This study employs stringent inclusion criteria to ensure the homogeneity of the sample and the scientific validity of the research findings. Specifically, the study targets full-time undergraduate students majoring in clinical medicine, aiming to ensure a consistent professional background and avoid the influence of different majors (such as nursing, pharmacy, etc.) on learning needs and cognitive characteristics. The age range is set between 17 and 19 years (average age 18.13 ± 0.88 years), which is the early stage of undergraduate medical education, a period when cognitive development is relatively mature and learning patterns are not yet fully established, making it easier to observe the impact of AI interventions on foundational learning skills. Individuals with severe learning disabilities (such as dyslexia, ADHD) or a history of mental illness (such as depression, anxiety requiring medication) are excluded to minimize potential confounding factors that could affect learning behavior and outcome assessment. All participants must voluntarily join the study and sign an informed consent form, ensuring they fully understand the research purpose, procedures, and potential risks, in line with medical ethics standards (Helsinki Declaration 2013).

To exclude potential bias from prior exposure to similar platforms, all participants completed a pre-enrollment technology usage survey. Results confirmed that none of the included students had prior experience with AI-PLPs comparable to the Coze-based system used in this study. This ensured that observed effects were attributable to the intervention itself rather than pre-existing familiarity.

This study included 40 eligible medical undergraduate students, using a strict randomization strategy to ensure balanced and comparable baseline characteristics between groups. Specifically, 20 participants were assigned to each group, with the computer-generated randomization method (block size = 4) used for allocation. The baseline academic performance data showed no statistically significant difference in pre-test knowledge reserves between the two groups (Experimental Group: 70.40 ± 8.96 points vs. Control Group: 70.20 ± 11.40 points, p = 0.950). In terms of demographic characteristics, the gender distribution was balanced (Experimental Group: male:female = 12:8; Control Group: 11:9, χ² = 0.06, p = 0.812), and age indicators also showed high consistency (Experimental group 18.10 ± 0.97 years vs. Control group 18.15 ± 0.81 years, t(38) = 0.36, p = 0.724). All participants voluntarily participated and signed informed consent forms. The complete baseline data are detailed in Table 1.

To ensure the experimental and control groups are balanced in key covariates, this study employs a stratified randomization design. Gender (male/female) and admission scores (top 50% vs. bottom 50%) are used as stratification factors.

This study employs a block randomization method (block size = 4) to ensure balance between groups. The generated random allocation sequence was sealed in opaque envelopes and managed by an independent third-party researcher.

The randomization resulted in perfectly balanced groups for gender and baseline scores (Table 1). No participants dropped out during the study.

The control group followed the traditional lecture-based model: 4 h/week of teacher-centered instruction using standardized textbooks (e.g., Systematic Anatomy). Learning reinforcement included weekly quizzes (multiple-choice, short-answer) graded uniformly by the teaching office. No digital tools or personalized feedback were used.

The experimental group used the Coze-based AI Personalized Learning Platform (AI-PLP) alongside 4 h/week of traditional instruction. The platform provided four core functions:

Dynamic Learning Path Optimization: Adjusted content difficulty/sequence every 48 h based on diagnostic tests (e.g., adding circulatory system micro-lessons if weaknesses detected).

Affective Computing Support: Used NLP to detect frustration (e.g., from interaction patterns) and triggered motivational messages.

Intelligent Resource Recommendation: Recommended personalized resources (e.g., animations, guidelines) from a 2,800-case database.

Immersive Clinical Simulation: Provided VR-based case training with AI mentor feedback.

Data was collected at:

Baseline (Week 0): Demographics, pre-test scores, learning behavior.

Intervention Period (Weeks 4, 8, 12): Platform logs, diagnostic tests, classroom recordings, engagement metrics.

Endpoint (Week 12): Post-test, satisfaction survey, motivation scales.

This study employs the standardized test bank of the Accreditation Council for Medical Education (LCME) to assess learning effectiveness. The tool includes three parallel test sets (A/B/C), covers Bloom’s taxonomy levels, and demonstrates high reliability (α = 0.89) and validity (CVI = 0.91). Scoring used IRT calibration and double-blind marking.

An adapted SERVQUAL scale assessed satisfaction across five dimensions (Table 2). The 20-item scale uses a 5-point Likert scale and showed excellent reliability (α = 0.84) and structural validity.

A comprehensive framework combined:

Objective Metrics: Automatically logged by the platform (study duration, resource downloads, simulation participation).

Subjective Metrics: Revised Schraw Scale (20 items) assessing metacognition, motivation regulation, digital resource use, and collaboration on a 6-point scale (CR = 0.91, AVE = 0.53).

Audio recordings of sessions were transcribed and analyzed using NVivo 12. Metrics included question/comment frequency and the proportion of contributions coded as “in-depth discussion” (involving analysis, evaluation, or synthesis), with high inter-coder agreement (Kappa = 0.85).

Missing Data (≤10%): Handled using Multiple Imputation by Chained Equations (MICE), generating 5 datasets (convergence: Gelman-Rubin < 1.01), pooled via Rubin’s rules.

Outliers: Identified via boxplot (IQR = 1.5) and validated through expert review (3 professors, k = 0.88), platform log checks, and student interviews. Contextually valid extremes (e.g., exam prep spikes) were retained. As shown in Figure 1.

Descriptive statistics: Means ± SD for continuous variables; frequencies (%) for categorical variables.

Group comparisons: Independent samples t-tests on primary outcomes (performance, satisfaction, study time, engagement). Bonferroni correction applied (α_adjusted = 0.0125 for 4 outcomes).

Effect sizes: Cohen’s d calculated for group differences (d ≥ 0.2 small, ≥ 0.5 medium, ≥ 0.8 large). As shown in Figure 2.

Correlations: Pearson’s r used (| r| ≥ 0.3 weak, ≥ 0.5 moderate, ≥ 0.7 strong).

ANCOVA adjusted for baseline scores (F(1,37) = 0.82, p = 0.371; Adjusted group difference remained significant: β = 2.61, p = 0.028).

Post hoc power analysis (G*Power 3.1, α = 0.05, d = 0.72): Power = 0.86 (>0.80 threshold).

Multiple Imputation vs. Complete Case effect size comparison (d_MI = 0.70 vs. d_Complete = 0.72) showed minimal bias. As shown in Figure 3.

The experimental group spent significantly more time on autonomous learning daily compared to the control group (49.25 ± 18.59 min vs. 34.80 ± 18.32 min; t = 2.042, p = 0.048). The effect size, Cohen’s d = 0.78 (95% CI [0.35, 28.55]), indicates that the difference is of moderate to large practical significance (>0.5 threshold). This finding confirms that the AI platform effectively extended students’ effective learning time through dynamic path optimization (Figure 7).

This study systematically evaluated the enhancement effect of an AI platform on classroom engagement behaviors through a combination of quantitative metrics and qualitative analysis. As shown in Figure 5. The experimental group demonstrated significant advantages in both the quality and depth of classroom participation: In terms of question frequency, the experimental group averaged 16.05 ± 3.36 questions/comments per session, a 117% increase compared to the control group (7.40 ± 3.57 times) (t = 7.89, p = 0.026, Cohen’s d = 2.46), indicating that AI intervention significantly stimulated students’ proactive thinking and willingness to engage in classroom interactions. Regarding discussion depth, the experimental group exhibited a 58% proportion of in-depth discussions (involving higher-order cognitive activities such as pathological mechanism analysis and treatment plan optimization), significantly higher than the control group’s 32% (Figure 8). NVivo 12 coding analysis revealed that keywords such as “evidence-based medicine” and “multidisciplinary integration” appeared 2.3 times more frequently in the experimental group’s discussions (p = 0.008), confirming that the AI platform’s clinical case simulation training (see section “2.3.2 Allocation method”) effectively promoted the development of students’ clinical reasoning and critical thinking skills.

Effect size analysis further elucidated the intervention intensity: The Cohen’s d for question frequency was 2.46 (>0.8), indicating an extremely large effect size; the between-group difference in discussion depth reached 26 percentage points (58% vs. 32%), demonstrating clear clinical educational significance. This “dual enhancement in quantity and quality” characteristic closely aligned with the transformation in classroom interaction patterns depicted in Table 6—students in the experimental group showed significantly higher engagement in dimensions such as knowledge application and argumentation (p < 0.01), forming a virtuous cycle of “high-frequency interaction and deep critical thinking.

This study employs dual validation through quantitative analysis and behavioral visualization to reveal the optimization effect of AI platforms on learning resource utilization: the experimental group demonstrated significantly higher literature reading volume compared to the control group (25.95 ± 7.01 papers vs. 17.50 ± 7.64 papers, t = 2.82, p = 0.008, Cohen’s d = 1.14), with targeted reading (literature directly related to current learning objectives) accounting for 83% (versus only 57% in the control group), confirming that the AI precision recommendation algorithm (see section “2.4.1 Control group”) significantly enhances resource acquisition efficiency. Platform log analysis further indicates that the experimental group increased average daily intensive reading time by 53 min (p < 0.001) and boosted literature note generation by 2.1 times (p = 0.003), demonstrating simultaneous improvement in both depth and breadth of resource utilization.

Visualization Evidence of Behavioral Patterns Shows Strong Correlation with Resource Utilization:

Figure 9 classroom discussion depth radar chart reveals that the experimental group significantly outperformed the control group in knowledge application (p = 0.012) and critical thinking (p = 0.004), confirming the facilitating effect of high-quality literature input on higher-order thinking.

Figure 10 learning behavior heatmap demonstrates the experimental group’s unique “trinity” learning model:

High questioning frequency (16.05 ± 3.36 times/class, ↑117% compared to the control group).

In-depth literature reading (25.95 articles/cycle, targeted reading ↑46%).

Sustained learning duration (49.25 min/day, ↑42%).

The Pearson correlation coefficient among these factors is r = 0.62 (p < 0.001), indicating a significant positive feedback loop between resource utilization efficiency and deep learning behaviors.

Mechanism Analysis: The AI platform dynamically generates personalized recommendation lists (matching rate > 90%) by tracking real-time learning behavior data (e.g., knowledge mastery, literature reading speed), improving the experimental group’s resource acquisition accuracy by 46% (p < 0.001). This closed-loop mechanism of “algorithm-driven, precision acquisition, and deep utilization (Area A in Figure 10’s heatmap) directly promotes knowledge internalization and cognitive leaps, providing a replicable digital solution for optimizing medical education resources.

Our findings must be contextualized within three constraints:

Learner heterogeneity: Effects were strongest for foundational knowledge (d = 0.92) versus clinical judgment (d = 0.47), echoing Sapici’s (34) concern about AI’s limitations in complex skill development.

Temporal decay: Skill retention dropped 22% at 12-week follow-up, necessitating longitudinal studies with booster interventions.

Algorithmic transparency: Unexplained path adjustments (15% of cases) may undermine trust – future work should integrate SHAP-value visualizations.

Ecological tradeoffs: While lab-controlled trials [e.g., Zhou et al. (14)] show lower effects, our real-world implementation achieved higher ecological validity at the cost of uncontrolled confounders.

Measurement gaps: Neural data (n = 10) lacked power to detect amygdala-prefrontal connectivity changes – a critical pathway for sustained engagement.