JMIR Publications

ABSTRACT

Background:

Wrist-worn and ring-based “stress” scores are now embedded in consumer wellness apps, workplace well-being platforms, and research mHealth pipelines. Many commercial implementations threshold mean heart rate (HR), a compact heart rate variability (HRV) index, or tonic electrodermal activity (EDA) and present the result as mental or cognitive stress. Psychophysiology and ambulatory sensing research have long shown that cognitive load and physical exertion can raise the same autonomic channels, so a level-only rule can be construct-ambiguous whenever exercise is possible during wear time.

Objective:

This digital-health methods study asks a design and reporting question: how much of the discriminative information for cognitive-versus-physical typing on public multimodal wearable evidence is carried by temporal dynamics, what those public quantities do and do not support, and what logging fields are reasonable to recommend before a cognitive-load interpretation is claimed.

Methods:

We synthesize published Random Forest importance summaries and protocol effect sizes from a PhysioNet wearable release of induced stress and structured exercise (100 sessions; 41 participants; classes STRESS, AEROBIC, and ANAEROBIC), together with related wearable stress corpora, ambulatory confounding studies, and reporting guidance. Temporal share of top-30 importance, the temporal-to-static ratio R_t/s, and absolute Cohen d for stress-versus-exercise and aerobic-versus-anaerobic contrasts were used as design quantities. No new participants were enrolled and no new classifier was trained.

Results:

Temporal predictors accounted for 36.74% of top-30 importance (R_t/s = 0.58), and trend slope ranked first. Stress-versus-exercise contrasts were large for both trend slope (|d| ≈ 1.53) and mean HR (|d| ≈ 1.59); aerobic-versus-anaerobic contrasts on the same axes remained small (|d| < 0.27). In this seated-stress versus structured-exercise protocol, a plain mean-HR contrast therefore separates classes at least as cleanly as trend slope. Published session morphology in the same source also reports earlier onset and peak under STRESS than under exercise, but those fields were not ranked in the importance table reused here.

Conclusions:

On these public quantities, trend slope is a first-rank companion to level features, not a replacement that uniquely rescues cognitive-versus-exertion typing. A four-field temporal logging card (onset, peak, recovery, trend slope) is offered as a design recommendation grounded in psychophysiology and in published morphology from the source evaluation—not as a finding that each of the four fields was importance-validated here. Level-only scores remain useful for coarse arousal and, on this protocol, for seated-stress versus exercise separation; they are under-tested for the harder free-living case of cognitive load with elevated HR versus light exertion.