Physiological Stress Detection and Real-Time Biofeedback Applicable to Alzheimer’s Patients via an Edge-Computing Multi-Sensor System

Abstract

In vulnerable demographics, such as patients struggling with Alzheimer’s disease, continuous physiological monitoring is necessary for identifying and mitigating stress responses. However, traditional stress monitoring systems often rely on computationally heavy cloud processing and lack localized detection mechanisms, which could lead to delayed caregiver responses. To address these limitations, this study describes the development of a low-latency sensor system which is capable of affective state classification and real-time biofeedback. The system architecture utilizes an ESP32 microcontroller to process live 1-second interval data streams from an electrodermal activity (EDA) sensor, a MAX30102 optical blood volume pulse (BVP) sensor, and an MPU6050 accelerometer across a 30-sample sliding window (the latter two sensors are I2C-based). Additionally, a Random Forest machine learning classifier, implemented directly on the system, evaluates 8 z-score scaled physiological and kinematic features (EDA mean, stdev, max, min, BVP heart rate, and accelerometer mean, stdev, max). To validate the model’s accuracy, physiological data is being collected on a healthy control group using the standardized Trier Social Stress Test (TSST) protocol to establish baseline, stress and amusement levels. When the model detects a consistent stress anomaly across 5 consecutive sampling windows, the system autonomously activates a pulse-width modulation (PWM) biofeedback motor to guide user breathing while sending a Bluetooth Low Energy (BLE) alert to a remote caregiver device. While currently being validated using standard protocols on a healthy control, future research entails refining the edge-computing system on clinical Alzheimer's populations