Live · updated just now

Computer model reads faces and body signals to track stress levels

Wednesday, 11 November 2026 · Mental health & wellbeing · Full report · Evidence: Early

What was found

Researchers built a computer model that tries to track a person's stress level moment by moment by combining three signals: video of the face, blood volume pulse (changes in blood flow that reflect heartbeat), and electrodermal activity (small shifts in skin sweat) 1. They tested it with strict safeguards against "leakage," a common problem where information from the test data slips into training and makes results look better than they are 1. Under these controls, the model's typical error was about 0.24 in five rounds of cross-checking and about 0.25 on a separate group of people kept aside for final testing 1. A key score called R², which shows how much better a model does than simply guessing the average, was negative in both cases (about -0.50 and -0.70) 1. This means the model predicted stress less accurately than a plain guess of the average stress level 1.

Who it applies to

The work concerns automated, continuous prediction of psychological stress from face video and body signals 1. It is a technical study of model design, not a study of people seeking help for stress.

How strong the evidence is

Early. This was a machine learning development study that tested whether the model works on people it had not seen before 1. It used 56 people from an existing research dataset called UBFC-Phys, with 45 used to build the model and 11 held back for testing 1. The group is small, and the model did not handle new individuals better than an average-based guess 1. A second dataset, AMIGOS, was explored but lacked checked, continuous stress measurements, which limits what it can show 1. Only the abstract was read, because the full paper is not openly available.

What it means in practice

When tested carefully, current models of this kind may not work reliably on new people 1. Differences between individuals appear to be a major hurdle that would need to be solved before such tools could be trusted in everyday settings 1.

What others say

A science news report highlighted that the strict anti-leakage approach revealed real weaknesses that looser testing might hide 2. It noted that by avoiding inflated scores, the study showed the model could not beat a simple average-based guess on people it had not encountered before 2.