The field of psychometrics faces a fundamental challenge rooted in the philosophical nature of mental states: the inherent tension between stable, long-term traits and fluctuating, moment-to-moment psychological states (1). Traditional psychometric assessments, while valuable, often provide a static “snapshot” of an individual’s mental health, which may not adequately capture the dynamic, real-time fluctuations of affective states such as anxiety and depression. This is particularly critical for university students, a demographic navigating a period of significant developmental and academic pressure. Recent large-scale epidemiological studies in China have highlighted the severity of this issue, with detection rates for depression and anxiety among university students reported at 9.8% and 15.5%, respectively (2). The mental health landscape for medical students can be even more demanding, characterized by a rigorous curriculum and high-stakes examinations, necessitating more sensitive and ecologically valid assessment methods (3). Consequently, accurately capturing the ebb and flow of state anxiety and state depression remains a critical research topic.

Within psychological research, measurement is broadly approached through observational, experimental, and psychometric testing methods (4). While observational and experimental methods offer high validity, their resource-intensive nature limits scalability. Psychometric testing, conversely, offers convenience and has been widely adopted. However, this convenience often comes at the cost of ecological validity, as a single questionnaire provides limited insight into a student’s psychological journey. This limitation reveals a significant gap: the need for a measurement paradigm that integrates the scalability of psychometric tests with the contextual richness of observational methods. To truly understand and potentially “correct” a student’s self-reported score, it is essential to look beyond the score itself and examine the underlying cognitive and behavioral patterns that drive these emotional states.

This perspective aligns with the principles of Cognitive-Behavioral Theory (CBT), which posits that emotions are shaped by an interplay of thoughts and behaviors. The continuous advancement of Information and Communication Technology (ICT) and Artificial Intelligence (AI) now offers an unprecedented opportunity to operationalize this holistic view at scale. The field is already witnessing significant progress through innovations like Computerized Adaptive Testing (CAT) (5, 6) and Automatic Item Generation (AIG) (7). The ubiquity of smartphones has further enabled the collection of rich, longitudinal data via social networks, allowing for the creation of dynamic user profiles that integrate network theory with psychological process models (8).

More recently, the emergence of powerful AI, particularly large language models like ChatGPT, is poised to act as a “human proxy” in interpreting complex behavioral data, heralding a quiet revolution in psychometrics (9, 10). The trajectory of psychological measurement is inevitably shifting towards capturing multimodal data from daily life via wearable devices—including facial expressions, voice patterns, and even brain imaging scans—long before Brain-Computer Interfaces (BCIs) might achieve true “mind-reading” (11). Crucially, this technological evolution is blurring the boundary between mental health assessment and clinical intervention (12). Modern intelligent applications are evolving from passive monitors into “digital health promoters,” capable of providing data-driven, adaptive interventions based on real-time psychometric insights (13, 14). These systems create a bidirectional feedback loop, where assessment informs personalized intervention, and the resulting behavioral changes continuously update the assessment itself. Such detailed personal profiles not only guide psychological interventions (15) but also serve as core indicators for evaluating their effectiveness and tracking mental health status over time (16). This technologically-empowered, integrated paradigm forms the foundation for the novel assessment framework proposed in this study.

Against the backdrop of the inherent challenges in traditional psychometrics and the opportunities afforded by technological innovation, this research aims to pioneer and implement a novel pathway for university student mental health assessment. Utilizing a WeChat mini-program platform, we deeply integrate information technology with psychological theory. Drawing inspiration from the principles of Ecological Momentary Assessment (EMA), which emphasize capturing experiences in real-time, real-world contexts (17), we constructed a dynamic, interactive, and continuous mental health assessment ecosystem.

The core objectives of this study were threefold. First, we aimed to evaluate the accuracy of an AI-driven dynamic assessment model, testing whether integrating daily cognitive and behavioral data could correct and enhance the performance of traditional static mental health scales. Second, we sought to explore the potential intervention effect of the system, investigating whether continuous engagement and feedback could lead to measurable improvements in students’ clinically-rated mental health. Third, we intended to assess the feasibility and user engagement of this approach, examining its practicality and acceptance within a real-world university setting.

To achieve these objectives, our conceptual strategy was to create a multidimensional, multimodal mental health assessment framework. This framework creatively integrates three distinct data streams: (1) traditional, standardized psychometric scales for baseline assessment; (2) motivation-driven “cognitive voting” to capture students’ thought tendencies and emotional responses to daily events (18); and (3) systematic “behavioral check-ins” for the idiographic measurement of actual behavioral patterns (19). By continuously collecting and integrating these data points, our system aims to construct dynamic, real-time mental health profiles for each student.

Ultimately, the goal of this study extends beyond passive assessment. By providing immediate feedback and gamifying personality attributes, we aim to actively engage students in their own mental health management. This process is designed to foster self-awareness and encourage positive behavioral or cognitive adjustments, thereby creating an integrated assessment and intervention loop. In essence, this research endeavors to develop and validate an intelligent and personalized tool that is more attuned to the contemporary needs of university mental health services, achieving greater immediacy, comprehensiveness, and personalization (20).

This study employed a longitudinal design to dynamically assess the psychological states of 177 undergraduate clinical medicine students over one and a half years. The research was conducted via a custom-developed WeChat mini-program, which served as the primary platform for data collection and user interaction. Our methodological framework integrated three core data streams: (1) periodic, systematic self-report scale screenings conducted at the beginning of each semester; (2) high-frequency ecological cognitive voting, comprising 87 instances distributed throughout the study period; and (3) event-driven behavioral check-ins, with 66 distinct instances recorded over the same period.

The core self-report scales included the Self-Rating Anxiety Scale (SAS), the Self-Rating Depression Scale (SDS), the Symptom Checklist-90 (SCL-90), the Eysenck Personality Questionnaire (EPQ), the Sixteen Personality Factor Questionnaire (16PF), and the Myers-Briggs Type Indicator (MBTI). Cognitive voting was designed as a method for students to express opinions and emotional attitudes regarding their campus life and current events through multiple-choice questions. Behavioral check-ins involved the systematic recording of specific events each student encountered, such as academic achievements, social conflicts, or personal milestones. These events were defined and logged by four designated “life teachers” (counselors), providing a semi-objective measure of students’ real-world experiences.

A battery of standardized psychometric scales was administered to collect baseline and periodic data on students’ mental health status. All selected instruments have been validated for use within the Chinese cultural context. The internal consistency for these scales, as reported in validation studies with Chinese university students or related populations, is detailed below.

Self-rating anxiety scale (SAS). The SAS is a 20-item self-report questionnaire used to measure the subjective severity of anxiety symptoms. The Chinese version of the SAS, introduced and validated by Wang (1984), has demonstrated good psychometric properties in Chinese populations (21). Subsequent validation studies among university students have reported high internal consistency, with Cronbach’s alpha coefficients typically around 0.88 (22).

This study employed a longitudinal design combining principles of psychometric testing, Ecological Momentary Assessment (EMA), and digital intervention. Over a one-and-a-half-year period, we conducted a dynamic assessment of psychological states involving 177 undergraduate clinical medicine students. The entire study was facilitated through a custom-developed WeChat mini-program, which served as the primary interface for data collection, feedback delivery, and student engagement. The research procedure involved an initial baseline assessment using standardized scales at the beginning of the first semester, followed by continuous data collection through weekly cognitive voting and event-contingent behavioral check-ins, alongside periodic reassessments with the full scale battery at the start of subsequent semesters.

To derive both quantitative and qualitative insights from the rich, multimodal data, we developed a novel conceptual framework to structure our assessment. This framework was designed to bridge the gap between high-level psychometric scores and the granular, real-world data collected daily. Drawing inspiration from psychometric modeling approaches such as the bifactor model (31), testlet models (32), and second-order factor models (33), as well as the time-series nature of Vector Autoregression (VAR) models (34), we established an intermediate hierarchical structure. This collaborative effort, involving four experienced university life teachers and three clinical psychologists, resulted in a model that could systematically capture and interpret the psychological and behavioral patterns reflected in students’ daily lives.

The structure consists of six primary factors: Anxiety Cognition, Anxiety Behavior, Depression Cognition, Depression Behavior, Personality Cognition, and Personality Behavior. Each primary factor was further subdivided into four sub-dimensions (e.g., negative events, developmental concerns, life discipline), resulting in a total of 24 cognitive and behavioral dimensions. The complete framework is detailed in Table 1.

This framework served two key purposes. First, it provided a structured schema to ensure that every designed cognitive vote and behavioral check-in could be systematically mapped to one or more specific dimensions. At the outset of the study, we also mapped all items from the psychometric tests to these 24 dimensions. Specifically, the 130 items from SAS, SDS, and SCL-90 were linked to the framework, primarily for weighting their original scores during the assessment process. The 328 items from EPQ, 16PF, and MBTI were also linked, serving as a reference for personality assessment and allowing students to observe continuous changes in their profiles. This entire mapping process (visualized in Figures 1–4, which are detailed later) created a comprehensive, interconnected data ecosystem.

Second, the framework acted as a reference for the dynamic adjustment algorithm. For instance, if a student repeatedly selected pessimistic options in cognitive voting, their scores on the SDS and related SCL-90 dimensions would be algorithmically adjusted. Similarly, demonstrated self-discipline in behavioral check-ins would influence their personality trait profiles derived from the MBTI and 16PF, which are understood to be shaped by cognitive development and cultural adaptation processes during university years (35). These traits, in turn, are predictive of significant life outcomes (36). This continuous, iterative design and adjustment process aimed to produce a psychological status description that more closely approximates the reality captured by observational methods.

To enhance assessment efficiency and accuracy, our framework leveraged a combination of Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG) techniques. This AI-driven approach automated and refined the process of linking ecological data to psychometric constructs. The specific commercial LLM API utilized was Alibaba’s Qwen series (model: qwen-max).

The application of AI models involved a multi-stage process:

Automated content mapping: The LLM played a crucial role in assisting our research team (comprising life teachers and clinical psychologists) in the initial design and ongoing management of the ecological assessment content. Specifically, it was used to automate the mapping of each cognitive voting item and behavioral check-in definition to the 24 dimensions of our conceptual framework. This process significantly enhanced the standardization and efficiency of our content management. Through the use of the LLM, mapping tasks that would have required hours of manual expert deliberation were completed in minutes, greatly improving the real-time scalability of the assessment. The results of this comprehensive mapping are visualized in Figures 1–8.

A core tenet of our study design was to move beyond passive measurement and actively promote students’ mental health through a multi-faceted, interactive environment. The mini-program was designed not only to evaluate psychological states but also to foster self-awareness and motivate positive change. This was achieved by integrating gamified feedback, community interaction, and a nuanced alert mechanism, all delivered within the students’ everyday life contexts (37).

This study was initially launched in the autumn semester of 2022, emerging from an urgent need to dynamically monitor student mental health during a campus-wide lockdown due to the COVID-19 pandemic in Hohhot. The traditional, infrequent survey-based assessments were deemed insufficient for the volatile psychological environment. Consequently, this research began as an enhanced mental health management initiative for a specific cohort of students.

A total of 177 undergraduate students (98 female, 79 male; mean age = 20.53 years) from four classes in their second (sophomore) and third (junior) years of a clinical medicine program at a medical university participated in the study. The sampling method employed was cluster sampling, as the initial cohort consisted of all students residing in a single dormitory building under unified management during the lockdown. All students in this building were initially enrolled in the program. The final sample of 177 participants represents those students who remained and provided continuous, complete data across the entire one-and-a-half-year (three academic semesters) study period. Students who left the cohort prematurely (e.g., due to changes in lockdown status or academic leave) were excluded from the final analysis due to incomplete longitudinal data.

The sample size was determined by the size of the accessible student cluster in the unique real-world context, rather than by a priori power analysis, which is appropriate for the exploratory nature of this study. Following detailed introductions by their life guidance teachers (counselors), all participants provided voluntary informed consent electronically via the WeChat mini-program, fully understanding the research objectives, procedures, and confidentiality measures. The demographic characteristics of the participants are detailed in Table 2.

The data collection for this longitudinal study officially commenced in September 2022 and concluded in March 2024, spanning three academic semesters. Prior to the official start, the WeChat mini-program had been fully developed and had undergone preliminary small-scale testing. The measurement process consisted of three primary components:

Periodic psychometric screening: At the beginning of each semester, participants were organized to complete a systematic self-report screening using the full battery of standardized scales (SAS, SDS, SCL-90, EPQ, 16PF, and MBTI). This provided periodic baseline data for each assessment wave.

All statistical analyses were conducted using Python (version 3.8.18) with the SciPy (version 1.10.1), Pandas (version 1.5.3), and Scikit-learn libraries. Data visualization was performed using Matplotlib (version 3.7.1) and Seaborn (version 0.12.2). The level of statistical significance for all tests was set at p<0.05. The analysis was designed to address two core research questions.

To assess the potential intervention effect of long-term participation in the dynamic assessment system, we analyzed the changes in clinician-rated Hamilton Anxiety (HAM-A) and Hamilton Depression (HAM-D) scores from the first semester (November 2022) to the third semester (March 2024).

As shown in Table 10 and visualized in Figures 16, 17, participating students exhibited statistically significant reductions in both anxiety and depression levels over the study period. The mean HAM-A score decreased by 15.2%, from 20.61 to 17.49. The mean HAM-D score showed an even more pronounced reduction of 40.0%, decreasing from 13.02 to 7.81. Paired-sample t-tests confirmed that both of these changes were statistically significant (HAM-A: p = .004; HAM-D: p<.001).

These findings are particularly noteworthy as they are based on independent clinical assessments rather than self-report measures. The results suggest that continuous engagement with the dynamic assessment system—which provides regular feedback and promotes self-monitoring—may have a tangible, positive therapeutic effect. This demonstrates the system’s potential not only as a more accurate measurement tool but also as a practical, scalable intervention for improving student mental health.

The central finding of this study is that integrating high-frequency, ecologically-valid data from students’ daily lives significantly enhances the accuracy of traditional psychometric scales for anxiety and depression. The superior performance of our dynamic model, evidenced by higher AUCs and a stronger correlation with clinician ratings, underscores a fundamental limitation of static assessments: their inability to capture the fluid, context-dependent nature of mental states. A single questionnaire provides a cross-sectional “snapshot,” which can be easily biased by a student’s mood on a particular day. In contrast, our dynamic system functions like a longitudinal “film,” continuously updating the assessment based on a stream of cognitive and behavioral data points, thereby creating a more robust, nuanced, and authentic psychological profile.

The clinical implications of this finding are profound, signaling a potential paradigm shift in mental health assessment. In an era where daily life is increasingly mediated by digital technology, platforms like WeChat or wearable devices are becoming vast, untapped repositories of behavioral and cognitive data. Our research provides a proof-of-concept that these digital footprints can be ethically harnessed to move beyond reactive, clinic-based assessments towards a model of proactive, continuous, and personalized mental health monitoring. While the tools developed in this study are still nascent, they point towards a future where “digital phenotyping” could become a cornerstone of clinical practice. Such systems could enable clinicians to detect subtle negative changes in a student’s state long before they escalate into a crisis, facilitating early and targeted interventions.

Furthermore, the significant reduction in clinically-rated HAM-A and HAM-D scores suggests that the assessment process itself can be therapeutic. By engaging students in daily self-monitoring and providing immediate, narrative-based feedback, our system creates an active assessment-intervention loop. This process aligns with the principles of measurement-based care, where continuous data is used to inform and guide treatment. The gamified and humanized feedback encourages self-reflection and may empower students to make small but meaningful adjustments to their daily routines and thought patterns, fostering a sense of agency over their own mental well-being. This suggests that the future of digital mental health tools lies not just in their predictive power, but in their ability to function as interactive companions that promote psychological resilience.

While this study was initially driven by practical needs rather than a single theoretical doctrine, its design and findings resonate strongly with several contemporary psychological frameworks. Our approach can be understood as an applied synthesis of principles rooted in the state-trait theory of personality, Ecological Momentary Assessment (EMA), and the emerging field of digital phenotyping.

At its core, our research directly addresses the classic state-trait distinction (1). Our fundamental premise is that traditional scales, while effective at capturing stable “traits,” are insufficient for tracking the moment-to-moment “states” that constitute daily emotional life. By continuously integrating daily cognitive and behavioral data, our dynamic model is a deliberate attempt to create a “state-sensitive” measurement tool. The finding that our dynamic scores are more predictive of clinical ratings supports the idea that an accumulation of state-level data provides a more valid picture of an individual’s current mental health status than a single trait-level snapshot.

The methodology itself is a direct application of Ecological Momentary Assessment (EMA). By collecting high-frequency data in the students’ natural environment, we mitigate the recall bias inherent in retrospective questionnaires. This ecological validity is crucial for understanding the real-world triggers and fluctuations of anxiety and depression. Furthermore, our work extends EMA by not just collecting data, but by using it to actively correct and inform psychometric scores in near real-time.

Finally, our study contributes to the burgeoning field of digital phenotyping. The ultimate vision, as articulated in literature on the future of psychometrics, is the use of multimodal data from daily life—captured via smartphones and wearable devices—to construct a comprehensive picture of mental well-being (11). Our WeChat mini-program serves as a direct, albeit early, implementation of this vision. By translating digital interactions (cognitive votes and behavioral check-ins) into psychometrically meaningful adjustments, we are building a data-driven “digital phenotype” of student mental health. The finding that this process may also be therapeutic aligns with theories of Self-Regulation, where the feedback from one’s own data is a critical mechanism for behavioral change and goal attainment.

A particularly noteworthy and complex finding of this study was the observed decline in the predictive power of our dynamic model for the SCL-90 over time. While the model initially showed a clear advantage over the traditional scale in predicting clinical ratings, this superiority diminished by the third semester. We propose that this phenomenon is not attributable to a single cause, but rather to a multifactorial interplay of methodological, psychological, and contextual factors.

First, the possibility of a reactivity effect cannot be discounted. Over the course of one and a half years, it is plausible that students became “assessment-savvy.” They may have discerned the underlying patterns of the system, consciously or unconsciously adjusting their responses in cognitive voting and behavioral check-ins to present themselves in a more favorable light or to “manage” their profiles. This form of learned behavior would introduce noise into the ecological data, decoupling it from their authentic psychological state and thereby weakening the model’s predictive accuracy.

Second, the decline may reflect the inherent limitations of our model in capturing the complexity of the SCL-90 construct and real-world life events. The SCL-90 measures a broad spectrum of psychopathological symptoms, which are often influenced by significant, discrete life events (e.g., academic failures, relationship breakdowns, the end of the COVID-19 lockdown). Our model, while effective at tracking the general ebb and flow of anxiety and depression, may lack the specificity to account for the impact of such major external shocks on more complex symptom dimensions like psychoticism or paranoia. The initial success of the model might have occurred during a period of relative environmental stability (i.e., the lockdown), while its later decline could coincide with a return to a more chaotic and unpredictable post-pandemic campus life, where a wider range of unmeasured confounding variables came into play.

Third, and perhaps counterintuitively, the model’s declining predictive power might be a paradoxical signal of its success as an intervention. The continuous feedback and self-monitoring process may have genuinely enhanced students’ self-regulation and coping skills. As students became more psychologically resilient, their mental states may have stabilized and exhibited less variance. Consequently, their daily cognitive and behavioral patterns would become less predictive of pathology simply because there was less pathology to predict. In this view, the “signal” (i.e., symptom fluctuation) weakened, making it harder for the model to make accurate predictions against a baseline of improved mental health. Disentangling these three potential explanations presents a significant challenge and underscores a crucial direction for future research.

Despite the promising findings, this study has several limitations that must be acknowledged. First, the research was conducted in the context of a unique historical event—the COVID-19 pandemic and subsequent campus lockdown. This “natural experiment” setting, while providing a compelling rationale for the study, introduces significant confounding variables. The observed improvements in student mental health could be partially attributed to the cessation of the lockdown and a return to normal campus life (history effect), or to the natural process of student maturation over the one-and-a-half-year period (maturation effect).

Second, our study design lacks a randomized controlled trial (RCT) framework. The comparison group, while similar in academic background, was not randomly assigned, which introduces potential selection bias. Furthermore, the act of being continuously monitored and engaged might have induced a Hawthorne effect, where participants’ behavior changed simply because they were aware of being studied. While our results provide strong preliminary evidence, a future RCT with randomized allocation to either the dynamic assessment group or a control group (receiving only traditional assessments) would be necessary to definitively establish the causal effect of our intervention.

Third, our data collection, while ecologically valid, was limited to self-report (cognitive votes) and semi-objective data (counselor-logged behavioral check-ins). Future research should aim to integrate more objective, passive data streams, such as smartphone sensor data (e.g., screen time, mobility patterns) or wearable device metrics (e.g., sleep patterns, heart rate variability). This would create a more comprehensive digital phenotype, reduce the potential for reactivity effects, and further enhance the model’s predictive power. Finally, the AI model itself, while effective, requires further refinement in its ability to interpret and adapt to complex, long-term psychological changes.