A review of 19 studies finds wearable AI for student stress is not yet accurate enough

Accuracy alone flattered the tools. Average sensitivity was 0.755, specificity 0.744 and F1-score 0.759, and the authors judged performance suboptimal, noting that AI has scored higher for cancers, heart disease and eye disorders in earlier reviews. They advise using wearable AI alongside clinical questionnaires.

The evidence base is thin: most included studies were small, results varied enormously between studies, and none tried to predict future stress or tell stress apart from depression or anxiety.

Two reviewers independently selected studies, extracted data and assessed bias with an adapted QUADAS-2 tool, with high agreement (kappa 0.87 for titles and abstracts, 0.92 for full texts). The review was registered with PROSPERO and followed PRISMA-DTA reporting guidance. Studies that used public datasets such as WESAD were excluded.

The included studies used 16 different wearable devices, most often Empatica E4 and Samsung Galaxy watches, worn mostly on the wrist (16 of 19). They tested 22 algorithms, most commonly support vector machines (14 of 19) and k-nearest neighbour (13 of 19). Most studies sorted students into two classes, stressed or not stressed (15 of 19).

Subgroup analyses found accuracy varied significantly with number of stress classes (P=.02), device type (P=.049), device location (P=.02), dataset size (P=.009) and ground truth (P=.001). The authors note that these factors overlap: the studies using electrodes, multiple body sites and objective labels were largely the same ones, so the groupings should be read with caution.

The authors say they cannot endorse immediate use in clinical or academic settings, citing suboptimal performance, small samples, few studies feeding the meta-analyses and high risk of bias in many studies. They call for studies of postgraduate and high school students, future stress prediction, and gold-standard cortisol-based ground truth.