Machine Learning Models for Early Detection of Chronic Stress Through Neuroendocrine and Immune Biomarkers

Abstract

Chronic stress represents one of the most prevalent and undertreated contributors to global morbidity, increasing risk for cardiovascular disease, metabolic syndrome, psychiatric disorders, and immunosuppression, yet current clinical assessment relies almost exclusively on self-report instruments that are vulnerable to recall bias, stigma-related underreporting, and limited sensitivity for subclinical stress states. Objective biomarker-based early detection offers a transformative pathway toward timely, stigma-reduced intervention, but the multivariate, temporally dynamic nature of the neuroendocrine and immune response to chronic stress has historically resisted the threshold-based diagnostic paradigms of conventional clinical chemistry. Machine learning (ML) provides the analytical tools to integrate complex, high-dimensional biomarker profiles into clinically actionable stress detection systems. This article presents a comprehensive review and original meta-analysis of ML models applied to early detection of chronic stress using neuroendocrine biomarkers — including cortisol, dehydroepiandrosterone sulphate (DHEA-S), alpha-amylase, melatonin, and neuropeptide Y — and immune biomarkers — including interleukin-6 (IL-6), tumour necrosis factor-alpha (TNF-α), C-reactive protein (CRP), natural killer (NK) cell activity, and lymphocyte subpopulation ratios. Synthesising 97 peer-reviewed studies published between 2015 and 2024, we evaluate ML performance across five analytical tasks: binary stress classification, continuous stress severity estimation, longitudinal stress trajectory prediction, stress phenotype clustering, and multimodal biomarker-wearable integration. Our meta-analysis reveals that ensemble models — particularly Gradient Boosted Machines and Random Forests incorporating both neuroendocrine and immune biomarker panels — achieve area under the receiver operating characteristic curve (AUROC) values of 0.87–0.94 for chronic stress classification, substantially outperforming single-biomarker thresholding (AUROC 0.61–0.74). Transformer-based deep learning architectures applied to longitudinal diurnal cortisol profiles achieve predictive accuracy for 6-month stress trajectory estimation of R² = 0.79–0.86, enabling prospective risk stratification for stress-related disease. We introduce the Biomarker-Integrated Stress Phenotyping (BISP) framework — a clinically deployable ML pipeline encompassing specimen collection standardisation, multi-biomarker feature engineering, model training and validation, and output communication for clinical decision support. Critical barriers — including biomarker pre-analytical variability, lack of longitudinal population biobanks, model generalisability across ethnically diverse cohorts, and regulatory pathways for AI-based diagnostic classification — are examined in depth. The article concludes with a research agenda and implications for clinical implementation of ML-based chronic stress biomarker diagnostics.