Dynamic prediction of psychological stress in college students based on AI and multimodal data fusion

For effective mental health intervention, dynamically predicting psychological stress in college students is essential. Traditional stress assessment approaches are limited in capturing rapid temporal variations in stress levels, reducing their effectiveness for real-time intervention. This research proposes a novel Artificial Intelligence (AI) -based framework, namely Snap-Drift Cuckoo Search-driven Transformer Encoder with Adaptive Gated Recurrent Unit (SDCS-TE-Adaptive GRU), for dynamic psychological stress prediction using multimodal physiological signals. A publicly available multimodal physiological stress dataset from Kaggle, containing 3,000 samples collected from 245 college students, was utilized in this research. The dataset includes heart rate variability (HRV), electroencephalography (EEG), electrodermal activity (EDA), and respiration signals. The proposed framework integrates transformer-based multimodal feature fusion with Adaptive GRU temporal learning to model long-range dependencies and dynamic stress variations effectively. In addition, Snap-Drift Cuckoo Search Optimization (SDCSO) was introduced for feature relevance selection and model optimization. Experimental results demonstrated that the proposed SDCS-TE-Adaptive GRU model achieved 96.5% accuracy, and 95% Area Under the Curve (AUC), with a low cross-entropy loss of 0.12 and memory consumption of 420 MB. The model also achieved a fast detection time of 1.35 s and training time of 18 s. Comparative analysis confirmed that the proposed framework outperformed existing machine learning (ML) and deep learning (DL) approaches for real-time stress prediction and personalized mental health monitoring.