Frontiers in Psychology：AI 支持的发展性阅读疗法混合方法可行性研究

Abstract

Introduction:

Multimodal data processing is increasingly important for understanding human affective experience in naturalistic settings. In this study, “multimodal” refers to the integration of heterogeneous human-centered evidence streams rather than synchronized perceptual or physiological sensing. We evaluated a low-threshold university-library support model integrating developmental bibliotherapy salons with AI chatbot-supported emotional expression.

Methods:

A convergent mixed-methods feasibility and acceptability design was conducted at a Chinese university. Of 180 baseline respondents, 45 voluntarily registered for a three-week intervention and 38 completed all sessions and post-intervention assessments; 68 non-participating students with complete baseline and follow-up assessments served as a descriptive comparison group. Evidence sources included psychological self-report data, participation records, AI-use feedback, participant evaluations, and semi-structured interviews with nine intervention participants. Quantitative indicators assessed recruitment, retention, acceptability, and exploratory SCL-90 change, while thematic analysis examined participants' affective experiences and perceived mechanisms.

Results:

All available intervention places were filled within two days, retention reached 84.4%, and 94.7% of completers reported moderate or considerable benefit. The intervention-group mean SCL-90 change was −0.12 points, with an approximate 95% confidence interval of [−0.23, −0.01]. Because allocation was voluntary and non-randomized, no causal group-by-time intervention effect is claimed. Qualitative findings suggested that bibliotherapy salons supported interpersonal resonance, cognitive reframing, and shared reflection, whereas the AI chatbot provided a private and non-judgmental channel for emotional disclosure.

Discussion:

AI-supported developmental bibliotherapy appears feasible and acceptable as a low-threshold university-library wellbeing service. The findings support feasibility, acceptability, and perceived mechanisms only; they do not establish treatment efficacy, clinical effectiveness, or a neuroscience mechanism. More broadly, the study illustrates how heterogeneous psychological, behavioral, and experiential evidence can be integrated in a naturalistic educational setting.

1 Introduction

University students are increasingly exposed to academic pressure, social uncertainty, career anxiety, and identity-related developmental challenges, making student wellbeing a central issue in higher education and campus service innovation (; Storrie et al., 2010). Psychological distress among university students has been associated with poorer academic functioning, reduced quality of life, and increased demand for campus-based support services (). Although university counseling services remain essential, many students still encounter barriers such as stigma, limited service capacity, long waiting times, and reluctance to disclose personal distress in formal clinical settings (; ). Recent studies suggest that digital mental health interventions and AI-driven conversational agents can provide scalable, low-threshold, and timely support for young people and university students, particularly when conventional services are difficult to access (). At the same time, neuroscience and perception science increasingly emphasize that emotional experience, wellbeing, and help-seeking behavior should be understood through multimodal signals, self-reported psychological states, user interaction behaviors, and contextual human experiences (). This perspective creates an important opportunity to rethink university libraries not only as knowledge-service spaces, but also as accessible environments for multimodal, human-centered wellbeing support.

Existing approaches to student wellbeing have mainly followed three directions: professional counseling, digital mental health interventions, and AI-supported conversational tools. Professional counseling offers strong human support but is often constrained by limited resources and stigma-related barriers (). Digital interventions improve scalability and can extend support beyond traditional service hours, but they may suffer from low engagement, limited personalization, and weak contextual embedding in students' everyday campus life (; ). AI chatbots provide anonymity, accessibility, and continuous interaction, making them promising tools for emotional disclosure and low-threshold support (). However, recent studies also report concerns about methodological heterogeneity, over-reliance on self-report, limited emotional warmth, privacy concerns, and uncertain long-term effects (; ). Evidence from university student populations further suggests that AI chatbot use should be interpreted carefully in relation to mental health outcomes, user dependence, and contextual service design (Zhang et al., 2025b). Therefore, a key research gap remains: how to design a feasible, acceptable, and ethically appropriate support model that integrates human-guided reading, AI-supported emotional expression, and mixed-methods data interpretation within a real university service environment.

To address this gap, this study proposes a multimodal AI-supported developmental bibliotherapy model situated in a university library. The proposed model combines face-to-face developmental bibliotherapy salons with AI chatbot-assisted emotional expression, thereby linking human interaction, reading-based reflection, AI-mediated dialogue, psychological self-report, and qualitative experiential data within a single feasibility framework. Rather than treating AI as a replacement for human support, the model positions AI as a supplementary channel that may lower the threshold for emotional disclosure, while bibliotherapy salons provide social resonance, shared interpretation, and guided reflection. Methodologically, this study adopts a convergent mixed-methods feasibility and acceptability design, integrating recruitment and retention indicators, participant evaluations, exploratory SCL-90 changes, AI-use feedback, and semi-structured interview data. This design responds to recent calls for application-oriented multimodal affective systems that move beyond algorithmic performance alone and examine user experience, contextual validity, and real-world implementation (; ). It also aligns with recent multimodal affective computing studies that emphasize cross-modal fusion, weakly correlated data, and human-centered interpretation under complex real-world conditions (; ).

A conceptual boundary is necessary. Questionnaires, participation records, self-reported AI use, evaluations, and interviews are not equivalent to synchronized perceptual or physiological modalities. They are heterogeneous data sources generated at different temporal resolutions and with different measurement assumptions. The present study therefore uses “multimodal evidence integration” as a pragmatic umbrella term, while treating “mixed-methods” and “heterogeneous evidence” as the more precise methodological descriptors. This distinction prevents the service evaluation from being confused with multimodal emotion recognition, computational sensor fusion, or neuroscience experimentation.

The main innovations and contributions of this study are summarized as follows:

• This study develops a library-based multimodal AI-supported developmental bibliotherapy model that integrates reading salons and AI chatbot interaction as complementary forms of low-threshold student wellbeing support. Unlike AI-only or reading-only interventions, the proposed service design combines a human-guided group component with a private digital reflection channel; this comparison is conceptual rather than the result of experimentally separated intervention arms.

• This study provides empirical evidence from a real university setting, showing that quantitative psychological indicators, intervention implementation records, participant evaluations, AI-use feedback, and qualitative experiential data can be jointly interpreted without overstating causal effectiveness.

• This study extends the application scope of multimodal data processing in neuroscience and perception science by demonstrating how affective, cognitive, behavioral, and contextual information can be organized within a human-centered library service model.

The remainder of this paper is organized as follows. Section 2 reviews developmental bibliotherapy, AI chatbot-based mental health support, and multimodal affective data processing. Section 3 describes the mixed-methods feasibility design, intervention procedure, measurement instruments, and data analysis strategy. Section 4 introduces the experimental environment, datasets, baselines, and evaluation protocol. Section 5 presents the empirical findings and supplementary analyses. Section 6 concludes the study and outlines future directions.

2 Related work

2.1 Developmental bibliotherapy and library-based wellbeing support

University students' psychological wellbeing has increasingly become a central concern in higher education, especially as students face academic pressure, interpersonal uncertainty, loneliness, career-related anxiety, and identity development challenges. In this context, university libraries have gradually been reconsidered as supportive campus spaces rather than merely repositories of academic resources (; ). Their openness, familiarity, voluntary accessibility, and non-clinical atmosphere make them suitable for providing preventive and developmental wellbeing services that may reach students who are reluctant to seek formal counseling (). From the perspective of campus space theory, libraries can also function as “third places” where students experience belonging, safety, and informal social connection outside classrooms and clinical service settings.

Developmental bibliotherapy provides a theoretically relevant pathway for library-based wellbeing support. Unlike clinical bibliotherapy, which is often delivered in therapeutic contexts, developmental bibliotherapy emphasizes reading, guided reflection, emotional identification, and discussion to help individuals respond to ordinary developmental challenges. Early bibliotherapy theory suggests that reading-based activities may support self-understanding through processes such as identification with textual characters, emotional catharsis, and insight generation. Later work further highlights the role of reader advisory and structured discussion in transforming reading from a passive activity into a reflective and supportive process (Swain et al., 2014).

Recent empirical studies also indicate that reading-based interventions can contribute to emotional regulation, stress management, communication skills, and psychological wellbeing. For example, bibliotherapy has been reported to reduce examination-related stress among college students when delivered in a structured intervention format (). Reading-centered interventions have also been associated with improvements in listening, communication, and reflective interpersonal understanding (). More recent evidence from university library settings suggests that reading therapy may have positive effects on college students' mental health, especially when the intervention is organized through accessible institutional resources (). Similarly, library-led bibliotherapy for stress management has shown practical feasibility in educational environments, supporting the value of librarians and campus professionals as non-clinical facilitators ().

However, most existing studies still treat bibliotherapy as an independent reading intervention. Although these studies provide important evidence for the value of reading-based wellbeing support, they rarely examine how bibliotherapy can be combined with AI-supported emotional expression, participant evaluation, and multimodal data interpretation. In particular, little is known about how university libraries can integrate human-guided reading salons and digital interaction tools into a single low-threshold support model. This gap motivates the present study, which situates developmental bibliotherapy within a broader multimodal AI-supported framework rather than treating reading, discussion, and digital support as separate services.

2.2 AI chatbots and digital mental health support for students

AI-based conversational agents have become an important direction in digital mental health research because they can provide accessible, scalable, and low-pressure interaction for users who may hesitate to seek face-to-face support. A recent systematic review and meta-analysis found that AI-based conversational agents have potential for promoting mental health and wellbeing, but their effects vary according to intervention design, user engagement, target population, and evaluation method. This suggests that AI chatbots should not be evaluated only as technical systems, but also as situated support tools whose value depends on how they are embedded in real service contexts.

Recent studies further show that AI chatbots may help alleviate psychological distress and support health-related behaviors among adolescents and young adults (). At the same time, generative AI mental health chatbots have raised new questions regarding response quality, safety, emotional appropriateness, and the boundary between supportive conversation and therapeutic intervention (Zhang et al., 2025a). These concerns are especially important in student populations, where users may engage with AI tools for emotional disclosure, self-reflection, companionship, or practical advice without clearly distinguishing between informal support and professional mental health care.

In college and university settings, recent studies have begun to examine chatbot feasibility, acceptability, and usage behavior more directly. Evidence from college student populations suggests that AI chatbots may provide a useful low-threshold support option, but existing studies remain heterogeneous in design, sample characteristics, outcome measures, and follow-up duration (). An open trial of a generative AI-powered mental wellness chatbot also indicates that students may accept AI-based support when it is convenient and responsive, although user trust, safety monitoring, and sustained engagement remain important concerns (). In addition, research on Chinese university students suggests that AI chatbot usage behavior is associated with mental health-related outcomes, highlighting the need to interpret chatbot engagement in relation to users' psychological states, technology habits, and support needs (Zhang et al., 2025b).

Taken together, these studies indicate that AI chatbots are promising but should not be understood as substitutes for human support. AI tools can provide anonymity, immediacy, and a non-judgmental channel for emotional expression, but they may lack relational warmth, contextual sensitivity, and deep interpersonal understanding. Therefore, the present study integrates AI chatbot interaction as a supplementary channel embedded within a human-guided developmental bibliotherapy program. This design preserves the accessibility of AI-supported emotional disclosure while retaining the social resonance, shared reflection, and interpretive depth of face-to-face reading salons.

2.3 Multimodal data processing for affective perception and human-centered support

Multimodal data processing has become a key research direction in affective computing, neuroscience, and perception science because human emotion and wellbeing are expressed through heterogeneous channels, including language, voice, facial behavior, physiological signals, self-reports, interaction behaviors, and contextual experiences. Recent surveys on multimodal emotion recognition have summarized advances in feature representation, cross-modal fusion, robustness, and application scenarios, while also identifying persistent challenges such as modality imbalance, missing modalities, privacy concerns, and limited ecological validity (). These challenges are directly relevant to real-world wellbeing-support studies, where data are often heterogeneous, partially observed, and embedded in complex social environments.

Recent multimodal affective computing studies have proposed different strategies to improve robustness under noisy, weakly correlated, or incomplete modality conditions. Transformer-based feature restoration has been used to enhance multimodal sentiment analysis when modality information is degraded or partially missing (Sun et al., 2023). Momentum distillation has also been introduced to improve multimodal learning by transferring complementary knowledge across tasks and modalities (). In addition, weakly correlated multimodal sentiment analysis has highlighted the difficulty of aligning heterogeneous information sources when different modalities do not provide equally strong or consistent affective signals ().

Other studies have focused on modality-invariant representation, cross-modal transfer, and large-language-model-enhanced fusion. Contrastive modality-invariant learning has been proposed to reduce the negative influence of missing modalities in multimodal emotion recognition (). Cross-modal dynamic transfer learning further explores how information from one modality can support another when affective cues are incomplete or unevenly distributed (). More recently, large-language-model-enhanced multimodal fusion has been investigated as a way to improve emotion interpretation by combining feature-level fusion with language-based reasoning ().

Uncertainty and missing-modality problems are also central to multimodal perception-oriented research. Uncertainty-aware lightweight multimodal emotion recognition has emphasized the importance of reliable prediction under ambiguous affective signals (). Multimodal physiological–behavioral integration, such as the combination of EEG and facial expressions, further demonstrates the value of integrating neurophysiological and perceptual information for emotion classification (). In addition, text-guided reconstruction methods for uncertain missing modalities show how one modality can help recover or compensate for incomplete information from another modality (Shi et al., 2025).

Despite these advances, many multimodal studies remain concentrated on algorithmic prediction, benchmark performance, and modality-level optimization. Comparatively fewer studies examine how multimodal thinking can inform human-centered wellbeing interventions in naturalistic educational environments. The present study does not aim to develop a new emotion recognition algorithm; instead, it extends the logic of multimodal data processing into a university library intervention by integrating reading-based human interaction, AI-supported emotional expression, psychological self-report, participant evaluation, and qualitative interview data. In this way, the study provides an applied perspective on how heterogeneous affective, behavioral, and experiential information can be jointly interpreted to understand student wellbeing and affective perception in a real-world support context.

For terminological precision, the manuscript distinguishes three levels of evidence: (1) conventional multimodal perceptual signals, such as speech, facial behavior, physiology, or neuroimaging; (2) digital interaction traces, such as timestamped usage logs or raw chatbot dialogue; and (3) heterogeneous service-evaluation evidence, such as questionnaires, attendance, evaluations, and interviews. The current study operates primarily at the third level and includes only self-reported summaries of the second level. Consequently, claims are limited to feasibility, acceptability, and perceived mechanisms in a naturalistic educational setting.

3 Method

3.1 Study design and multimodal feasibility framework

This study adopted a convergent mixed-methods feasibility and acceptability design to evaluate a multimodal AI-supported developmental bibliotherapy model implemented in a university library. The study was not designed as a randomized clinical efficacy trial. Instead, it aimed to examine whether the proposed model could be recruited, delivered, completed, accepted, and meaningfully experienced in a real-world educational environment. This design choice is consistent with the early-stage nature of the intervention, where the central methodological question is not whether the model produces definitive causal effects, but whether it is feasible and acceptable enough to justify more rigorous future evaluation.

3.1.1 Operational definition, primary aim, and evidentiary boundary

The primary aim was to estimate implementation feasibility and participant acceptability. Exploratory symptom change was a secondary, hypothesis-generating outcome. The term “multimodal data” refers here to the coordinated interpretation of structurally different evidence sources, not to synchronized sensor channels. No latent multimodal representation was learned, no modality-level features were fused, and no prediction model was trained. The unit of integration was the study-level inference: recruitment and retention described implementation, SCL-90 scores described self-reported symptom burden, participant evaluations described acceptability, and interviews contextualized perceived mechanisms and limitations.

The corresponding hierarchy of inference was prespecified as follows. Feasibility indicators could support claims about whether the program was deliverable; acceptability indicators could support claims about perceived usefulness and burden; within-group SCL-90 change could describe temporal change among completers; and interview themes could describe participants' reported experiences. None of these components, alone or in combination, could establish that the intervention caused psychological improvement. This hierarchy guided the revised wording throughout the Results, Discussion, and Conclusion sections.

The methodological innovation of this study lies in integrating heterogeneous evidence sources into a unified feasibility-oriented framework. These sources include implementation indicators, psychological self-report, AI-use feedback, participant evaluation, and qualitative narratives. For each participant i, the complete multimodal evidence space is defined as

where represents bibliotherapy-salon participation data, denotes AI chatbot interaction feedback, indicates psychological self-report data, represents qualitative interview data, and denotes evaluation and acceptability data. This formulation reflects the core assumption of the present study: student affective experience and wellbeing-related responses should be interpreted through multiple complementary data sources rather than a single outcome measure.

The mixed-methods structure followed a convergent design. Quantitative and qualitative data were collected during the same intervention period, analyzed separately, and integrated during the interpretation stage. This structure is expressed as

where denotes the quantitative dataset, denotes the qualitative dataset, and ⊕ indicates interpretive integration rather than statistical pooling. This distinction is important because the study did not treat interview narratives as numerical labels, nor did it reduce participant experience to questionnaire scores. Instead, the two evidence streams were used to explain and complement each other.

Feasibility was conceptualized as a multidimensional construct involving recruitment, enrollment, completion, implementation stability, and acceptability. Formally, feasibility was defined as

where R denotes recruitment feasibility, E denotes enrollment responsiveness, C denotes completion and retention, S denotes session implementation stability, and A denotes acceptability. This definition allowed the study to evaluate the proposed model as a service process rather than only as an outcome-generating intervention.

Exploratory psychological change was calculated using pre- and post-intervention psychological scores. For participant i, the individual change score was defined as

A negative value of ΔP_i indicates a reduction in self-reported psychological symptoms. However, because participants self-selected into the intervention rather than being randomly assigned, this change score was interpreted as exploratory evidence rather than causal evidence of intervention effectiveness.

To strengthen interpretive rigor, quantitative findings, participant evaluations, and qualitative themes were triangulated. For a given finding k, the triangulated interpretation was represented as

where Q_k denotes quantitative evidence, V_k denotes participant evaluation evidence, and I_k denotes qualitative interview evidence. A finding was considered more credible when these three sources converged in a theoretically coherent direction. This triangulation strategy is particularly suitable for feasibility research because implementation success, acceptability, and perceived mechanisms cannot be fully captured by a single statistical indicator.

3.2 Participants, recruitment, and intervention protocol

Participants were recruited through university announcements, the library's official social media account, and student online communities. These recruitment channels were selected because they were already embedded in students' everyday campus information environment and were consistent with the low-threshold positioning of the intervention. A total of 180 students completed the baseline survey. After receiving information about the intervention, 45 students voluntarily registered for the three-week program. Among them, 38 completed all three developmental bibliotherapy salons and both pre- and post-intervention assessments, forming the intervention group. In addition, 68 students who did not complete the intervention but finished both baseline and follow-up questionnaires were included as a descriptive comparison group.

Participant flow was reconstructed explicitly. Of the 180 baseline respondents, 45 registered for the intervention and 135 did not register. Among the 45 registrants, 38 completed the intervention and post-test, whereas seven did not complete the full protocol, corresponding to an intervention attrition rate of 15.6%. Among the 135 non-registrants, 68 completed the follow-up assessment and 67 did not contribute complete pre-post data. Thus, the complete-case sample of 106 participants represents 58.9% of the baseline respondents. The source package used for this revision did not contain verified individual-level reasons for the seven intervention withdrawals or the 67 non-intervention follow-up losses; no reason distribution is therefore asserted. The participant recruitment, intervention completion, and analytic flow are summarized in Figure 1. The corresponding participant counts, percentages, and analytic roles are detailed in Table 1.

Figure 1

Table 1

Reconstructed participant flow and analytic status.

The baseline recruitment pool was defined as

where s_i denotes the i-th student who completed the baseline survey. This baseline pool provided the initial population from which intervention participants and descriptive comparison participants were identified.

The final complete-case quantitative sample was composed of the intervention group and the comparison group:

where denotes the intervention group and denotes the descriptive comparison group. The corresponding group sizes were

This complete-case structure ensured that both baseline and follow-up data were available for all participants included in the quantitative analysis.

Recruitment responsiveness was measured by the proportion of intervention places filled during the registration period. It was calculated as

where N_reg denotes the number of registered students and N_slot denotes the number of available intervention places. This indicator reflects students' initial willingness to participate in a library-based AI-supported wellbeing activity.

Retention was calculated as

where N_complete denotes the number of registered participants who completed all intervention sessions and post-intervention assessments, and N_reg denotes the total number of registered participants. Retention is particularly important for this study because the intervention required repeated participation across three weeks rather than one-time attendance.

Because participant-level baseline data for the seven non-completers were not included in the supplied manuscript package, a formal completer-versus-dropout comparison could not be reproduced in this revision. Attrition is therefore treated as a potential source of systematic bias rather than assumed to be random. If non-completers had greater initial distress, lower engagement, stronger privacy concerns, or less favorable experiences, the completer-only acceptability estimate could be upwardly biased. Conversely, if withdrawal was mainly logistical, bias could be smaller. The revised interpretation therefore reports retention as an implementation outcome and does not equate retention with intervention effectiveness.

The intervention consisted of three developmental bibliotherapy themes selected according to student needs identified from the baseline survey and the practical expertise of librarians and campus counselors. The theme set was defined as

where B₁ represents stress and anxiety management, B₂ represents self-understanding and personal growth, and B₃ represents intimate relationships and social support. These themes were designed to cover both emotional distress and broader developmental concerns.

Each participant's intervention sequence combined weekly bibliotherapy salon participation with AI chatbot-supported interaction. This sequence was represented as

where B_t denotes the bibliotherapy salon in week t, and A_t denotes AI chatbot-supported interaction during the same week. This formulation emphasizes that the two intervention components were not independent add-ons, but parallel and complementary support channels embedded within the same intervention period.

Each bibliotherapy salon followed a structured process:

where r_t denotes selected reading, h_t denotes facilitator guidance, d_t denotes group discussion, and s_t denotes reflection and summary. The structured process was designed to support identification, emotional expression, cognitive reframing, and interpersonal resonance. Reading materials provided symbolic and narrative resources, facilitators guided interpretation, group discussion promoted social resonance, and reflection helped participants connect reading content with personal experience.

AI chatbot participation was represented as

where a_ij denotes the j-th self-reported AI chatbot interaction of participant i. Because backend usage logs were unavailable due to privacy restrictions, AI engagement was treated as self-reported supportive evidence rather than objective behavioral tracking. This decision preserved participant privacy while still allowing the study to examine how students perceived and used the AI component.

To describe the conceptual relationship between the human-guided and AI-supported components, the integrated intervention exposure of participant i was defined as

Here, B_i represents bibliotherapy exposure, A_i represents AI-supported interaction, and α is a conceptual weighting parameter. This equation was not used to estimate a fixed statistical weight. Instead, it clarifies the methodological assumption that the intervention effect, if any, should be understood as emerging from the complementary interaction between reading-based human support and AI-assisted emotional disclosure.

3.3 Measures and quantitative analysis

The quantitative component focused on feasibility indicators, acceptability indicators, and exploratory psychological changes. Feasibility indicators included recruitment responsiveness, retention, session completion, and implementation stability. Acceptability indicators included satisfaction, perceived benefit, preferred themes, and participants' responses to the intervention format. These indicators were analyzed descriptively because the study aimed to evaluate feasibility and acceptability rather than establish causal clinical efficacy.

Participant satisfaction was summarized using the average satisfaction score:

where S_i denotes the satisfaction score of participant i, and N denotes the number of participants who completed the evaluation. Satisfaction was used to assess whether participants considered the intervention format acceptable and appropriate.

Perceived benefit was summarized as

where denotes the perceived benefit rating of participant i. Perceived benefit was interpreted as an acceptability-related indicator rather than a clinical outcome. This distinction prevented the study from overstating subjective evaluation as objective treatment effectiveness.

Psychological wellbeing was assessed using the Symptom Checklist-90. Let x_ij denote the score of participant i on item j. The total mean score was calculated as

This mean-score approach allowed the overall psychological symptom level to be compared before and after the intervention. Lower scores indicate fewer self-reported psychological symptoms.

The average within-group change among intervention participants was calculated as

where N_int denotes the number of intervention completers. This indicator summarizes the average direction and magnitude of pre-post change within the intervention group.

The standard deviation of the change scores was computed as

This value reflects the variability of participants' individual changes and provides the denominator for the paired-sample test and effect size estimation.

The paired-sample statistic was calculated as

where t describes whether the average within-group change differs from zero. Because the intervention was non-randomized, this statistic was used only for exploratory description.

The corresponding standardized effect size was calculated as

Cohen's d was used to describe the magnitude of within-group change. It was not interpreted as definitive evidence of intervention efficacy because the design did not rule out self-selection, maturation, regression to the mean, or other uncontrolled influences.

Baseline comparability between the intervention and comparison groups was examined using the standardized mean difference:

where denotes the baseline mean psychological score of the intervention group, and denotes the baseline mean psychological score of the comparison group. The pooled standard deviation was calculated as

These two equations were used to assess whether the two groups were similar at baseline. Any baseline imbalance was considered important when interpreting exploratory changes.

Using the reported baseline means, group sizes, and the reported between-group test, the baseline SCL-90 standardized mean difference was reconstructed as approximately 0.39, with the intervention group showing greater symptom burden. The approximate 95% confidence interval was [−0.01, 0.79]. Although the interval includes zero, the point estimate is methodologically meaningful and should not be dismissed because the corresponding p value was 0.059. The baseline difference is consistent with self-selection: students experiencing greater distress may have been more motivated to enroll. Therefore, regression to the mean and baseline-dependent change remain plausible alternative explanations for the observed within-group decrease.

A group-by-time model would ordinarily be preferred for comparative outcome analysis:

where β₃ represents the differential temporal change and Z_i contains baseline covariates. However, the supplied source package did not include the comparison group's post-test mean, variance, or participant-level records. A valid estimate of β₃ could therefore not be reproduced and is not reported. This omission is made explicit to prevent the within-group result from being misread as a controlled intervention effect.

The internal consistency of the SCL-90 was assessed using Cronbach's alpha:

In this equation, K denotes the number of items, denotes the variance of item j, and denotes the variance of the total score. Cronbach's alpha was used to examine whether the psychological self-report instrument demonstrated sufficient internal consistency in the current sample.

The observed α =.974 indicates very high internal consistency, but it is not evidence of construct validity and may partly reflect item redundancy in a 90-item instrument. Accordingly, the revised manuscript does not interpret this coefficient as proof that the SCL-90 captures all relevant dimensions of student wellbeing or that the total mean score is free from common-method bias.

To summarize feasibility and acceptability in a transparent way, a descriptive feasibility-acceptability index was defined as

where R_cap denotes recruitment capacity, R_ret denotes retention, S_avg denotes average satisfaction, and B_avg denotes average perceived benefit. The weights were constrained by

The index was used as an organizing concept rather than a definitive score. Its purpose was to make explicit how different implementation and acceptability indicators jointly informed the feasibility judgment.

The FAI was not assigned a numerical value because the component scales were not harmonized, the weights were not prospectively justified, and no validation sample was available. It is therefore retained only as a conceptual protocol template and is not used as an outcome, a hypothesis test, or evidence supporting feasibility. The empirical feasibility conclusion is based directly on separately reported recruitment, retention, and acceptability indicators. Future validation would require prespecified normalization, expert- or data-derived weights, external criteria, and sensitivity analysis across plausible weighting schemes.

3.4 Qualitative analysis and mixed-methods integration

The qualitative component was designed to explain how participants experienced the multimodal AI-supported developmental bibliotherapy model and why they perceived it as useful, limited, or acceptable. Nine participants from the intervention group were selected through purposive sampling to reflect diversity in gender, major, grade, intervention role, and AI-use frequency. Semi-structured interviews focused on motivations for participation, experiences of bibliotherapy salons, experiences of AI chatbot interaction, perceived changes, and suggestions for improvement.

The interview subsample was intended to maximize variation rather than statistical representativeness. The revised reporting follows COREQ principles as closely as the available records permit (Tong et al., 2007). The minimum interview guide comprised the following domains: (1) reasons for registering; (2) expectations before the first salon; (3) helpful and unhelpful aspects of reading and discussion; (4) circumstances in which the chatbot was used; (5) perceived differences between chatbot disclosure and human conversation; (6) any discomfort, privacy concern, or inappropriate response; (7) perceived changes in emotion, thinking, or behavior; and (8) suggestions for service improvement. Interviews were analyzed as experiential accounts, not as independent proof of efficacy. The interview questions and their corresponding analytic purposes are summarized in Table 2.

Table 2

Interview guide and analytic purpose added for qualitative transparency.

The qualitative corpus was denoted as

where q_i represents the transcript of participant i. Each transcript was anonymized before analysis, and participant identifiers were replaced with pseudonyms to protect confidentiality.

The initial coding process was expressed as

where C denotes the set of initial codes, and ϕ(·) represents the interpretive coding process. Coding began with repeated reading of the transcripts, followed by the identification of meaningful units related to emotional expression, reading experience, AI interaction, interpersonal resonance, cognitive reframing, and perceived change.

After initial coding, codes were organized into broader themes:

Here, Θ denotes the set of final themes, and θ_l denotes the l-th theme. Theme development emphasized not only what participants said explicitly, but also how they described the relationship between reading, discussion, AI interaction, and perceived wellbeing-related changes.

To enhance qualitative reliability, two trained research assistants independently coded the transcripts. Intercoder agreement was assessed using Cohen's kappa:

In this equation, P_o denotes observed agreement, and P_e denotes agreement expected by chance. Disagreements were resolved through discussion, and the first author served as arbiter when necessary. This process helped reduce subjective coding bias and improve the trustworthiness of thematic interpretation.

The coding strategy was hybrid. Deductive sensitizing concepts were drawn from the intervention logic, including disclosure, resonance, reframing, activation, and AI limitations; inductive codes were added when participants described experiences not captured by those concepts. The codebook was iteratively revised after independent coding, and disagreements were resolved by discussion rather than by mechanically accepting the majority code. The first author documented interpretive decisions and considered how involvement in intervention design could shape theme development. Because only nine interviews were available, the study does not claim population-level saturation. Instead, the sample is described in terms of information power: the interviews were focused on a narrow intervention, but the small and self-selected sample limits thematic breadth and negative-case coverage.

No verbatim participant quotations were inserted during this revision because the uploaded source package did not contain the anonymized transcripts. Fabricating quotations would compromise qualitative trustworthiness. Representative quotations should be selected from the verified transcripts for each major theme, with pseudonymous identifiers and removal of potentially identifying details. The final audit trail should also state interview duration, setting, recording method, transcription procedure, interviewer training, prior relationships, software used, and whether participants reviewed transcripts or themes.

Mixed-methods integration was conducted after the quantitative and qualitative analyses were completed separately. The final integrated interpretation was defined as

where Ψ denotes the integrated interpretation, denotes feasibility evidence, ΔP denotes exploratory psychological change, and Θ denotes qualitative themes. This formulation reflects the central logic of the study: feasibility, psychological trends, and participant experience should be interpreted together.

Convergence across evidence sources was represented as

where 𝕀(·) is an indicator function, Q_k denotes the k-th quantitative finding, θ_k denotes the corresponding qualitative theme, and ~ indicates interpretive consistency. This equation was used conceptually to describe whether quantitative findings and qualitative themes supported the same conclusion.

Finally, the overall methodological target was summarized as

This equation captures the methodological goal of the present feasibility study. Rather than maximizing only clinical symptom reduction or algorithmic prediction accuracy, the proposed model aims to optimize feasibility, acceptability, and contextual validity in a real university library environment.

4 Experimental setup

4.1 Environment

The experiment was conducted in a real university library environment rather than in a laboratory setting. This design choice is consistent with the aim of examining whether a multimodal AI-supported developmental bibliotherapy model can be feasibly implemented and accepted as a low-threshold support service in students' everyday academic environment. The intervention was delivered over three consecutive weeks, from April 27 to May 17, 2025. Three developmental bibliotherapy salons were organized in library discussion rooms, and each session lasted approximately 75–90 min. The human-support environment involved university librarians, campus counselors, and student reading leaders. The AI-support environment was based on an AI chatbot platform that allowed text- and voice-based emotional expression, emotional journaling, and access to mental-health-related educational resources. Because the platform's privacy policy did not permit access to backend interaction logs, AI-use information was collected through participant self-report and voluntary screenshot checks.

4.1.1 AI component reproducibility, safety, and privacy specification

The AI component was supportive rather than diagnostic or therapeutic. Participants could use text or voice interaction for reflection, emotional expression, journaling, and access to general educational resources. The chatbot was not treated as a substitute for professional care, and no clinical diagnosis or medication recommendation was an intended study function. The study records available for analysis did not document the platform name, provider, underlying model/version, standardized prompt, safety-filter configuration, or whether human monitoring occurred; these fields are therefore reported as not documented rather than inferred. The available and unavailable information relevant to AI-system reproducibility, safety, monitoring, privacy, and adverse-event reporting is summarized in Table 3.

Table 3

AI intervention reproducibility and safety reporting matrix.

A defensible safety workflow for this type of service includes clear disclosure of AI limitations, instructions not to use the chatbot for emergencies, visible campus counseling and emergency contacts, a protocol for responding to severe SCL-90 indicators or direct reports of self-harm, and documentation of adverse or concerning chatbot responses. The authors should confirm which elements were implemented contemporaneously. Voice interaction may involve third-party processing beyond the university; the final manuscript must therefore specify data controller responsibilities, storage location, retention period, access control, and whether participants could opt for text-only use.

4.2 Data sources and analytic corpora

To reflect the multimodal and mixed-methods nature of the study, two datasets were constructed. In the revised terminology, these are analytic data sources or corpora rather than benchmark datasets: they were assembled for service evaluation, are not presented as publicly reusable machine-learning datasets, and contain sensitive human-subject information. Dataset 1, the Library Feasibility and Psychological Self-Report Dataset, contains complete-case quantitative records from 106 students, including 38 intervention participants and 68 descriptive comparison participants. Formally,

where x_i represents demographic information, g_i group status, and SCL-90 scores, e_i evaluation indicators, and a_i self-reported AI-use information. Dataset 2, the Interview-Based Perception Dataset, consists of semi-structured interview data from nine intervention participants:

where t_j denotes the anonymized transcript, m_j participant metadata, and c_j assigned codes.

4.3 Baselines

Since the present work is a feasibility and acceptability study rather than a randomized controlled efficacy trial, the baselines were defined as comparative research settings and analytical reference models. Table 4 summarizes the reference settings used to contextualize the proposed model.

Table 4

Baseline and reference settings used to contextualize the proposed model.

4.4 Evaluation protocol

Four categories of indicators were used: implementation feasibility, participant acceptability, exploratory psychological change, and qualitative experiential interpretation. Implementation feasibility was evaluated through recruitment responsiveness, registration capacity, session completion, and retention. Acceptability was assessed through participant evaluation, perceived benefit, preferred themes, and willingness to engage with the intervention format. Exploratory psychological change was assessed using pre- and post-intervention SCL-90 scores. Qualitative analysis followed thematic analysis (; Tong et al., 2007). The final mixed-methods interpretation was based on convergence across three evidence sources:

where Ω denotes the integrated interpretation, quantitative evidence, qualitative evidence, and baseline literature.

The comparison group was not an untreated control group created by random allocation. Its role was descriptive: to show who remained in the survey cohort without entering the program and to contextualize baseline differences. The primary analysis therefore reports feasibility and acceptability directly, and the symptom analysis is limited to a completer-based within-group estimate with a confidence interval. Covariate-adjusted causal effects, propensity-score estimates, and group-by-time effects are not presented because the necessary individual-level data and a defensible assignment mechanism were not available in the uploaded package.

5 Experiments and analysis

5.1 Overall performance comparison and analysis

The overall performance was evaluated from implementation performance and exploratory psychological outcome. The primary focus was not to claim causal clinical effectiveness, but to examine whether the proposed model could be feasibly implemented, accepted by students, and meaningfully experienced in a real university library environment.

The intervention-group mean change was . From t(37) = 2.26, the standard error of the change was approximately 0.053, yielding a 95% confidence interval of approximately [−0.23, −0.01]. This interval quantifies uncertainty around the completer-only change but does not control for the comparison-group trajectory. The reported standardized within-group effect (d = 0.37) is small and should be interpreted as descriptive. The 94.7% perceived-benefit figure corresponds to 36 of 38 completers; the remaining two participants (5.3%) reported limited benefit. “Moderate” and “considerable” refer to the two upper response categories of the post-program perceived-benefit item, whereas “limited” denotes the lower category.

Figures 2, 3, together with Tables 5, 6, show that the proposed model achieved strong implementation performance in the university library setting. Rapid recruitment indicates that the library-based format had a low participation threshold, while the retention rate of 84.4% is encouraging for a multi-session wellbeing-support activity delivered in a non-clinical campus setting. The SCL-90 decrease from 1.64 to 1.52 provides preliminary support for a possible reduction in self-reported symptoms, although this should be interpreted cautiously because the study was non-randomized. Available baseline characteristics and group-comparability information are summarized in Table 7; unavailable summary values and covariates are reported as missing rather than inferred.

Figure 2

Figure 3

Table 5

Overall implementation and acceptability performance of the proposed model.

Table 6

Exploratory psychological outcome analysis based on SCL-90 scores.

Table 7

Baseline characteristics and group-comparability information available in the supplied source package.

5.2 Conceptual component-level mechanism analysis

Because the intervention was not randomized into separate component groups, the ablation analysis was conducted at the design and mechanism level. It evaluates what type of evidence or support function would be weakened if one component were removed.

This section is not an empirical ablation study. Participants were not assigned to salon-only, chatbot-only, library-free, or mixed-methods-free variants. The table therefore summarizes hypothesized component functions and participant-perceived mechanisms; it does not estimate component effects. The resulting conceptual component-level mechanism analysis is summarized in Table 8; it should be interpreted as a design-level comparison rather than an empirical ablation or dismantling experiment. Statements about the “full model” are interpreted as design propositions requiring future factorial, dismantling, or multi-arm evaluation.

Table 8

Conceptual component-level mechanism analysis of the proposed model; no experimental dismantling was conducted.

The integrated design was intended to provide a broad support structure by combining human-guided developmental bibliotherapy, AI-supported emotional expression, and mixed-methods evaluation; however, the present data cannot demonstrate that it is superior to any isolated component. Compared with AI-only tools, the proposed model preserves human interaction and group reflection. Compared with conventional reading therapy, it adds an AI-supported channel that extends emotional expression beyond the face-to-face session.

5.3 Participant attrition, selection bias, and sensitivity of interpretation

Seven of 45 registrants did not complete the full intervention, and 67 of 135 non-registrants did not provide complete follow-up data. Because verified withdrawal reasons and individual baseline profiles were not supplied, missingness cannot be assumed to be completely at random. The complete-case analysis may overrepresent participants with greater availability, stronger engagement, more favorable experiences, or fewer privacy concerns. The high perceived-benefit proportion should therefore be understood as the proportion among completers, not among all registrants or all baseline respondents.

The voluntary allocation mechanism also creates confounding by indication: students with higher baseline distress may have been more likely to seek the intervention. The baseline SCL-90 point estimate supports this possibility. Regression to the mean could then produce a pre-post reduction even without an intervention-specific effect. A future confirmatory study should prespecify a primary outcome, use randomized or matched allocation, retain reasons for non-participation and withdrawal, conduct intention-to-treat analysis, use multiple imputation or inverse-probability weighting for missing outcomes, and report a group-by-time estimate with covariate adjustment.

5.4 Mechanism-oriented and mixed-methods analysis

Figure 4 and Table 9 show that the two intervention components served distinct but complementary functions. AI mainly supported private emotional disclosure and low-pressure interaction, while the bibliotherapy salons supported interpersonal resonance and cognitive reframing. This two-layer structure explains why the integrated model is more suitable than a single-channel intervention.

Figure 4

Table 9

Mechanism-oriented analysis of the proposed model.

5.5 Supplementary experimental results

Figures 5–7 summarize the supplementary descriptive results on role-specific AI engagement, bibliotherapy-theme preferences, and perceived benefit after the intervention. The role-AI engagement matrix shows that AI chatbot engagement was not restricted to a small subset of highly active users; both general participants and reading leaders reported repeated use during the intervention period. The theme-preference distribution indicates that self-understanding and stress reduction were equally important, suggesting that students participated not only because of distress but also because of broader developmental needs. The perceived-benefit distribution further supports the acceptability of the proposed model, with 36 of 38 completers reporting moderate or considerable benefit.

Figure 5

Figure 6

Figure 7

5.6 Robustness, error analysis, and clustering visualization

Robustness was evaluated from implementation stability, measurement reliability, baseline comparability, qualitative coding consistency, and AI-use uncertainty. Table 10 summarizes the evidence. Figures 8–11 provide the exploratory PCA, hierarchical clustering, normalized centroid profiles, and participant-feature heatmap used only to visualize case-level heterogeneity in the nine-participant interview subsample.

Table 10

Robustness and sensitivity analysis of the mixed-methods evidence.

Figure 8

Figure 9

Figure 10

Figure 11

The robustness table should be read conservatively. High Cronbach's alpha supports score consistency but may also indicate redundancy; non-significant age and gender tests do not prove equivalence; and the baseline SCL-90 difference remains relevant despite p =.059. Most importantly, neither coding agreement nor cross-source convergence removes self-selection or common-method bias.

The PCA, dendrogram, centroid plot, and heatmap are retained only as exploratory visual summaries of nine coded cases. With n = 9, component loadings, cluster number, centroids, and apparent separation are highly sensitive to coding choices, scaling, and individual observations; they cannot be validated, generalized, or used to infer stable participant subtypes. The labels “facilitator-engaged,” “high-AI expressive,” and “moderate-use” are descriptive annotations applied to this small sample rather than empirically established classes. No hypothesis test, predictive accuracy, stability resampling, or external validation is claimed. The substantive qualitative interpretation is based on thematic analysis, while these graphics serve only as transparent displays of case-level heterogeneity.

6 Conclusion

This study proposed and evaluated a multimodal AI-supported developmental bibliotherapy model in a university library context. By integrating developmental bibliotherapy salons, AI chatbot-supported emotional expression, psychological self-report, participant evaluation, and qualitative interviews, the study examined whether a low-threshold, human-centered, and AI-assisted support model could be feasibly implemented and accepted by university students. Different from purely clinical interventions or algorithm-centered affective computing systems, the proposed model emphasizes the complementary relationship between human-guided reading activities and AI-mediated private emotional disclosure. The university library is therefore positioned not as a substitute for professional counseling services, but as an accessible support environment that can promote self-reflection, emotional expression, interpersonal resonance, and developmental growth.

The empirical findings suggest that the proposed model demonstrated good feasibility and acceptability. Recruitment reached full capacity within two days, the retention rate was 84.4%, and 94.7% of completers reported moderate or considerable benefit. Exploratory quantitative analysis showed a small reduction in SCL-90 scores among intervention participants after the three-week program. Qualitative findings explained this pattern by revealing perceived mechanisms including emotional disclosure, cognitive reframing, interpersonal resonance, and behavioral activation. Additional analyses, including component-level ablation, robustness analysis, error analysis, engagement visualization, and clustering-based profiling, further supported the internal coherence of the proposed model.

In the revised interpretation, the component table is conceptual rather than ablative, and the PCA/clustering figures are purely exploratory visualizations. Neither analysis strengthens causal inference. The most defensible conclusions are that the program could be delivered, most completers evaluated it positively, and participants described distinct perceived functions for group reading and private chatbot interaction.

The study also contributes to the broader topic of multimodal data processing in neuroscience and perception science. This contribution is deliberately narrow: the study does not provide neural, physiological, perceptual-sensor, or computational-fusion evidence. Its primary contribution is a feasibility study of heterogeneous mixed-methods evidence integration in a library-based, AI-supported student wellbeing service. Although this work does not aim to develop a new emotion recognition algorithm, it extends multimodal thinking into a real-world wellbeing-support scenario. The model integrates heterogeneous forms of evidence, including self-reported psychological symptoms, AI-use feedback, participant evaluations, interview narratives, and engagement-profile visualization. This provides an application-oriented example of how multimodal affective and perceptual information can be organized to understand students' wellbeing experiences in context. In particular, the findings suggest that AI-supported systems may be most useful when embedded within human-centered service ecosystems rather than deployed as isolated digital tools.

Several limitations should be acknowledged. First, the intervention was voluntary and non-randomized, so the exploratory psychological changes cannot be interpreted as causal evidence of clinical effectiveness. Second, the study was conducted at a single university, and the sample was relatively small, especially for the qualitative and clustering analyses. Third, AI chatbot use was measured mainly through self-report because backend usage logs were unavailable due to privacy restrictions. Fourth, the intervention lasted only three weeks, so long-term sustainability and delayed effects remain unknown. Future research should use randomized controlled or quasi-experimental designs, recruit larger and more diverse samples, incorporate privacy-preserving AI-use logs, and add longitudinal follow-up.

Additional limitations include incomplete documentation of the AI platform and safety configuration, absence of verified reason-level attrition data in the supplied package, incomplete demographic summary statistics, lack of a reproducible comparison-group follow-up estimate, and absence of source transcripts for quotation verification. These limitations restrict reproducibility and prevent adjustment for academic year, discipline, previous service use, and prior chatbot exposure. A confirmatory study should preregister the allocation and analysis plan, distinguish feasibility from efficacy outcomes, document adverse events, use a complete participant-flow diagram, and follow COREQ for qualitative reporting.

7 Ethics, informed consent, risk management, and adverse events

The study flow diagram indicates that informed consent was obtained at baseline. However, the uploaded source package does not contain the name of the approving ethics committee, approval number, approval date, approved protocol, or the exact written-consent language. The final statement should also confirm whether participation was voluntary, whether withdrawal was permitted without penalty, and how participants were informed that the AI chatbot was not a substitute for professional mental-health care.

Because the study involved psychological distress assessment and AI-mediated emotional disclosure. At minimum, it should specify how severe symptoms, suicidality, or acute distress were identified; who reviewed concerning information; which campus counseling, clinical, or emergency services received referrals; and how urgent risk was escalated. The manuscript must also report the verified number of adverse events and concerning chatbot responses. Absence of an event log in the supplied package cannot be interpreted as evidence that zero events occurred.

For privacy protection, it should specify anonymization or pseudonymization, encryption, access permissions, storage duration, deletion procedures, and whether screenshots were redacted before analysis. Platform-provider access to text or voice data, cross-border processing, and reuse for model training should be disclosed where applicable.

Statements

Data availability statement

The original contributions presented in the study are included in the article/Ssupplementary material, further inquiries can be directed to the corresponding author.

Author contributions

CZ: Data curation, Formal analysis, Investigation, Writing – original draft. WK: Methodology, Writing – original draft, Writing – review & editing.

Funding

The author(s) declared that financial support was received for this work and/or its publication. This work was supported by the 2026 Teaching Reform Research Project of the Library and Information Work Committee of Jiangsu Universities (Project No. 2026JTYB17), under the project “A Study on the Developmental Bibliotherapy Service Model in Academic Libraries from the Perspective of ‘Three-Wide Education'.”

Acknowledgments

The authors sincerely thank the researchers, institutions, and open-data communities that have made psychological, educational, conversational, and multimodal neuroscience datasets publicly available. These open-source resources have greatly supported reproducible research and the development of human-centered AI, affective perception, and student wellbeing studies. The authors also thank the university library staff, student reading leaders, and participants who supported the implementation and evaluation of the developmental bibliotherapy activities.

Conflict of interest

The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Generative AI statement

The author(s) declared that Generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

References

• 1

  BraunV.ClarkeV. (2006). Using thematic analysis in psychology. Qual. Res. Psychol. 3, 77–101. doi: 10.1191/1478088706qp063oa

  • CrossRef
  • Google Scholar
  • View reference in article

• 2

  BrewsterL. (2014). The public library as therapeutic landscape: a qualitative case study. Health Place26, 94–99. doi: 10.1016/j.healthplace.2013.12.015

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 3

  BruffaertsR.MortierP.KiekensG.AuerbachR. P.CuijpersP.DemyttenaereK.et al. (2018). Mental health problems in college freshmen: prevalence and academic functioning. J. Affect. Disord. 225, 97–103. doi: 10.1016/j.jad.2017.07.044

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 4

  CampbellF.BlankL.CantrellA.BaxterS.BlackmoreC.DixonJ.et al. (2022). Factors that influence mental health of university and college students in the uk: a systematic review. BMC Public Health22:1778. doi: 10.1186/s12889-022-13943-x

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 5

  CastilloL. G.SchwartzS. J. (2013). Introduction to the special issue on college student mental health. J. Clin. Psychol. 69, 291–297. doi: 10.1002/jclp.21972

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 6

  ChandraumakanthamO.GowthamN.ZakariahM.AlmazyadA. (2024). Multimodal emotion recognition using feature fusion: an llm-based approach. IEEE Access12, 108052–108071. doi: 10.1109/ACCESS.2024.3425953

  • CrossRef
  • Google Scholar
  • View reference in article

• 7

  CoxA.BrewsterL. (2020). Library support for student mental health and well-being in the uk: Before and during the covid-19 pandemic. J. Acad. Libr.. 46:102256. doi: 10.1016/j.acalib.2020.102256

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 8

  EbertD. D.MortierP.KaehlkeF.BruffaertsR.BaumeisterH.AuerbachR. P.et al. (2019). Barriers of mental health treatment utilization among first-year college students: First cross-national results from the who world mental health international college student initiative. Int. J. Methods Psych. Res. 28:e1782. doi: 10.1002/mpr.1782

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 9

  FengY.HangY.WuW.SongX.XiaoX.DongF.et al. (2025). Effectiveness of ai-driven conversational agents in improving mental health among young people: systematic review and meta-analysis. J. Med. Internet Res. 27:e69639. doi: 10.2196/69639

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 10

  GülerS. E.AkbulutF. P. (2025). Multimodal emotion recognition: emotion classification through the integration of EEG and facial expressions. IEEE Access 13, 24587–24603. doi: 10.1109/ACCESS.2025.3538642

  • CrossRef
  • Google Scholar
  • View reference in article

• 11

  HamdanW.Al DuaijyL.Al SawyY. (2021). Effectiveness of bibliotherapy in alleviating examination stress on college students: a quasi-experimental trial. Health Inf. Libr. J. 38, 172–181. doi: 10.1111/hir.12312

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 12

  HongS.KangH.ChoH. (2024). Cross-modal dynamic transfer learning for multimodal emotion recognition. IEEE Access12, 14324–14333. doi: 10.1109/ACCESS.2024.3356185

  • CrossRef
  • Google Scholar
  • View reference in article

• 13

  KalatehS.Estrada-JimenezL. A.Nikghadam-HojjatiS.BarataJ. (2024). A systematic review on multimodal emotion recognition: building blocks, current state, applications, and challenges. IEEE Access12, 103976–104019. doi: 10.1109/ACCESS.2024.3430850

  • CrossRef
  • Google Scholar
  • View reference in article

• 14

  LattieE. G.AdkinsE. C.WinquistN.Stiles-ShieldsC.WaffordQ. E.GrahamA. K. (2019). Digital mental health interventions for depression, anxiety, and enhancement of psychological well-being among college students: systematic review. J. Med. Internet Res. 21:e12869. doi: 10.2196/12869

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 15

  LinR.HuH. (2023). Multi-task momentum distillation for multimodal sentiment analysis. IEEE Trans. Affect. Comput. 15, 549–565. doi: 10.1109/TAFFC.2023.3282410

  • CrossRef
  • Google Scholar
  • View reference in article

• 16

  LiuR.ZuoH.LianZ.SchullerB. W.LiH. (2024a). Contrastive learning based modality-invariant feature acquisition for robust multimodal emotion recognition with missing modalities. IEEE Trans. Affect. Comput. 15, 1856–1873. doi: 10.1109/TAFFC.2024.3378570

  • CrossRef
  • Google Scholar
  • View reference in article

• 17

  LiuW.LiW.RuanY.-P.ShuY.ChenJ.LiY.et al. (2024b). Weakly correlated multimodal sentiment analysis: New dataset and topic-oriented model. IEEE Trans. Affect. Comput. 15, 2070–2082. doi: 10.1109/TAFFC.2024.3396144

  • CrossRef
  • Google Scholar
  • View reference in article

• 18

  LiuY.ZhaoY.ZhuS.PanS.ZhangL.PanS.et al. (2025). A meta-analysis of the effectiveness of reading therapy for college students' mental health problems in university libraries. J. Acad. Libr. 51:103052. doi: 10.1016/j.acalib.2025.103052

  • CrossRef
  • Google Scholar
  • View reference in article

• 19

  Madrid-CagigalA.KealyC.PottsC.MulvennaM. D.ByrneM.BarryM. M.et al. (2025). Digital mental health interventions for university students with mental health difficulties: a systematic review and meta-analysis. Early Interv. Psych. 19:e70017. doi: 10.1111/eip.70017

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 20

  MorinR. A.LaVertuA. E. (2025). Bibliotherapy for stress management: a wellness intervention for first-year medical students. J. Med. Libr. Assoc. 113:158. doi: 10.5195/jmla.2025.1830

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 21

  NyakharS.WangH. (2025). Effectiveness of artificial intelligence chatbots on mental health &well-being in college students: a rapid systematic review. Front. Psych. 16:1621768. doi: 10.3389/fpsyt.2025.1621768

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 22

  Organization W. H. (2022). World Mental Health Report: Transforming Mental Health for all. Geneva: World Health Organization.

  • Google Scholar
  • View reference in article

• 23

  RadoiA.CioroiuG. (2024). Uncertainty-based learning of a lightweight model for multimodal emotion recognition. IEEE Access12, 120362–120374. doi: 10.1109/ACCESS.2024.3450674

  • CrossRef
  • Google Scholar
  • View reference in article

• 24

  RamseyE.AagardM. C. (2018). Academic libraries as active contributors to student wellness. Coll. Undergrad. Libr. 25, 328–334. doi: 10.1080/10691316.2018.1517433

  • CrossRef
  • Google Scholar
  • View reference in article

• 25

  Reyes-PortilloJ. A.SoA.McAlisterK.NicodemusC.GoldenA.JacobsonC.et al. (2025). Generative ai-powered mental wellness chatbot for college student mental wellness: Open trial. JMIR Format. Res. 9:e71923. doi: 10.2196/71923

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 26

  SherrillW. W.LawsonK. N.BednarH. S. (2020). The act of listening well: improving communication skills through bibliotherapy. Health Commun. 35, 1837–1840. doi: 10.1080/10410236.2019.1663584

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 27

  ShiP.HuM.NakagawaS.ZhengX.ShiX.RenF. (2025). Text-guided reconstruction network for sentiment analysis with uncertain missing modalities. IEEE Trans. Affect. Comput. 16, 1825–1838. doi: 10.1109/TAFFC.2025.3541743

  • CrossRef
  • Google Scholar

• 28

  StorrieK.AhernK.TuckettA. (2010). A systematic review: students with mental health problems-a growing problem. Int. J. Nurs. Pract. 16, 1–6. doi: 10.1111/j.1440-172X.2009.01813.x

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 29

  SunL.LianZ.LiuB.TaoJ. (2023). Efficient multimodal transformer with dual-level feature restoration for robust multimodal sentiment analysis. IEEE Trans. Affect. Comput. 15, 309–325. doi: 10.1109/TAFFC.2023.3274829

  • CrossRef
  • Google Scholar

• 30

  SwainD. K.SwainC.RautarayB. (2014). Bibliometric analysis of the journal of educational media and library science from 2008 to 2012. PEARL J. Libr. Inf. Sci. 8, 9–14. doi: 10.5958/j.0975-6922.8.1.002

  • CrossRef
  • Google Scholar

• 31

  TongA.SainsburyP.CraigJ. (2007). Consolidated criteria for reporting qualitative research (COREQ): a 32-item checklist for interviews and focus groups. Int. J. Qual. Health Care19, 349–357. doi: 10.1093/intqhc/mzm042

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 32

  ZhangQ.ZhangR.XiongY.SuiY.TongC.LinF.-H. (2025a). Generative AI mental health chatbots as therapeutic tools: systematic review and meta-analysis of their role in reducing mental health issues. J. Med. Internet Res. 27:e78238. doi: 10.2196/78238

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 33

  ZhangX.LiZ.ZhangM.YinM.YangZ.GaoD.et al. (2025b). Exploring artificial intelligence (AI) chatbot usage behaviors and their association with mental health outcomes in Chinese university students. J. Affect. Disord. 380, 394–400. doi: 10.1016/j.jad.2025.03.141

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

Keywords

affective perception, developmental bibliotherapy, human-centered AI, multimodal data processing, student wellbeing

Citation

Zhou C and Kang W (2026) Multimodal data processing for affective perception in AI-supported developmental bibliotherapy: a mixed-methods feasibility study in university libraries. Front. Psychol. 17:1927988. doi: 10.3389/fpsyg.2026.1927988

Received

04 July 2026

Revised

04 August 2026

Accepted

15 September 2026

Published

06 October 2026

Volume

17 - 2026

Reviewed by

Yan Zhu, Shanghai Dianji University, China

Pu Bin, The Hong Kong University of Science and Technology, Hong Kong SAR, China

Updates

Copyright

© 2026 Zhou and Kang.

This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

*Correspondence: Chenyang Zhou, Zhou_Chenyang@jssnu.edu.cn

Disclaimer

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.