Network and machine learning analysis of childhood trauma, mental health, and AI-based emotional support needs in adolescents from underdeveloped regions Authors: Jianqiang Wang, Shaoxiong Li, Jing Lv, Yanxin Zhang, Zheng Liu, Ziyan Wang, Yuhan Peng, Congcong Liu, Xuezhu Yu, Yuru Du, Youdong Li Published: 2026 ABSTRACT AbstractBackgroundThe current intervention efficacy of generative conversational artificial intelligence (GCAI) on overall mental health issues remains limited, which may be related to the complex and unclear relation between GCAI emotional support needs (GCAI-ESN) and childhood trauma and mental health issues.MethodsWe used the Childhood Trauma Questionnaire, the Mental Health Inventory of Middle-school students, and whether there were GCAI-ESN to assess 14,380 adolescents. Machine learning (ML) combined with SHapley Additive exPlanations (SHAP) analysis was employed to identify the key predictors influencing the GCAI-ESN model. Based on the identified predictors, an undirected network was constructed and a Bayesian network analysis was conducted.ResultsThe prevalence of childhood trauma, mental health issues, and GCAI-ESN among adolescents in underdeveloped regions was 32.43% (95% CI: 31.66%-33.19%), 16.78% (95% CI: 16.17%-17.39%), and 39.53% (95% CI: 38.73%–40.33%), respectively. The SHAP analysis of four machine learning models identified 15 key predictors for GCAI-ESN. In the undirected network model, easily anxious ‘MH6’ (EI = 1.07) and depressed ‘MH5’ (EI = 1.07) are the core nodes, while GCAI-ESN (BEI = 3.09) and family said hurtful things ‘CTQ14’ (BEI = 1.00) are the bridge nodes. Bayesian network analysis indicates that ‘CTQ14’ is the node with the most outgoing potential predictive relationships.ConclusionsThis study identified the key predictors of GCAI-ESN among adolescents in underdeveloped regions and determined ‘MH6’, ‘MH5’, and ‘CTQ14’ as potential intervention targets when the GCAI provides emotional support, offering concrete focal points for future research.Supplementary InformationThe online version contains supplementary material available at 10.1186/s40359-026-04924-5.Keywords: Generative Conversational AI, Emotional support needs, Mental health, Childhood trauma, Underdeveloped regions, Machine learning Introduction In recent years, mental health issues among adolescents have become increasingly prevalent and severe, with significant regional disparities. The prevalence of mental health conditions in the adolescent population is approximately 14% [1]. Rural adolescents exhibit a higher prevalence of mental health conditions compared to their urban counterparts and experience a higher incidence of childhood trauma [2–4]. A meta-analysis reveals that the treatment rate for any mental disorder in low and middle-income countries is only 6% [5], and rural adolescents have fewer opportunities to access mental health services compared to those in urban and suburban areas [6]. Adolescents in these underdeveloped regions often struggle to access effective treatment due to the absence of safety nets and economic constraints, perpetuating a cycle of poverty and mental health issues in these areas [7, 8]. Childhood trauma, mental health, and emotional support needs Research indicates that childhood trauma serves as a significant contributing factor to mental health issues among adolescents [9, 10]. Childhood trauma refers to adverse experiences occurring before the age of 18, including physical abuse, emotional abuse, sexual abuse, physical neglect, or emotional neglect [11]. Adolescents with a history of childhood trauma often lack adequate family, peer, or social support, are more vulnerable to negative experiences such as peer victimization, and are more likely to develop insecure attachment patterns [12], leading to heightened sensitivity to social evaluation [13]. Concurrently, among various types of childhood trauma, emotional abuse and emotional neglect demonstrate the strongest associations with depressive symptoms [14, 15]. Numerous studies indicate that emotional support serves as a protective factor against childhood trauma and mental health issues in adolescents [16–18]. According to Self-Determination Theory, humans possess an innate tendency toward self-determination. Individuals engage in actions because these actions fulfill their basic psychological needs, which are crucial for both short-term well-being and long-term development [19]. Adolescents with childhood trauma and mental health issues may have a heightened need for care and compassion from others to alleviate their distress or facilitate recovery from traumatic experiences. However, personal privacy concerns and limited public resources may restrict many adolescents’ opportunities to seek professional emotional support, leaving them with the need but without the access. Therefore, it is necessary to explore novel approaches to mitigate the interplay between childhood trauma and mental health issues by providing enhanced emotional support. Therefore, it is necessary to explore novel approaches to mitigate the interplay between childhood trauma and mental health issues by providing enhanced emotional support. The potential of AI-based emotional support Generative conversational artificial intelligence (GCAI) offers an emerging technological pathway to address the aforementioned issues. GCAI emotional support needs (GCAI-ESN) refers to an individual’s intrinsic motivation to seek supportive dialogue with GCAI to obtain emotional solace and coping strategies. With advancements in Large Language Models (LLMs), GCAI has acquired the capability to generate human-like responses, positioning it not merely as a tool but also as a virtual companion capable of providing emotional support and psychological companionship [20, 21]. A large-scale survey on motivations for using mental health conversational AI revealed that the expectation of receiving emotional support represents the most prominent user demand [22]. Existing research also indicates that users with higher levels of social isolation or mental health issues are more likely to be attracted to GCAI, form emotional connections with it, and exhibit stronger dependency [22, 23]. Among them, individuals with more severe social anxiety symptoms show significantly higher levels of trust and dependence on GenAI chatbots; such users tend to perceive GCAI as a predictable, controllable, and low-risk interactive environment, thereby reinforcing their emotional attachment [24]. To better provide emotional support to individuals with mental health issues, GCAI-based chatbots are being continuously optimized in areas such as empathy and compassionate response [25–27]. Thus, AI-based chatbots hold great potential in providing emotional support. However, some studies suggest that the overall effect of chatbots on improving mental health outcomes and loneliness is not significant [28–30], which may be related to the complex reasons why adolescents seek GCAI-ESN. Based on this, the present study, from the perspective of childhood trauma and mental health issues, aims to systematically analyze their interactive associations with GCAI-ESN using machine learning (ML) and network analysis. ML and network analysis Compared to traditional statistical methods, ML holds distinct advantages in addressing issues such as model overfitting, non-linearity, and complex prediction problems, earning recognition within psychological research [31]. The development of eXplainable Artificial Intelligence (XAI) has further propelled the application of ML, as it can elucidate the prediction mechanisms of complex ML models [32]. Combining ML with the SHapley Additive exPlanations (SHAP) interpreter from XAI enables the visualization of predictor importance values, facilitating the selection of key predictors through ranking [33]. Network analysis serves as a methodological tool to further reveal interactive associations among multiple variables. However, it is limited to examining correlations between nodes without inferring directionality. Therefore, we further incorporated Bayesian network analysis to explore potential directional relations among variables, presenting them in the form of a Directed Acyclic Graph (DAG). The current study This study employs ML to identify key predictors of GCAI-ESN, uses network analysis to examine their interconnections with childhood trauma and the mental health network structure, and applies DAG to identify potential directional associations. This integration holds significant theoretical value for informing the development of emotionally supportive chatbots. Methods Procedure Convenience sampling was employed to conduct an online survey via the Wenjuanxing platform from March to November 2024 in a less developed region of Hebei Province, China. Our study included middle and high school students who were currently in school, with an age range of 12 to 18 years old. A total of 19,166 questionnaires were collected. Based on an estimated reading speed of 400 characters per minute for Chinese junior high school students and the total number of items in the questionnaire [34], the average completion time for the electronic survey was calculated to be approximately 9 min. Questionnaires with completion times deemed too short (< 5 min) or too long (> 20 min) were excluded, as these were considered to reflect inattentive responding. Questionnaires exhibiting obvious response patterns (select one option for all questions) were also excluded. Consequently, a total of 14,380 valid electronic questionnaires were retained, yielding an effective response rate of 75.03%. In the survey, the investigators distributed electronic questionnaires to teachers at the schools within the survey sites, who subsequently passed them on to the potential participants. Each participant voluntarily engaged in the survey and received the survey results upon completion of the questionnaire. The survey procedure strictly adhered to the Declaration of Helsinki and was approved by the Ethics Committee of the authors’ institution (No. 20220933). Measurement Our study collected general information, Childhood Trauma Questionnaire-Short Form (CTQ-SF), Mental Health Inventory of Middle-school students (MMHI-60), and GCAI-ESN. General information included gender, age, whether an only-child, current residence, per capita monthly household income, and whether AI-related applications had been used. The Chinese version of the CTQ-SF, developed by Bernstein and revised by He [35, 36], was used to assess childhood trauma. The scale measures five dimensions, including emotional abuse, physical abuse, sexual abuse, emotional neglect, and physical neglect, with 5 items for each dimension. Using a Likert 5-point scale, a score of ≥ 13 for emotional abuse, ≥ 10 for physical abuse, ≥ 8 for sexual abuse, ≥ 15 for emotional neglect, or ≥ 10 for physical neglect indicates the presence of childhood trauma (Cronbach’s α = 0.72; Kaiser-Meyer-Olkin (KMO) = 0.85). The MMHI-60, developed by Wang, was used to assess mental health issues of adolescents [37]. The scale measures ten dimensions, including obsessive-compulsive symptoms, paranoia, hostility, interpersonal sensitivity, depression, anxiety, academic pressure, maladaptation, emotional instability, and psychological imbalance, with 6 items for each dimension. Using a Likert 5-point scale, a total average score of 2 to 2.99 indicates mild mental health issues, 3 to 3.99 indicates moderate issues, 4 to 4.99 indicates severe issues, and 5 indicates very serious issues (Cronbach’s α = 0.98; KMO = 0.99). Regarding GCAI-ESN, participants were asked, “Do you need to communicate your troubles or distress to obtain emotional support and advice from generative conversational artificial intelligence (also known as a chatbot)?” with the options “Yes” (value = 1) and “No” (value = 0). Statistical analysis Firstly, descriptive statistics was conducted using Positron. Count data were described using frequency and percentage, while continuous data were expressed by mean ± standard deviation (M ± SD). Secondly, given that the independent variables comprised general information and items from two scales, resulting in a numerous and complex feature set, we constructed four ML models using ensemble algorithms: LightGBM, CatBoost, XGBoost, and Gradient Boosting Regression (GBR). The SHAP analysis was then applied to generate summary plots to visualize the contribution of each independent variable to the prediction of the GCAI-ESN. To identify factors that were consistently important across different models, we determined the intersection of these top-20 lists, thereby defining a set of key predictors [33]. Furthermore, to interpret the influence of these key predictors, we generated partial dependence plots (PDPs) based on the model that achieved the highest Area Under the Curve (AUC). These plots visualize the marginal effect of each key predictor on the predicted GCAI-ESN. Thirdly, a partial correlation network model was constructed with the GCAI-ESN and its key predictors serving as nodes. The “qgraph” package was employed for network analysis, using Expected Influence (EI) as the metric to identify the core nodes in the network structure, with higher EI values indicating greater influence of that node in the network [38]. Bridge Expected Influence (BEI) was calculated using the “networktools” package, considering the top 10% of BEI values as bridge nodes in the network structure [39]. The “bootnet” package was used to perform 1000 Bootstraps for each node to test the stability of the network [40], and assessed using the Correlation Stability coefficient (CS-coefficient). The CS-coefficient should not be below 0.25, is preferably above 0.50, and above 0.70 is considered optimal. Fourthly, a Bayesian network was estimated using the “bnlearn” package and presented in the form of a DAG [41]. DAG is a graph used to represent potential directional relations among variables [42]. In the DAG, the strength and direction probability of the arcs are represented by ‘strength’ and ‘direction’, respectively. These parameters can take values in the range of 0 to 1. The greater the ‘strength’, the stronger its contribution to the network. A larger ‘direction’ indicates a clearer potential direction of association. Given the clear temporal precedence of childhood trauma over mental health issues, directional constraints were imposed in the model to allow paths exclusively from childhood trauma to mental health outcomes. Our study employed the bootstrap method to perform 10,000 resamples of the original dataset, and after each resampling, the hill-climbing algorithm was used to estimate the network structure [41]. Specifically, if the direction between two variables is consistent in more than 51% of the 10,000 resamples, this directionality is considered to be determined in the final DAG [43]. To investigate the stability of the estimated DAG, we conducted a bootstrap simulation with 1000 bootstrap samples. To ensure the generalizability of the network model, only edges with a strength greater than 0.5 and a direction certainty greater than 0.85 were retained [44]. Results Descriptive statistics There were 9,043 males (49.48%) and 9,050 females (50.52%), with a mean age of 14.71 years (SD = 1.5). Full demographic details are presented in Table 1. Of these participants, 16.78% (95% CI: 16.17%-17.39%) exhibited mental health issues of varying severity. Furthermore, 32.43% (95% CI: 31.66%-33.19%) reported experiencing childhood trauma. The prevalence rates of various dimensions of both childhood trauma and mental health issues are presented in Table S1 and S2. Notably, only 8.54% (95% CI: 8.10%–9.00%) had used AI-related applications, whereas 39.53% (95% CI: 38.73%–40.33%) expressed a need for GCAI-based emotional support. Table 1Baseline characteristics of participantsBaseline characteristicsN (%)Gender Male7115 (49.48)Age (M ± SD)14.71 ± 1.50Only-child Yes665 (3.7) No17,428 (96.3)Current residence Urban4502 (31.31) Rural9878 (68.69)Per capita monthly household income < 20004027 (28.00) 2000 ~ 50007918 (55.06) 5001 ~ 10,0002075 (14.43) > 10,000360 (2.50)Used AI applications Yes1228 (8.54) No13,152 (91.46) Baseline characteristics of participants The screening and validation of key predictors We first explored the key predictors that affect the GCAI-ESN. The dataset was partitioned into training, validation, and test sets using a 6:2:2 ratio. Figure 1A presents the AUC values for the test sets of the four ML models. The results indicate that the GBR model achieved the highest AUC score (AUC = 0.747), suggesting it possesses the strongest generalization performance. Furthermore, by identifying the common factors among the top 20 contributors from each of the four models, we found that 15 predictors were consistently identified as important across all models (Fig. 1B and F). These key predictors comprised three items related to depression; two items each for obsessive-compulsive symptoms, anxiety, emotional instability, and emotional abuse; and one item each for paranoia, interpersonal tension, psychological imbalance, and emotional neglect. Based on the analysis of the test set using the GBR model, PDPs were generated for these 15 key predictors (Fig. 2). With the exception of one item related to felt unimportant at home ‘CTQ5’, all other 14 predictors demonstrated a positive predictive effect on the GCAI-ESN score when their values exceeded 1. Additionally, the plots indicated that a moderate level of childhood trauma (a value of 2 or 3) was associated with an increased need for emotional support, whereas a severe level of childhood trauma (a value of 4 or 5) was linked to a decreased need for such support (Fig. 2G and I, and 2O). Identification of key predictors via SHAP analysis across four ML models on the test set. (A) AUC values of the four ML models. (B) Venny diagram illustrating the overlap among the top 20 predictors identified by the four models. (C-F) Summary plots for the top 20 predictors in the GBR, CatBoost, LightGBM, and XGBoost models PDPs for the 15 key predictors. (A) MH5: depressed; (B) MH2: mood swings; (C) MH13: tearful; (D) MH18: feels misunderstood and unsupported; (E) MH11: thoughts always differ; (F) MH7: motivation fluctuates; (G) CTQ3: subjected to verbal abuse; (H) MH3: double-check homework; (I) CTQ14: family said hurtful things; (J) MH16: self-blame; (K) MH48: persistent fear of failing exams; (L) MH6: easily anxious; (M) MH28: upset by others’ better grades; (N) MH34: unexplained fear; (O) CTQ5: felt unimportant at home Undirected network analysis Undirected network analysis was utilized to explore the core and bridge nodes within the network structure formed by GCAI-ESN and its 15 key predictors (Fig. 2A). The analysis revealed that the core nodes within this network structure are easily anxious ‘MH6’ (EI = 1.07) and depressed ‘MH5’ (EI = 1.07), while GCAI-ESN (BEI = 3.09) and family said hurtful things ‘CTQ14’ (BEI = 1.00) serving as the bridge nodes (Fig. 3B-3C). Furthermore, the edge between easily anxious ‘MH6’ and depressed ‘MH5’ exhibited the strongest weight (weight = 0.29). All other edge weights are presented in Table S3. Regarding the stability of the undirected network model, the narrow confidence intervals for the edge weights suggested high stability of the network structure (Figure S1). The CS-coefficients for both EI and BEI were 0.75 and 0.59, respectively, which also indicated high stability for both indices (Figure S2). Estimation of the undirected network model and identification of central nodes. (A) Undirected network structure. (B) Centrality index of EI for each node. (C) Bridge centrality index of BEI for each node Bayesian network analysis We proceeded to employ the GCAI-ESN and its 15 key predictors to estimate a Bayesian network, thereby elucidating the potential directional relations between nodes in the undirected network. The core nodes easily anxious ‘MH6’ and depressed ‘MH5’ in the undirected network model were influenced by subjected to verbal abuse ‘CTQ3’ (Strength = 0.94, Direction = 1.00) and family said hurtful things ‘CTQ14’ (Strength = 1.00, Direction = 1.00), respectively. Notably, we identified subjected to verbal abuse ‘CTQ3’, family said hurtful things ‘CTQ14’, mood swings ‘MH2’, feels misunderstood and unsupported ‘MH18’, tearful ‘MH13’, persistent fear of failing exams ‘MH48’, motivation fluctuates ‘MH7’ as direct predictors for GCAI-ESN (Fig. 4). All edge strengths and directional indicators for these arcs are detailed in Table S4. Bayesian network displayed as a DAG, with edge thickness corresponding to strength values. Discussion Our study found that the prevalence of childhood trauma, mental health issues, and GCAI-ESN among adolescents in underdeveloped regions was 32.43% (95% CI: 31.66%-33.19%), 16.78% (95% CI: 16.17%-17.39%), and 39.53% (95% CI: 38.73%–40.33%), respectively. The prevalence rate of mental health issues was close to that reported in prior studies [1, 45]. Using SHAP analysis on the test sets of four ML models, we identified 15 key predictors. These predictors were then used as nodes to conduct both undirected and directed network analyses, aiming to identify core nodes, bridge nodes, and their associations with other predictors. We found that the response rate for GCAI-ESN was 39.53%, yet only 8.54% of adolescents had prior exposure to GCAI-type applications. This indicates that GCAI remains underutilized among adolescents in underdeveloped regions. Research suggests that areas with lower economic levels often exhibit stronger stigma associated with mental illness, meaning adolescents with mental health issues may struggle to access effective support or treatment [46]. Concurrently, other studies indicate that compared to other large language models, GPT-4o demonstrates superior emotional understanding and empathy. AI counselors designed based on such capabilities can effectively alleviate individual depression, anxiety, and loneliness in the short term [47]. Given its high accessibility and capacity to provide emotional support, GCAI could serve as a crucial tool for bridging the gap between demand and access, suggesting its potential for broad application among adolescents in underdeveloped regions. Among the four ML models, 15 predictors demonstrated key predictive effects on GCAI-ESN. Items from the domains of depression, obsessive-compulsive symptoms, anxiety, paranoia, interpersonal sensitivity, and psychological imbalance all exhibited positive predictive contributions to GCAI-ESN when their scores exceeded 1. However, this pattern was not observed for the three predictors from the childhood trauma. Notably, for ‘CTQ5’, its positive predictive contribution to GCAI-ESN declined at the highest score point. This suggests that mental health issues may be the direct catalyst for adolescents’ emotional support needs, while they might have difficulty recognizing the impact of childhood trauma, particularly among those with more severe trauma. Drawing on classical psychoanalysis theory, fundamental childhood experiences, internal conflicts, and psychological trauma are considered etiological factors for adult psychosomatic diseases and even mental disorders [48]. This implies that when adolescents seek consultation or emotional support from GCAI, assistance should extend beyond addressing immediate mental health concerns. Adopting an integrative perspective to help them understand both the presenting issues and their origins may yield more enduring positive effects. In the undirected network model constructed from GCAI-ESN and its 15 key predictors, ‘MH6’ and ‘MH5’ emerged as core nodes, which aligns with findings from prior research [49]. This indicates that symptoms of anxiety and depression represent prominent and significant manifestations when childhood trauma, mental health issues, and GCAI-ESN co-occur, playing a crucial role in connecting other nodes within the network model. Furthermore, the edge weight between these two nodes was the strongest, indicating the closest connection. This suggests that they represent potential intervention targets for alleviating the connections among nodes within the network. Additionally, GCAI-ESN and ‘CTQ14’ functioned as bridge nodes, indicating their important bridging roles in connecting the three clusters. However, GCAI-ESN existed as a single node without forming a distinct cluster with other nodes, which may lead to an overestimation of its bridging effect. Therefore, ‘CTQ14’ should be considered the key bridge node linking childhood trauma and mental health issues, and regarded as a potential intervention target to mitigate the risk of childhood trauma contributing to mental health issues. In the DAG based on the Bayesian network, ‘MH6’ and ‘MH5’ were directly predicted by ‘CTQ3’ and ‘CTQ14’, respectively, with ‘CTQ14’ also being the node emitting the most potential predictive connections. Both ‘CTQ3’ and ‘CTQ14’ are items reflecting emotional abuse, describing experiences such as being sworn at, insulted, or told hurtful things by family members. Adolescents may interpret these as forms of rejection or disapproval from their families. Helping adolescents recognize that such remarks may not stem from malice but rather from familial motivation or long‑established parenting patterns could be effective in mitigating the impact of childhood trauma. This aligns with the core premise of cognitive‑behavioral therapy: it is not the event itself but the individual’s interpretation and perception of the event that leads to anxiety, depression, and other distressing emotions. Research has shown that internet-based cognitive behavioral therapy (CBT) can effectively reduce symptoms of depression and anxiety in individuals [50]. This suggests that while providing emotionally supportive interactions grounded in psychotherapeutic principles, GCAI should also aim to help mitigate the adverse effects of childhood trauma. In summary, this study conducted a large-scale survey on childhood trauma, mental health issues, and GCAI-ESN among adolescents in underdeveloped regions. By employing machine learning and two network analysis methods, we preliminarily identified ‘MH6’, ‘MH5’, and ‘CTQ14’ as potential intervention targets when GCAI provides emotional support. Future research could develop specialized, evidence-based intervention modules targeting these nodes, such as CBT-based emotion regulation or anxiety management exercises. Building upon this, chatbots should integrate psychoeducational content to help adolescents understand the link between traumatic experiences and current emotional distress, while also providing a safe space for disclosure and cognitive restructuring. This approach would enable more precise support for adolescents with a history of childhood trauma and mental health issues. However, our study has certain limitations: First, it only surveyed students who were currently enrolled in school, excluding those who exhibited severe school refusal behaviors or were on leave from school. This may have led to an underestimation of the prevalence of mental health issues, as their mental health issues tend to be more pronounced. Second, although the sample size of our study is relatively large, the survey method was not a nationally random sampling, and thus the generalizability of the results requires further validation. Third, the survey instrument used in our study was the self-rating questionnaire, which is inherently subjective and may introduce reporting bias. Moreover, GCAI-ESN, which is assessed with a single binary item, captures a subjectively perceived or anticipated need for emotional support rather than actual AI usage. Future research should develop an item pool for adolescent GCAI-ESN through qualitative interviews and construct a multidimensional scale for comprehensive assessment. Fourthly, although our study included some demographic characteristics as covariates, there were still unmeasured confounding variables that may affect adolescent mental health, such as parental mental health and school environment. Fifth, although DAG can infer potential directional associations based on cross-sectional data, the lack of long-term longitudinal follow-up in this study makes it difficult to establish clear causal relationships between variables. Furthermore, the substantial gap between reported need (39.53%) and actual usage (8.54%) warrants further investigation. Sixth, although we employed four machine learning algorithms, the highest AUC value was only 0.747. This suggests that the independent variables have limited explanatory power for GCAI-ESN and indicates that its antecedents are likely more complex, necessitating further exploration in future research. Seventh, while we have proposed potential intervention targets, we have not tested their effectiveness; future experimental studies are required for validation. Conclusions Using four ML algorithms, we identified 15 key predictors of GCAI-ESN, including ‘MH5’, ‘MH2’, ‘MH13’, ‘MH18’, ‘MH11’, ‘MH7’, ‘CTQ3’, ‘MH3’, ‘CTQ14’, ‘MH16’, ‘MH48’, ‘MH6’, ‘MH28’, ‘MH34’, and ‘CTQ5’. By treating these predictors as nodes, undirected and directed network models were constructed. Based on the analysis of central and bridge nodes, as well as DAG, ‘MH6’, ‘MH5’, and ‘CTQ14’ were preliminarily identified as potential intervention targets when GCAI provides emotional support. Supplementary Information Supplementary Material 1. Acknowledgements We express our profound appreciation to Professor Alf Gerlach and Ulrich Sollmann from Germany for their expert guidance during our research. Authors’ contributions JqW and SxL managed, analyzed, interpreted, and annotated the data, designed the study, and drafted the manuscript. JL and YxZ collected and managed the data and drafted the manuscript. ZL, ZyW, YhP, CcL, and XzY collected the data and drafted the manuscript. YrD and YdL interpreted the data, designed the study, drafted the manuscript, supervised the project, and provided resources. All authors reviewed and edited the manuscript and critically revised it for important intellectual content. Funding The study was supported by Hebei Province Social Science Fund (HB25SH017). Data availability Data: The data of this study are available from the corresponding author upon reasonable request.Code: All original code for this study is available in the GitHub public repository at https://github.com/JiangqiangW/ML-and-network-analysis.git. Declarations Ethics approval and consent to participate Our survey was conducted anonymously, with the questionnaire collecting no private information such as names or identification numbers. Our study strictly adhered to the Helsinki Declaration and was reviewed and approved by the Medical Ethics Committee of the First Hospital of Hebei Medical University (No. 20220933), with a waiver for written informed consent granted, and oral informed consent was obtained from students prior to the survey. No manuscript content was generated by artificial intelligence (AI), but AI tools were used for grammar checking to improve readability. Consent for publication Not applicable Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Jianqiang Wang and Shaoxiong Li contributed equally to this work and share first authorship. Contributor Information Yuru Du, Email: 59003890@hebmu.edu.cn. Youdong Li, Email: 1026087958@qq.com. References Associated Data Supplementary Materials Supplementary Material 1. Data Availability Statement Data: The data of this study are available from the corresponding author upon reasonable request.Code: All original code for this study is available in the GitHub public repository at https://github.com/JiangqiangW/ML-and-network-analysis.git.