Applying neural ordinary differential equations for analysis of hormone dynamics in Trier Social Stress Tests.

ABSTRACT

Introduction: This study explores using Neural Ordinary Differential Equations (NODEs) to analyze hormone dynamics in the hypothalamicpituitary-adrenal (HPA) axis during Trier Social Stress Tests (TSST) to classify patients with Major Depressive Disorder (MDD). Methods: Data from TSST were used, measuring plasma ACTH and cortisol concentrations. NODE models replicated hormone changes without prior knowledge of the stressor. The derived vector fields from NODEs were input into a Convolutional Neural Network (CNN) for patient classification, validated through cross-validation (CV) procedures. Results: NODE models effectively captured system dynamics, embedding stress effects in the vector fields. The classification procedure yielded promising results, with the 1x1 CV achieving an AUROC score that correctly identified 83% of Atypical MDD patients and 53% of healthy controls. The 2x2 CV produced similar outcomes, supporting model robustness. Discussion: Our results demonstrate the potential of combining NODEs and CNNs to classify patients based on disease state, providing a preliminary step towards further research using the HPA axis stress response as an objective biomarker for MDD.