Six adults came to the lab on three consecutive days. On one day they simply rested, on another they pedaled a bike to exhaustion, and on the third they went through an ordeal that makes the heart race without a single movement: they had to give a speech and solve problems in front of a stern committee. Each wore a wristband that recorded heart rate, its variability (the spread of intervals between beats), skin electrical activity, and movement. Saliva was collected five times per session to measure cortisol — the hormone released into the blood during stress.

The recordings were sliced into ten-minute chunks and labeled: rest, physical stress, recovery from physical stress, psychological stress, recovery from psychological stress. Then a computer program was trained to guess the person's state using only the wristband data. On average, it got it right 78% of the time. It recognized physical exertion and recovery well, but it confused psychological stress and its recovery with rest: out of ten such states, it correctly identified only five or six.

When five numbers describing cortisol levels in saliva over the same time window were added to the wristband data, accuracy jumped to 94%. Recognition of psychological stress improved especially: the program correctly identified it in 83% of cases, and recovery from it in 88%. Cortisol helped separate states that previously blurred together: psychological stress no longer hid behind rest.

So wrist sensors alone are not enough to tell anxiety from calm. But add saliva analysis, even occasionally, and the picture becomes clear. This paves the way for more honest stress monitoring systems that won't mistake pre-exam jitters for lounging on the couch.