AI-driven dynamic psychological measurement: correcting university student mental health scales using daily...

This study aimed to evaluate an Artificial Intelligence (AI)-driven dynamic psychological measurement method for correcting traditional mental health scales. We sought to validate its feasibility using daily behavioral and cognitive data from university students and assess its potential as an intervention tool.

A total of 177 university students participated in a one-and-a-half-year study. Using a WeChat mini-program, we collected data from cognitive voting (87 instances), behavioral check-ins (66 instances), and standardized psychological scales (SAS, SDS, SCL-90). Scale scores were dynamically adjusted using Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG) techniques. Paired-sample t-tests, MANOVA, and Cohen's d were used to compare the performance of the dynamic model against traditional scales. Intervention effects were validated using the Hamilton Anxiety Rating Scale (HAM-A) and Hamilton Depression Rating Scale (HAM-D).

The dynamic assessment demonstrated superior performance in identifying both anxiety (SAS: dynamic model AUC = 0.95 vs. traditional AUC = 0.86) and depression (SDS: dynamic model AUC = 0.93 vs. traditional AUC = 0.82). Over three semesters, participating students showed significant decreases in clinically-rated anxiety scores on the HAM-A (15.2% reduction; 95% CI for mean difference [1.00, 5.25], p = 0.004) and depression scores on the HAM-D (40.0% reduction; 95% CI for mean difference [2.71, 7.71], $p<0.001$ ). High student engagement was observed (cognitive voting participation: 79%; behavioral check-ins: 42%). While the dynamic adjustment for the SCL-90 was initially effective ( $R^2=0.34$ ), its specificity later decreased, potentially due to interference from life factors (dynamic model MSE = 102.74 vs. traditional MSE = 84.17).

AI-driven dynamic assessment provides superior accuracy for anxiety (SAS) and depression (SDS) scales over static methods by effectively capturing psychological fluctuations. The significant reductions in clinically-rated anxiety and depression suggest the system may function as an integrated assessment-intervention loop, fostering self-awareness through continuous feedback. High user engagement confirms the method's feasibility. However, the model's diminished specificity for the complex SCL-90 scale over time highlights challenges in handling intricate, long-term symptom patterns. This research supports a shift towards continuous "digital phenotyping" and underscores the need for rigorous validation, multimodal data integration, and robust ethical considerations.