Review of Stress Detection Methods Based on TinyML

Abstract

Stress has become a serious public health issue with significant effects on physical and mental health, highlighting the need for reliable methods for early and continuous detection. Wearable sensors enable non-invasive monitoring of physiological signals related to stress. However, cloud-based solutions face challenges such as latency, energy consumption and privacy restrictions, which limit their suitability for resource-constrained wearable devices. Tiny Machine Learning (TinyML) represents an alternative approach by enabling on-device stress detection in low-power embedded systems. This paper provides an overview of TinyML-based stress detection approaches published between 2020 and 2025. It describes the most common physiological signals (photoplethysmography (PPG), electrocardiography (ECG), heart rate variability (HRV), electrodermal activity (EDA), Heart Rate (HR), skin temperature, respiration (RESP) and triaxial accelerometer), as well as commonly used benchmark datasets (Wearable Stress and Affect Detection (WESAD), Nurse Stress Dataset, Montreal Stress Test (MIST), ISRUC and SWELL), and hardware platforms (ESP32, RP2040-based Raspberry Pi Pico, STM32 series microcontrollers, Arduino Nano 33 BLE Sense and heterogeneous platforms such as InfiniWolf). A comparative analysis evaluates performance, memory usage, latency and deployment feasibility of different TinyML architectures. The results show that lightweight models, particularly deep neural networks (DNNs) and one-dimensional convolutional neural networks (1D CNNs) combined with optimization techniques such as quantization, allow for an optimal trade-off between accuracy and resource usage. Future research directions and key challenges for developing scalable and reliable TinyML-based stress detection systems for wearable healthcare applications are also discussed.