Many AI models rely on supervised learning: you provide the system with hundreds or thousands of examples, each labeled—in our case “stress” or “no stress”—and the model learns to recognize patterns on its own. The challenge, however, is that assigning these labels is not always straightforward.

Stress is not directly observable or easily measurable. That is why we opted for a controlled laboratory setup using the Trier Social Stress Test, a method specifically designed to induce stress. We invited 30 employees from Howest and AZ Groeninge, each of whom went through three phases individually: a resting phase, a stress phase, and a recovery phase.

On paper, this seems simple: during the stress phase, everyone experiences stress, and during the other phases, they do not. In practice, however, this assumption did not hold. Some participants were already nervous during the resting phase, while others remained surprisingly calm during the stress test.

This demonstrates how misleading assumptions can be. If you train a model on incorrectly labeled data, it will learn the wrong patterns.

To introduce more objectivity, we collaborated with AZ Groeninge Hospital to measure cortisol levels through saliva samples. Cortisol is widely regarded as a reliable indicator of stress. However, it also has limitations as a ground truth. First, cortisol responds with a delay: it typically peaks 15 to 20 minutes after a stress event. Second, translating cortisol values into binary labels is inherently arbitrary. If a 20% increase indicates stress, what about someone with a 19% increase?

Why not simply ask participants whether they felt stressed? We did, but self-reporting has its own limitations. People differ in how they report experiences, and some may hesitate to admit they were stressed.

The solution: “label propagation”

The solution lay in a technique called label propagation (implemented in the python scikit learn package). This approach combines multiple data sources to achieve more accurate labeling. First, we distinguished between certain and uncertain data. Data points where both the experimental phase and cortisol measurements aligned were considered reliable.

For the uncertain cases, we applied a clustering algorithm. This algorithm grouped physiologically similar moments without prior knowledge of whether they represented stress. These clusters were then linked to the reliable data to assign labels.

The results were striking. A quarter of the data labeled as “stress phase” turned out not to contain stress. Additionally, a small portion of the resting phase—especially in the first minutes—did show clear stress signals. This again highlights how misleading simple assumptions can be.

By assuming that everyone was relaxed during the resting phase and stressed during the stress phase, we introduced incorrect labels in more than 10% of the total dataset. This naturally confused the model. In many cases, the heart rate was low, yet the model was told the person was stressed. Once we corrected the labeling, the performance of the AI model improved significantly: accuracy increased from 86% to 94% when detecting stress in new individuals.