Why Nonlinear Models Revolutionize Physiological Monitoring?

Wearable devices like the Empatica E4 bracelet record signals such as skin conductance, temperature, and movement. These data change in complex ways when we are focused, stressed, or exercising. Researchers compared linear and nonlinear models on three public datasets. Results show nonlinear models achieve 0.89‑0.98 accuracy and up to 0.99 ROC‑AUC, while linear models stay below 0.73. Even in Leave‑One‑Subject‑Out validation, nonlinear models keep good generalization. Feature analysis reveals that combining EDA, temperature, and accelerometer data is most informative. SHAP explanations highlight physiologically meaningful patterns for stress, cognitive load, and physical effort. In short, our bodies behave non‑linearly, so advanced algorithms capture nuances better. This work offers a unified benchmark for building more reliable monitoring systems in health, sport, and education.

Physiological signals captured by wearable sensors are typically non‑stationary and exhibit temporal and nonlinear dependencies. Skin conductance (EDA) reflects sudoriferous gland activity, controlled by the sympathetic nervous system, and varies in response to emotional or cognitive stimuli. Skin temperature, on the other hand, is influenced by vasodilation and thermoregulation, processes dependent on both internal (metabolism) and environmental factors. The accelerometer (ACC) provides three‑dimensional movement data, useful for distinguishing static from dynamic activities. In linear models, the relationship between these variables and physiological state is approximated by linear combinations of features, limiting the ability to capture complex interactions. Nonlinear models—deep neural networks, gradient‑boosted decision trees, and RBF‑kernel SVMs—can learn highly nonlinear mapping functions, allowing them to model threshold effects, signal synergies, and temporal dynamics. The Leave‑One‑Subject‑Out (LOSO) approach highlights inter‑individual variability: differences in sudoriferous response, thermoregulation, and movement patterns require models capable of generalizing beyond the training subject. Ablation analysis shows that removing EDA or ACC drastically reduces performance metrics, confirming the crucial role of multimodal fusion. SHAP interpretability visualizes local feature importance, revealing, for instance, that EDA spikes combined with brief accelerations indicate acute stress, while slow temperature changes paired with continuous motion signal moderate exercise. These findings suggest that future monitoring platforms should integrate adaptive preprocessing pipelines, subject‑specific normalization, and robust nonlinear models to enhance accuracy and transferability in real‑world settings.