Research on Integrating Narrative Simulation and Visual AI into Teaching Management to Build Mental Health Capacity in Universities

Summary

This study integrates narrative simulation and visual AI into teaching management systems to improve university students' mental health, emotional resilience, and engagement.

Abstract

University students increasingly face mental health challenges, including anxiety, depression, and stress, yet most higher education environments lack proactive systems for emotional monitoring and support. This study aimed to design, implement, and evaluate a hybrid teaching management system that integrates narrative simulation with visual artificial intelligence (AI) to promote mental health literacy and emotional resilience among students. The system includes a narrative decision-making module simulating stress scenarios and an AI-powered emotion recognition tool (based on facial expression detection) embedded in classroom settings. A mixed-methods design was employed with 332 undergraduate students and 15 faculty members from multiple universities. Pre- and post-intervention surveys, usage logs, and real-time emotional data from visual AI were collected. Quantitative data were analyzed using descriptive statistics, paired t-tests, ANOVA, and multiple linear regression. Results indicated statistically significant improvements in students' mental health scores (p < 0.01), emotional awareness, and decision-making confidence. This integrated approach demonstrates both usability and scalability, offering instructors early emotional insight and students a reflective learning environment. The method is best suited for institutions equipped with AI-capable classrooms and trained ethical oversight.

Introduction

Mental health has become a pressing concern in higher education, with rising levels of anxiety, depression, and burnout reported among university students worldwide¹. These psychological burdens, often linked to academic overload, social pressures, and financial uncertainty, negatively impact academic performance, retention, and long-term well-being². While many institutions have implemented counseling services and crisis interventions, such approaches tend to be reactive and limited in reach³. There is a growing demand for proactive, scalable, and embedded strategies that support students' psychological resilience within everyday learning environments.

Recent advances in educational technology offer new opportunities to integrate emotional support directly into teaching systems. Two promising approaches are narrative simulation and visual artificial intelligence (AI). Narrative simulation refers to interactive, story-driven modules that place students in emotionally complex scenarios, encouraging reflection, empathy, and decision-making. Grounded in experiential learning theory, these simulations allow learners to rehearse responses to challenges such as test anxiety, time management, or peer conflict in a safe, guided environment^4,5. Experiential Learning Theory further suggests that reflection and active experimentation are crucial for transforming experience into durable learning, making narrative simulation a viable means to foster emotional intelligence and adaptive coping.

Visual AI uses machine-learning models to infer students' emotional states in real time, e.g., from facial expressions, gaze, or posture⁶. When deployed ethically and with informed consent, such systems can function as early-warning tools, identifying stress or disengagement before escalation⁷. Embedded within teaching-management platforms, visual AI can provide timely feedback to instructors, supporting personalized academic or emotional responses. This integration aligns with Cognitive-Behavioral Theory, which emphasizes that increasing awareness of maladaptive cognitions and restructuring them can improve emotional outcomes.

This approach also resonates with contemporary views of teaching management as the orchestration of student experience across digital and in-person settings. By embedding mental-health support into this infrastructure, resilience training becomes part of the academic process rather than an optional add-on⁸. It further accords with the Transactional Model of Stress and Coping, which frames emotion regulation as a dynamic interplay between appraisal and coping strategies; here, visual AI contributes to primary appraisal (detecting cues), whereas narrative simulation scaffolds secondary appraisal (exploring coping alternatives) in a psychologically safe space.

Despite growing interest, few empirical studies have evaluated the combined use of visual AI and narrative simulation for mental-health support in higher education. Moreover, key constructs such as mental-health literacy, "the knowledge and beliefs that aid in the recognition, management, or prevention of mental disorders"⁹, and emotional resilience, the capacity to adapt to psychological stress¹⁰, are often under-operationalized. This study addresses these gaps by designing and testing an integrated teaching-management system that pairs narrative simulation with visual-AI components to promote early identification of distress and foster student self-regulation.

Objective and key contribution

This study explores how narrative simulation and visual AI can be meaningfully integrated into teaching-management systems to address mental-health concerns among university students. Rather than treating mental health as an external service, the goal is to embed emotional engagement and early psychological support within routine learning. Combining interactive decision-making scenarios with real-time emotional monitoring affords students opportunities for reflective practice while enabling instructors to respond sensitively to signs of distress.

The system is conceptually grounded in experiential learning, highlighting practice, feedback, and emotional involvement as drivers of behavior change. Narrative simulations present familiar academic stressors (e.g., time pressure, peer tension) that require emotionally salient choices. In parallel, a visual-AI tool calibrated for facial-expression inference monitors moment-to-moment changes during teaching sessions. The approach also draws on Emotional-Intelligence theory, which identifies self-awareness, empathy, and regulation as core competencies of resilience.

The research involved 332 undergraduate students and 15 faculty members across three subject areas. Using a mixed-methods design, we assessed whether anxiety, depression, and stress decreased, and whether students' emotional awareness and decision confidence improved. Statistical modeling and qualitative analyses jointly inform how emotion-aware technologies can support student well-being, not as a replacement for professional care, but as a preventive, scalable component of everyday academic life.

System overview

This study is structured into five main sections. The introduction presents the background and motivation for addressing mental health challenges in higher education. The second section reviews existing literature on narrative simulation, visual AI, and their applications in educational contexts. The third section outlines the research design, including the mixed-methods approach, participant recruitment, and analytical techniques. The fourth section presents the key findings, both quantitative and qualitative, focusing on changes in students' emotional resilience and mental health awareness. Finally, the conclusion discusses the broader implications of the findings and suggests future directions for the integration of emotion-aware technologies into teaching practice.

Related work

Universities have trialed diverse approaches to support mental health, including peer-led outreach, digital campaigns, and classroom interventions¹¹. Campus-wide messaging combined with student peer support can improve service engagement and awareness, though tailoring across diverse populations remains challenging¹². At the population level, research often relies on self-report surveys grounded in diagnostic criteria¹³, which limits causal inference and overlooks real-time indicators of strain^14,15.

A growing line of work examines emotional intelligence (EI) as a predictor of student well-being; traits such as self-awareness, clarity, and emotional repair are associated with better outcomes, with belongingness frequently identified as a mediator¹⁶. Yet many findings are based on one-time surveys, lacking dynamic feedback for sustained support. Incorporating EI-informed principles into narrative simulation offers a route to deepen empathy, reflection, and self-regulation.

In parallel, educational researchers increasingly integrate AI into learning environments¹⁷. Visual-AI tools that detect facial expressions or stress-related micro-expressions have been deployed in examinations and tutoring, with promising results for adaptive feedback¹⁸. Narrative simulations are well-established in nursing/medical education for practicing emotionally charged decisions¹⁹, but these applications are often domain-specific and seldom integrated into general teaching-management systems²⁰. A small body of work proposes frameworks for student-AI collaboration and AI-driven instructional design^21,22, emphasizing ethics, implementation gaps, and the need for scalable models that deliver pedagogical and psychological value beyond novelty.

This study advances the field by integrating narrative simulation with real-time affective computing to create a proactive, embedded response to mental-health challenges in university learning environments.

Research gap and practical implementation considerations

Although digital tools for student mental health are proliferating, many studies rely on self-report and short-term interventions outside classroom contexts, limiting long-term impact and timeliness. Few efforts connect constructs like EI or mental-health literacy to real-time feedback or observable classroom behavior. Visual AI and narrative simulation have been explored separately, but rarely combined into a single system that both tracks and responds to emotional states during academic activity. This study addresses that gap by pairing simulation-based emotional reflection with AI-driven recognition inside the learning environment itself.

Practical implementation considerations

Effective deployment depends on infrastructure (stable internet, integrated cameras), ethical oversight, and algorithm calibration to protect participants and reduce bias, particularly when facial analysis is involved. Faculty training is also necessary to interpret AI-generated cues responsibly and to integrate them into pedagogical decisions. Consequently, the approach best fits universities with established digital infrastructures and institutional ethics committees, where it can function as a proactive, scalable framework for student mental-health empowerment.

Protocol

Prior to data collection, ethical approval was obtained from the Institutional Review Board of Nanjing University of Posts and Telecommunications (Approval Code: NJUPT-IRB-2024-0312). All participants provided written informed consent after being briefed on the study scope, the voluntary nature of participation, the use of visual AI for emotion tracking, and data confidentiality procedures. Participation could be discontinued at any time without penalty. The software and equipment used are listed in the Table of Materials.

1. Study preparation

1. Definition of the operational timeline and end-to-end flow
   Two assessment points were set: T1 (pre-intervention), immediately after the first exam, and T2 (post-intervention), following brief feedback delivered by instructors guided by AI emotion cues. One narrative simulation was scheduled after the exam period, and a second simulation several days later to assess changes in coping. The end-to-end pipeline, encompassing classroom camera input, emotion-recognition API, instructor feedback, survey collection, data integration, and analysis, is illustrated in Figure 1A.
2. Recruitment and documentation of participants
   A total of 347 participants were recruited through voluntary enrollment from existing courses, including 332 undergraduate students (ages 18-22; 140 male, 192 female) and 15 faculty members (ages 27-45; 10 female, 5 male). Academic-year and departmental distributions were recorded. Student and faculty demographics are presented in Table 1 and Table 2, respectively.
3. Preparation of classrooms and materials
   Each room was equipped with a computer (Intel i5 or higher, ≥ 8 GB RAM), an HD webcam (≥ 720p), and stable internet connectivity (≥ 10 Mbps). The webcam was positioned at eye level under uniform front lighting to ensure clear facial visibility at ≥ 640 × 480 px. A notice indicated that emotion tracking was active. Framing and lighting were verified before each session, and software configuration was initiated once detection stability was confirmed.

2. Visual AI configuration

1. Provisioning of the emotion-recognition service
   A cloud account was created, and the emotion-recognition API was enabled. The API key and endpoint URL were generated, with credentials stored in encrypted form under role-based access control. For reproducibility, Microsoft Azure Face API (Emotion, v3.2, East Asia region) was employed. The client authenticated via HTTPS using a bearer token and returned per-face JSON objects containing {faceId, faceRectangle, emotions: {anger, fear, sadness, disgust, happiness, surprise, neutral}, timestamp} at approximately 25-30 fps.
2. Connection of the camera and calibration of detection
   The classroom webcam stream was linked to the API client (Python or REST). The sampling rate was set to 25-30 fps with a frame resolution of 640 × 480 px. A test stream was initiated, and connectivity was confirmed by receiving an HTTP 200 OK response with valid timestamped JSON. The camera angle and lighting were adjusted until faces were detected within one second of frame entry. The final camera position and lighting parameters were documented. Calibration was evidenced by a confirmation screen displaying bounding boxes, the 200 OK indicator, and JSON fields, as illustrated in Figure 1B.
3. Definition of classification thresholds and logging rules
   Each JSON output included ISO 8601 timestamps, emotion labels, and confidence values. A negative-affect composite per frame was defined as the mean of {anger, fear, sadness, disgust} confidence scores. A stress flag was triggered when the composite remained ≥ 0.75 for at least 5 s (rolling window). A 10 min pilot run was conducted to optimize thresholds, targeting a per-session false-positive rate of ≤5%. Logging was configured to aggregate session-level summaries instead of storing identifiable frames.
4. Implementation of privacy and data-minimization controls
   Pseudonymous identifiers (S001-S332 for students and F001-F015 for faculty) were assigned, with the ID key file maintained on an encrypted server under role-based access control. Local video caching was disabled. When temporary caching was required for quality assurance, cached frames were deleted within 24 h after verification. Per-session detection accuracy, average latency, and error rates were recorded for later presentation in the Results section.
5. Verification of AI readiness (procedural checkpoint)
   Stable face detection at 25-30 fps, receipt of HTTP 200 OK responses, valid JSON fields, documented threshold settings, and secure credential storage were confirmed. The entire pipeline was reported as having passed readiness checks consistent with the calibration depiction in Figure 1B.

3. Survey administration

1. Configuration of instruments and anchors
   A 5-point Likert scale was employed to measure depression, anxiety, and stress, using items adapted from DASS-21 to align with the academic context. Item codes and anchors followed the instrument layout presented in Table 3.
2. Administration of T1 (pre-test)
   The survey was administered immediately after the initial exam and prior to any AI-informed feedback. A secure online form was deployed with all fields set as mandatory and duplicate submissions prevented. The dataset was exported in CSV format with fields for participant ID, timepoint (T1), and item scores.
3. Delivery of brief instructor feedback
   Within 24 h after T1, non-clinical guidance was provided to students identified by AI. Feedback included short relaxation exercises (3-5 min), study-plan adjustments, or time-management suggestions. No clinical notes were recorded in the research dataset. Students exhibiting severe distress were referred to campus counseling services in accordance with institutional policy.
4. Administration of T2 (post-test)
   The same survey was re-administered after the second exam session and subsequent feedback. Data were exported in CSV format with timepoint (T2) and matched to T1 responses using participant IDs.
5. Locking of survey data (procedural checkpoint)
   Response rates and missing data patterns were verified, and the survey database was locked before data merging. Descriptive statistics were planned for inclusion in the Results section.

4. Narrative simulation exercises

1. Execution of Simulation 1
   An academic-stress scenario was presented, depicting situations such as receiving a low grade or perceiving unfair feedback. Each student was instructed to select one of three response options: internalizing frustration, publicly blaming the instructor, or reflecting and seeking clarification. A duration of 3-5 min was allocated for response selection, followed by a written rationale of 150-200 words. The simulation interface displaying the timing prompt and response options is shown in Figure 1C.
2. Conduction of short interviews
   Immediately after Simulation 1, short interviews lasting 5-7 min were conducted to explore participants' appraisals, coping intentions, and perceived emotional triggers. Interview notes were recorded and linked to participant IDs.
3. Execution of Simulation 2
   A second simulation was conducted several days later, following the same procedure as Simulation 1 but using a parallel scenario. This session aimed to observe changes in emotional awareness and coping behavior. The choice, rationale, and interview notes were recorded as in Simulation 1.
4. Confirmation of simulation records (procedural checkpoint)
   Each participant was verified to have one recorded choice, one rationale containing at least 150 words, and one corresponding interview note. A coding summary template was prepared for subsequent thematic analysis.

5. Exporting AI logs and merging datasets

1. Export of session-level emotion summaries
   Session-level emotion summaries were exported in CSV format, including participant ID, session ID, ISO 8601 timestamp, emotion label, confidence value, and the stress-flag indicator. Raw video frames were not saved.
2. Merging of AI logs with survey files
   AI log files were merged with T1 and T2 survey datasets using participant ID and session date as key variables. Derived variables, such as ΔAnxiety = Anxiety_T2 − Anxiety_T1, were computed to assess change over time. An audit of 5% of merged records was performed to verify alignment of identifiers and chronological consistency.
3. Confirmation of merge integrity (procedural checkpoint)
   One-to-one ID matching and correct temporal ordering, with T1 preceding T2 for each participant, were confirmed. Any identified discrepancies were resolved prior to conducting statistical analysis.

6. Quantitative analysis (SPSS workflow)

1. Import of data and definition of variables
   Data files were opened in SPSS 27.0, and CSV datasets were imported. Variable labels and types were defined, and missing-value rules were established. Master files were saved in .sav format for subsequent analyses.
2. Assessment of reliability and descriptive statistics
   Cronbach's α was computed for each construct to assess internal consistency. Descriptive statistics, including mean (M) and standard deviation (SD), were generated for all variables. The resulting outputs were retained for inclusion in the Results section.
3. Execution of paired-samples t-tests
   Paired-samples t-tests were conducted to compare T1 and T2 scores for anxiety, depression, and stress. Results included t, degrees of freedom (df), p-value, Cohen's d, and 95% confidence interval (CI), all of which were presented in the Results section.
4. Execution of repeated-measures ANOVA
   Time (T1, T2) was specified as a within-subject factor in a repeated-measures ANOVA. Mauchly's test of sphericity was evaluated, and the Greenhouse-Geisser correction was applied when assumptions were violated. The analysis reported F, df, p, η², and 95% CI values for inclusion in the Results.
5. Fitting of multiple linear regression models
   Multiple linear regression models were fitted with T2 outcomes regressed on T1 scores, AI stress-flag rate, and narrative choice (dummy-coded). Model results were summarized by reporting β, standard error (SE), t-value, p-value, and R².
6. Maintenance of unit and notation standards
   All numerical values adhered to SI unit conventions, with a space between number and unit (e.g., 10 min, 25 fps, 640 × 480 px). Equations were inserted using Word's Equation tool where appropriate, and statistical notation was kept consistent throughout all sections.

7. Thematic analysis workflow

1. Assembly of the qualitative corpus
   Interview notes were transcribed, and narrative rationales were compiled into a single corpus labeled by participant ID and timepoint (Sim1 or Sim2). This corpus served as the dataset for qualitative analysis.
2. Development of the codebook and training of coders
   A codebook was developed focusing on emotional engagement, decision-making quality, and emotional-intelligence indicators, including self-awareness, empathy, and emotional regulation. Two coders were trained on 10% of the corpus to ensure consistent interpretation of codes and analytic criteria.
3. Establishment of interrater reliability and completion of coding
   Interrater reliability was assessed with a target Cohen's κ of at least 0.75. Any discrepancies were discussed and resolved before finalizing the codebook. The remaining corpus was then fully coded. Theme frequencies were summarized, and a representative thematic map was prepared for presentation in the Results section. One representative quotation was selected to illustrate each core theme.

8. Quality control, troubleshooting, and archiving

1. Resolution of technical issues
   When face-detection rates were low, front lighting was improved and the camera angle adjusted. Frame rates were verified to remain within 25-30 fps. In cases where excessive stress flags were observed, the detection threshold was increased to 0.80 or the detection window was extended to at least 8 s to stabilize results.
2. Assurance of data consistency
   Participant IDs and timepoints were validated before data analysis. Coders were recalibrated after every 50 transcripts to prevent coding drift. A session-level quality-control checklist was maintained and signed by both the session lead and the data steward to document procedural integrity.
3. Archiving of datasets and documentation
   All analytic files were de-identified by replacing direct identifiers with participant IDs. Consent forms were stored separately from research data. Final datasets, SPSS syntax files, codebooks, and quality-control logs were archived on a secure server under restricted access.

Results

Descriptive analysis

Pre-post comparisons indicated consistent declines in adverse affect following implementation. Post-test means decreased to 2.87 for anxiety, 2.74 for depression, and 2.69 for stress, from pre-test means of 3.25, 3.11, and 3.08, respectively (Figure 2). Descriptive outcomes are summarized in Table 4, and the item-level instrument structure is provided in Table 3. To evaluate AI reliability wit...

Discussion

The study demonstrates that when narrative simulation is combined with visual AI, emotional awareness becomes a shared process rather than a one-directional intervention. Students were not merely responding to a digital tool; they were engaging in a structured emotional reflection that connected cognitive appraisal with embodied experience. Across the dataset, both quantitative and qualitative results revealed a consistent pattern of reduced anxiety and improved emotional regulation, suggesting that the protocol can func...

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
Hardware	Facial Recognition Camera	Logitech (or similar)	C920 HD Webcam
Instrument	Interview Transcript Collection	Manual entry	Audio recordings/text
Questionnaire	Modified DASS-21 Scale	Adapted by authors	5-point Likert format
Software	Microsoft Azure Emotion API	Microsoft	Emotion API v3 (Cloud)
Software	SPSS Statistics	IBM	Version 26
Software	Narrative Simulation System	Self-developed	Web-based (HTML5)
System Platform	University Teaching Management System	Internal platform	Integrated with AI module