From Clean Labs to Noisy Lives: Real-World Stress Detection Using Spectrogram-Based Transformers on PPG Signals

Anice Jahanjoo, Yiting Wei, Soheil Khooyooz, Mostafa Haghi, Nima TaheriNejad

Confirmation: I have read and agree with the IEEE BHI 2025 conference submission's policy on behalf of myself and my co-authors.

Keywords: Attention mechanism, biomedical signal processing, photoplethysmography (PPG), real-world data, stress detection

TL;DR: We propose a transformer-based model for stress detection using spectrogram images of real-world PPG signals, highlighting the performance gap between lab and natural settings and emphasizing the need for robust methods under real-life conditions.

Abstract:

Accurate detection of human stress levels is crucial
for mental health monitoring and has wide-ranging applications
in workplace wellness and healthcare. While Photoplethysmog-
raphy (PPG) signals have been increasingly utilized to analyze
physiological states, most existing studies are limited to controlled
experimental settings. This paper addresses this gap by collecting
PPG data from employees during real-world work conditions
using a wearable device, thereby enhancing the validity and
applicability of stress detection systems. We propose a novel
stress classification method based on a multi-head attention
transformer model, capturing both temporal and frequency-
domain features in the PPG signal. We apply Short-Time Fourier
Transform (STFT) to extract spectral representations of the PPG
data. A transformer architecture is then employed to model
complex dependencies and subtle variations in physiological
signals via self-attention mechanisms and stacked encoder layers.
Experimental evaluation demonstrates that our proposed method
achieves a classification accuracy of 76.17% and an F1-score
of 76.42% in stress detection task, outperforming machine
learning baselines and state-of-the-art methods. These findings
highlight the effectiveness of transformer-based approaches in
stress classification and reveal the substantial performance gap
between laboratory-controlled results and real-world outcomes.

Track: 1. Biomedical Sensor Informatics

Registration Id: TXNDQSSZY5S

Submission Number: 68