Abstract:The negative effects of stress on well-being demonstrate the need for real-time detection. The increasing prevalence of wearables with AI capabilities to continually monitor vital signs like heart rate and blood pressure highlights their growing value in promptly identifying stress. This paper presents a new approach for real-time stress detection employing an LSTM-based deep learning model, a technique with promising outcomes for time-series data analysis. Our model is trained on raw PPG signals from the WESAD dataset covering diverse stress scenarios. By learning patterns and fluctuations over time, it can effectively distinguish stressed and non-stressed states. The training uses only the raw PPG signals, segments them into windows, and creates labeled data for supervised learning. To enable real-time detection, we explore deploying our trained model on STM32H7xx microcontrollers equipped with a Cortex-M7 core offering low-power and hardware acceleration. We implement the LSTM model leveraging their capabilities for efficient inference. This implementation process involves optimizing the model and converting it into a format compatible with the microcontrollers. Within this study, we employ key TensorFlow toolkit optimization methods, including quantization-aware training (QAT), pruning, prune-preserving quantization-aware training (PQAT), and post-training quantization (PTQ), along with the TensorFlow Lite (TFL) toolkit, to evaluate and compare the outcomes obtained from applying these methods to the baseline model. Our goal is to select the most effective approach for the processor, enabling real-time detection. Through the utilization of these techniques, our objective is to reduce the size of the model and the necessary processing resources, such as RAM size, while ensuring that a high level of accuracy is maintained. Our results show the capability of the optimized LSTM model to accurately detect stress from raw PPG data on resource-constrained, low-power STM32H7xx MCUs. The final optimized model requires only 170 Kbytes of RAM, a nearly 12 times reduction in size, while still achieving a high accuracy of 87.76% when performing inference on the microcontroller.

Wavelength Selection for Real-Time Detection of Human Stress Based on StO2

CoSTHR: A Heart Rate Estimating Network with Adaptive Color Space Transformation

Tracing Human Stress from Physiological Signals using UWB Radar

Personalized Stress Monitoring using Wearable Sensors in Everyday Settings

An Integrated Human Stress Detection Sensor Using Supervised Algorithms

A Handheld Technology for Optical Monitoring of Transcutaneous Blood Oxygen Saturation

A Transductive Model-based Stress Recognition Method Using Peripheral Physiological Signals

StressNet: Detecting Stress in Thermal Videos

Real-Time Psychological Stress Detection According to ECG Using Deep Learning

Research on Multiple Spectral Ranges with Deep Learning for SpO2 Measurement

Measurement of stress-induced sympathetic nervous activity using multi-wavelength PPG

Fast Multi-Level Mental Stress Identification from Bispectrum-based Heart Rate Variability Feature

Evaluating Feature Selection for Stress Identification

Real-Time Stress Detection from Raw Noisy PPG Signals Using LSTM Model Leveraging TinyML

Wavelength Optimization and Experimental Validation of Near-Infrared Tissue Oxygenation Detection

Global Stress Detection Framework Combining a Reduced Set of HRV Features and Random Forest Model

Hyperspectral Imaging for Non-invasive Blood Oxygen Saturation Assessment

An Adaptive System for Wearable Devices to Detect Stress Using Physiological Signals

Three-wavelength system for practical application in skin oximetry: simultaneous equations with prediction-correction approach

Mental Stress Detection: Development and Evaluation of a Wearable In-Ear Plethysmography

Mental Stress Recognition based on Non-invasive and Non-contact Measurement from Stereo Thermal and Visible Sensors