Volatile Organic Compounds for Stress Detection: A Scoping Review and Exploratory Feasibility Study with Low-Cost Sensors

TL;DR

This study reviews VOCs as a novel modality for emotion detection, maps existing evidence, and explores the feasibility of using low-cost sensors combined with physiological data to identify stress, highlighting promising results and key challenges.

Contribution

It provides a comprehensive review of VOC-based affective computing and demonstrates initial feasibility of low-cost sensors for stress detection, identifying critical technological gaps.

Findings

VOC signatures reflect stress and affective states but with heterogeneity.

Low-cost TVOC sensors combined with physiological data can classify stress with 77.3% accuracy.

Significant interindividual variability and calibration needs were identified.

Abstract

Volatile organic compounds (VOCs) represent a novel but underexplored modality for emotion recognition. This paper presents a systematic evidence synthesis and exploratory investigation of VOC-based affective computing using low-cost sensors. Study 1, a systematic scoping review following PRISMA-ScR guidelines, analyzed 16 studies from 610 records across breath, sweat, skin, and urine biosources. Evidence indicates that stress and affective states are reflected in VOC signatures (aldehydes, ketones, fatty acids, sulfur compounds), though with considerable heterogeneity. Current research relies predominantly on laboratory-grade GC-MS or PTR-MS, while wearable sensors provide pattern-level outputs without compound-specific identification - a critical gap for practical systems. Study 2 (n=25) investigated whether low-cost TVOC sensors (BME688, ENS160) combined with physiological monitoring (HR, HRV, GSR) can detect laboratory-induced stress. Exploratory analysis revealed that high cardiovascular reactors exhibited elevated TVOC during arithmetic stress (d=1.38), though requiring replication in larger samples. Substantial interindividual variability emerged (CV>80%), with coupling patterns moderated by baseline emission levels and temporal lags of 30-80 seconds. Random Forest-based multimodal classification achieved 77.3% accuracy (5-fold CV). SHAP analysis indicated VOC sensors contributed 24.9% of model performance. Leave-one-subject-out validation yielded 65.3% accuracy, highlighting the need for individual calibration. This work provides three contributions: (1) comprehensive mapping of VOC biomarker evidence and technological gaps, (2) initial demonstration that low-cost sensors can capture stress-related VOC patterns in multimodal fusion, and (3) identification of key implementation challenges. Findings require replication in larger samples (n>=50).