SCT framework. Group sex participation indicates broader patterns of sexual exploration, while marital status reflects the impact of traditional social norms and familial expectations³³. Married MSM often navigate complex sexual behaviors driven by social discrimination and family obligations³⁴, illustrating the influence of environmental pressures on behavioral choices. Substance use emerges as a significant risk factor, potentially facilitating sexual risk behavior through diminished inhibition and maladaptive coping mechanisms³⁵, demonstrating the interplay between personal and behavioral factors within the theoretical framework.However, these findings should be interpreted within the context of NGO-based recruitment and self-reported behavioral data, which may influence representativeness and response accuracy. These issues are discussed in greater detail in the limitations section.

In practical application, risk scores derived from this model are intended to be used as confidential behavioral risk profiles rather than individual labels. Risk feedback should be delivered through community-based counseling and supportive interventions, with strict avoidance of stigmatizing language or surveillance-oriented framing. This approach aligns with harm-reduction principles and protects MSM-W from secondary social harm.

Limitations

This study has several important limitations. First, all behavioral, psychological, and sexual health variables were obtained through self-reported questionnaires, which may introduce recall bias and social desirability bias—particularly for sensitive behaviors such as substance use and sexual practices. Although we employed rigorous instrument development procedures, including expert consultation, pilot testing, and optimization of item wording to enhance clarity and reduce response burden, inherent limitations of self-reported data remain and may influence measurement precision.

Second, the use of NGO-based recruitment may have introduced selection bias. Individuals who are more socially connected or more engaged with community organizations may have been overrepresented in the sample, potentially limiting the representativeness of MSM populations who are less connected, more closeted, or reluctant to engage with community services. Furthermore, although our sampling covered six geographically diverse urban regions in China, generalizability to MSM in rural areas or regions without active NGO networks remains uncertain. Broader regional sampling and probability-based recruitment strategies are warranted in future studies.

Third, despite the implementation of rigorous internal validation procedures—including simple cross-validation, five-fold cross-validation, 5 × 5 nested cross-validation, and 2,000-sample bootstrap optimism correction—the model has not yet undergone external or temporal validation. Although internal validation demonstrated strong robustness and low optimism (0.0065), external validation remains essential to confirm model transportability across different settings, populations, and sociocultural environments. Ongoing multi-site data collection with our NGO partners will provide an independent dataset for future external validation.

Fourth, although the model incorporated 28 candidate predictors guided by Social Cognitive Theory, the final nine-variable model may not capture all relevant sociocultural or environmental determinants shaping MSM–women sexual behavior. Factors such as stigma-related stress, familial expectations, migration context, or broader social-structural influences may also contribute to behavioral decision-making but were not fully captured in our dataset. Future research should evaluate additional constructs, including community-level factors, stigma environments, and relationship dynamics, to improve model completeness and interpretability.

Finally, as with all cross-sectional predictive modeling studies, causal inferences cannot be drawn. While SCT provides a theoretical foundation for interpreting interactions among psychological, behavioral, and environmental factors, these associations should not be interpreted as unidirectional or deterministic pathways. Longitudinal designs are needed to further explore the dynamic and reciprocal mechanisms underlying heterosexual engagement among MSM.

Conclusions

Drawing from a multicenter survey of 2,403 MSM across six Chinese cities, this study developed and validated a theory-guided predictive model for male-to-female sexual behavior by integrating Social Cognitive Theory with machine-learning–based variable selection. The Mean Decrease Accuracy approach yielded a parsimonious nine-variable model with strong discriminative ability, which remained stable after 5 × 5 nested cross-validation. The model further demonstrated good probabilistic accuracy, acceptable calibration, and moderate PR-AUC suitable for imbalanced behavioral outcomes. Predictor contributions reflected SCT-aligned psychological (anxiety, depression, self-esteem), behavioral (group sex, substance use), and demographic (age, education, marital status, sexual orientation) mechanisms, illustrating how reciprocal psychological–environmental influences shape cross-gender sexual behaviors. Subgroup analyses confirmed stable performance across age, education, marital status, and migration groups, and baseline comparisons showed added predictive value beyond demographics alone.

The resulting nomogram provides a practical, theory-informed tool for early behavioral risk identification and has potential utility in targeted, non-stigmatizing HIV/STI prevention programs. Although the model demonstrated strong internal validity, generalizability beyond the six study regions requires external validation, which is currently underway through ongoing multi-site data collection. All analytic scripts have been documented, and the model’s outputs are intended to support community-based counseling rather than serve as individual risk labels, aligning with harm-reduction and stigma-free communication principles.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

S.L. and Y.G. conceptualized and designed the study, conducted data analysis, and drafted the first version of the manuscript. H.X., Q.G., and G.X. were responsible for data collection and investigation. J.C. contributed to literature review, validation, and visualization. J.Z. provided software support, data analysis assistance, and visualization. Y.W. and Y.C. supervised the study, managed the project, and revised the manuscript into its final version.

Funding

Development and Clinical Application of a Multimodal Early Warning Model for Neurological Syphilis and Associated Paralytic Dementia (AF7320007, YG2025ZD29). Public Health Research Project from Shanghai Disease Control and Prevention Administration (2024GKM30). Key discipline projects of Shanghai Three-Year Action Plan for Public Health under Grant (GWVI-11.1-29). National health commission of the PRC (Grant No. Z155080000004). Tongren Hospital Talent Launch Project (TR2023rc11, TR2024RC10). Changning District Medical and Health Research Special Project (CNKW2022Y08).The Research Fund of Shanghai Tongren Hospital, Shanghai Jiaotong University School of Medicine (TRYJ2024LC13,TRYJ2022LC10).Science and Technology Commission Shanghai Municipality for the Seroepidemiological Study of Novel Coronavirus Pneumonia in Key Populations (No. 20JC1410204) .

Data availability

The data presented in this study are available on request from the corresponding author. Due to privacy and ethical restrictions, the data are not publicly available to protect the participants’ privacy.

Declarations

Competing interests

The research was conducted in the absence of any commercial or financial relationships that could be construed as a potential Competing interests.

Ethical approval

This study was conducted in accordance with the guidelines of the Declaration of Helsinki and approved by the Ethics Committee of Shanghai University of Medicine & Health Sciences(Approval Number: 2023-MSMMPOX-22-310222197604080237).

Consent for publication

Not Applicable.

Informed consent

Statement: Informed consent was obtained from all subjects participating in the study.