Tools
Tools
Bharot, Nitesh, Sirohi, Kartik, Syed, Muslim Jameel, Muntean, Cristina Hava and Verma, Priyanka (2026) P2INNs: Physio-Psychiatry Informed Neural Networks for Depression Risk Prediction. IEEE Access, 14. pp. 132467-132482. ISSN 2169-3536
Preview |
PDF
Download (2MB) | Preview |
Abstract
Physics-Informed Neural Networks (PINNs) incorporate prior knowledge into learning by embedding differential equation constraints into the training objective. Motivated by the limited adoption of such approaches in computational psychiatry, we propose Physio-Psychiatry Informed Neural Networks (P2INNs) for depression-related prediction tasks. P2INNs integrate clinically motivated differential constraints derived from empirical physiological literature, capturing established relationships between sleep, physical activity, and affective risk. These constraints act as structured inductive biases during training, guiding the network toward physiologically plausible solutions while preserving flexibility to model inter-individual variability. We evaluate the proposed framework across two independent case studies using the StudentLife and Depresjon datasets, spanning regression and classification settings with heterogeneous affective targets. Experimental results demonstrate that P2INNs consistently improve predictive performance in low-data regimes and exhibit increased robustness to hyperparameter variation compared to standard neural networks and conventional regularization techniques. We further compare against monotonic neural networks and sign-only gradient penalties, demonstrating that the full ODE-form constraint provides measurable advantages beyond enforcing monotonicity alone. While classical machine learning baselines remain competitive on this low-dimensional feature set, P2INNs offer a principled mechanism for embedding physiological priors into neural architectures, with potential to scale to richer and temporally structured inputs.
Actions (login required)
 |
View Item |