Towards lightweight stress monitoring on biometric data for IoMT environments

Background and objective: Stress is a physiological response mechanism that enables humans to react to perceived threats through a fight-or-flight response. While beneficial in acute situations, prolonged exposure to stress can lead to significant physical and mental health issues, making early and...

Descripción completa

Detalles Bibliográficos
Autores: Montoya-Peña , Carlos, Gaitán-Guerrero, Juan Francisco, Espinilla, Macarena, López, Jose L.
Tipo de recurso: artículo
Estado:Versión publicada
Fecha de publicación:2026
País:España
Institución:Universidad de Jaén
Repositorio:RUJA. Repositorio Institucional de la Producción Científica de la Universidad de Jaén
OAI Identifier:oai:ruja.ujaen.es:10953/7487
Acceso en línea:http://sciencedirect.com/science/article/pii/S0169260726000556?via%3Dihub
https://hdl.handle.net/10953/7487
Access Level:acceso abierto
Palabra clave:Stress monitoring
Biometric data
HRV
Respiratory signals
Healthcare
Machine learning
Internet of medical things
004
004.3
004.4
004.6
Descripción
Sumario:Background and objective: Stress is a physiological response mechanism that enables humans to react to perceived threats through a fight-or-flight response. While beneficial in acute situations, prolonged exposure to stress can lead to significant physical and mental health issues, making early and reliable detection essential. Although many existing approaches achieve high accuracy by relying on numerous physiological signals and features, such solutions are often unsuitable for Internet of Medical Things (IoMT) applications that increasingly rely on edge computing paradigms. In these scenarios, stress detection models must operate directly on resource-constrained devices with limited computational and energy budgets. Therefore, this work proposes a lightweight and efficient methodological framework for stress detection, specifically designed for edge-based IoMT deployment. Methods: Eight supervised Machine Learning (ML) algorithms were evaluated: Random Forest (RF), LightGBM, CatBoost, XGBoost, Logistic Regression (LR), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), and a Multilayer Perceptron (MLP). All models were trained using Heart Rate Variability (HRV) and respiratory features extracted from the WESAD dataset. The proposed framework combines population-level training with subject-specific adaptation and evaluates model performance under progressive dimensionality reduction using subsets of 15, 10, 8, 6, and 4 features. Results: The proposed two-stage framework demonstrates that subject-specific adaptation significantly improves stress detection performance. XGBoost achieved the highest balanced accuracy (95.1% 4.7%) using 10 features, outperforming the configuration with all 15 variables. Crucially, the study identifies a reduced set of 6 features as the optimal deployment configuration; despite its further reduced feature set, it showed no statistically significant performance loss compared to the 10-feature model (95% CI: −0.0078, 0.0068) and maintained a 99.6% probability of outperforming the best models from all other architectures evaluated. Conclusions: The results show that accurate and personalized stress detection is feasible using reduced feature sets, enabling efficient, interpretable, and real-time deployment of ML models in wearable and IoMT-based monitoring systems.