A cross-sectional survey was conducted between March and April 2026 in Sichuan Province and Chongqing Municipality, in accordance with the Declaration of Helsinki and approved by the Institutional Ethics Committee (NSMC Ethics Project Review [2025] No.040). The study population comprised three cohorts: employer (demand side), undergraduate preventive medicine and public health (supply side), and faculty/administrators (training side). Data were collected digitally via the Wenjuanxing platform (https://www.wjx.cn) from 7 universities and 141 employer organizations (including CDCs, hospitals, health centers, government departments, research institutes, and medical enterprises).

After obtaining electronic informed consent from all participants, eligible respondents met predefined inclusion criteria (belonging to one of the three target populations and able to complete the survey). Exclusion criteria included implausibly short completion times (<120 s), logical contradictions, and patterned responses. Of 1,058 initial responses, 27 were excluded, yielding a final effective sample of 1,031 participants (150 employers, 680 students, 201 educators), with an effective response rate of 97.4%.

Based on literature review and expert consultation, initial item pools were generated and refined through two rounds of pilot testing (cognitive interviews, n = 11; small-sample pilot, n = 45). Items with low relevance, unclear wording, or substantial overlap were deleted, resulting in final instruments: Employer (25 items), Student (20 items), and Teacher (22 items), all using a 5-point Likert scale.

The Employer Demand Scale (25 items, total α = 0.920) covered five dimensions: Talent Demand Intensity, Current Status Assessment, Perceived Competency Deficits, Training Expectations, and Industry-Education Cooperation Willingness.

The Student Self-Assessment Scale (20 items, total α = 0.901) covered five dimensions: Cognition/Attitude, Basic Knowledge, Practical Application Skills, Innovation/Collaboration, and Continuous Learning.

The Teacher Competency Scale (22 items, total α = 0.874) covered five dimensions: Curriculum System, Teacher Competency/Innovation Mechanisms, Teaching Resources, Student Ability Evaluation, and Future Strengthening Directions.

Confirmatory Factor Analysis confirmed good psychometric properties, with standardized factor loadings ranging from 0.647 to 0.860, McDonald’s ω from 0.778 to 0.901, and corrected item-total correlations >0.30.

All analyses were performed using R version 4.4.0, with a two-sided significance level of 0.05.

Supplementary File 1 presents the baseline demographic and professional characteristics of the 1,031 stakeholders included in the final analysis.

On the demand side, the 150 employer representatives were drawn from a broad spectrum of public health sectors, notably hospitals (19.3%), CDCs (16.0%), and research institutes (16.0%). Most employer respondents served as technical or operations heads (64.7%) at municipal (38.7%) or district/county (22.0%) levels. Furthermore, a significant proportion of these organizations had already initiated AI technology integration, either at the initial trial (44.0%) or partial implementation (32.7%) stages.

Among the 680 students, the gender distribution was balanced (51.8% male, 48.2% female). Participants spanned various academic stages, with a robust representation across core public health disciplines, including occupational and environmental health (24.3%), epidemiology and biostatistics (20.7%), and nutrition and food hygiene (20.4%). Notably, a substantial majority of students (75.7%) reported having prior AI-related training or coursework.

For the training supply side, the educator cohort (n = 201) was predominantly composed of academic staff with 89.6% holding either a master’s or doctoral degree. The respondents represented diverse academic ranks and teaching experiences, and 85.1% indicated prior involvement in AI-related teaching or curriculum delivery.

Initial descriptive statistics indicated that overall evaluations across the three cohorts—employers, students, and educators—were moderately high. Normality was confirmed for all continuous variables (|skewness| < 2, |kurtosis| < 7). Specifically, employers reported the highest scores regarding the “perceived competency deficits” of current students (3.75 ± 0.88). In contrast, students self-evaluated their “practical application skills” most favorably (3.73 ± 0.97), whereas educators expressed the highest confidence in the current “teaching resource supply” (3.88 ± 0.85).

To pinpoint the structural loci of these educational discrepancies, an Importance-Performance Analysis (IPA) matrix was constructed, with empirical grand means established as the cross-origin thresholds following standard IPA practice (Figure 2). Robustness was assessed via bootstrap resampling and sensitivity analyses using alternative thresholds, which confirmed the stability of the core directional pattern. As detailed in Table 1, the data revealed perceptual misalignments between employer priorities and student self-assessed competencies. Most critically, “epidemic trend prediction and risk assessment” emerged as the highest-rated employer priority (3.84 ± 1.06), yet students’ self-reported performance in this domain was comparatively lower (3.72 ± 1.17), placing this competency in the high-priority “Concentrate Here” quadrant (Quadrant II).

Conversely, while students rated their “tool customization and model construction” capacity relatively highly (3.77 ± 1.22), direct demand for this specific technical skill from employers was comparatively moderate (3.65 ± 1.14), mapping it to Quadrant IV. All employer-student differences were small in magnitude; however, bootstrap analysis indicated that the directional pattern—particularly the higher priority assigned to epidemic prediction by employers—remained stable, supporting interpretation of these findings as relative priority differences rather than absolute gaps.

Given the aggregate performance deficits identified in the IPA, an individual-centered Latent Profile Analysis (LPA) was subsequently conducted to decompose the unobserved heterogeneity within students’ practical AI application skills (Items C1–C6). After systematically comparing the goodness-of-fit indices for models specifying one to four latent classes, the 2-class model was identified as the optimal parsimonious solution (Supplementary File 2). This configuration demonstrated the lowest Bayesian Information Criterion (8679.390) and sample-size adjusted BIC (8619.063), alongside robust classification accuracy (Entropy = 0.999).

The 2-class model was robust to alternative specifications, including ordinal latent class analysis and VVI covariance, and outperformed a one-factor CFA model (ΔBIC = 2971.7), supporting the categorical distinction over a single continuous trait.

Based on the conditional mean distributions extracted from the optimal model (Figure 3), the digital skill proficiency among the student cohort exhibited distinct polarization, yielding two distinct subpopulations. The High-Order Application Group (Class 1, n = 158, 23.2%) maintained consistently high performance across all six skill indicators (mean range: 4.18–4.30), demonstrating comprehensive readiness to integrate AI into public health scenarios. Conversely, the Foundation-Weak Group (Class 2, n = 522, 76.8%) scored substantially lower across all dimensions (mean range: 1.96–2.21). This indicates that the digital literacy of the vast majority of current public health students appears limited, lacking the substantive technological capacity required by hiring organizations.

To evaluate how the aforementioned competency deficits influence external stakeholders, a multiple linear regression model was formulated to ascertain factors associated with employers’ willingness to participate in collaborative academic-practice training. The overall model yielded a statistically significant fit (F = 18.89, p < 0.001, Adjusted R² = 0.194) (Table 2).

To account for potential confounders, the employer regression model was expanded to include organization type, organization size, AI adoption stage, and respondent role as covariates. After controlling for these variables, Demand_X1 (B = 0.363, p < 0.001) and Deficit_X2 (B = 0.211, p = 0.023) remained significant positive predictors of cooperation willingness, confirming the robustness of the core findings. Among covariates, only government departments showed significantly lower willingness compared to CDC (B = −0.589, p = 0.027). Model R² = 0.303 (Adjusted R² = 0.219).

The regression parameters indicated that demand-side perceptions of educational inadequacies were positively associated with cooperation willingness. Specifically, both a heightened “talent demand intensity” (β = 0.327, p < 0.001) and a higher perception of current students’ “competency deficits” (β = 0.235, p = 0.009) served as significant positive predictors. These findings suggest that overarching talent scarcity and the explicit observation of skill mismatches are associated with greater willingness to proactively intervene in the university pedagogical process.

Internally, the successful reform of AI curricula relies heavily on faculty capabilities. A moderation model was subsequently constructed to test the boundary conditions governing the translation of educators’ “personal classroom AI application ability” into their “ability to mentor student research projects,” utilizing “institutional innovation mechanisms” as the moderating variable (Supplementary File 3).

To address potential confounding in the teacher model, academic rank, years of teaching, and prior AI training were added as covariates. The interaction term (classroom ability × innovation mechanism) remained significant after adjustment (B = −0.125, p = 0.013), consistent with the unadjusted model. All covariates were non-significant (p > 0.05). Model R² = 0.418 (Adjusted R² = 0.388).

Regression diagnostics for the teacher moderation model showed all VIF values below 2, indicating no multicollinearity. The Breusch-Pagan test was significant (p < 0.001), suggesting mild heteroscedasticity; however, ordinary least squares estimates are robust to moderate violations. All Cook’s distance values were below 1 (maximum = 0.175), confirming no influential outliers. Residual analysis supported the overall model assumptions. No severe violations of regression assumptions were detected.

While the main effect remained significantly positive, the core interaction term (Classroom Capacity × Innovation Mechanism) yielded a significantly negative coefficient (β = −0.120, t = −2.468, p = 0.014). To elucidate the trajectory of this negative moderation, a simple slope analysis was executed (Figure 4A). Simple slope analysis revealed that the positive association between classroom AI application ability and research mentorship was significant at low (slope = 0.408, p < 0.001) and mean (slope = 0.285, p < 0.001) levels of institutional innovation, but became non-significant at high levels (slope = 0.162, p = 0.101). Johnson-Neyman analysis further indicated that the simple slope remained significant across the entire observed moderator range, as the transition point (W = 2.378) fell outside the observed range (−2.826 to 1.174) (Figure 4B). Ceiling effects were ruled out as an alternative explanation, as only 32.1% of teachers in the high innovation group scored the maximum on research mentorship ability (mean = 4.21, SD = 0.62).

The visualization revealed a diminishing association: when institutional innovation support were relatively weak (Mean - 1 SD), educators’ personal classroom capabilities efficiently translated into research mentorship abilities (indicated by a steep, positive slope). Conversely, under high levels of institutional innovation support (Mean +1 SD), the translation efficiency flattened significantly.

This study acknowledges certain limitations. Constrained by the cross-sectional design, strict causal inferences cannot be established, despite the deployment of regression and moderation models. Second, the assessment of student skills and faculty capabilities relied primarily on self-reported scales, which may be susceptible to social desirability bias. Third, the analyses were based on aggregated data from multiple institutions rather than matched university-employer-student dyads. Future research may benefit from designs that link specific universities, employers, curricula, and students, or from longitudinal designs that examine whether collaboration and institutional support predict objectively measured competency gains.

Based on empirical evidence from Sichuan and Chongqing, Western China, this study suggests that the sustainable integration of AI into public health may be associated with a structural misalignment between academic training emphasizing computational techniques and frontline demands centered on public health problem-solving. Furthermore, LPA identified polarization, with 76.8% of students in the Foundation-Weak group and 23.2% in the High-Order Application group, suggesting a need for precision-stratified pedagogy. Rigidly structured administrative frameworks may be associated with lower educators’ effectiveness in translating their AI proficiency into practical research guidance. To address these competency gaps in these regions, universities may consider implementing real-world data-driven project-based learning and further refine faculty evaluation metrics. This multi-stakeholder assessment, guided by the conceptual framework, highlights the value of integrating demand, supply, and training perspectives in examining AI workforce education. Given the cross-sectional nature of this study, all associations reported should be interpreted as correlational, not causal. Further research in other geographic and institutional contexts is needed to assess the generalizability of these findings.