Deep learning-based physiological risk stratification in night-shift hospital workers.

This cross-sectional study of 2,250 hospital workers attempted to identify potential physiological risk structures for night workers by utilizing the nonlinear manifold learning framework. Unlike conventional mean-centered linear analysis, this study applied the Potential of Heat-diffusion for Affinity-based Transition Embedding (PHATE) for topological visualization and Neural Additive Models (NAM) to predict interpretable risk. The ‘cholesterol homeostasis profile’ was defined as the main latent indicator. Analysis showed that the systemic metabolic load (PHATE-1) was higher after age correction, despite the night-shift group being younger than the non-night shift group (p < 0.001). This axis showed a positive correlation with body mass index (r = 0.71) and triglyceride (r = 0.56). The NAM model captured the risk slope with an average AUC of 0.864 (range 0.832–0.901) and confirmed a nonlinear risk transition when it exceeded the triglyceride 2 mmol/L threshold. The predicted risk of night workers in the highest risk cluster (cluster 1) was 0.71, and the metabolic load and cholesterol homeostasis profile risk were increased. Ultimately, the PHATE-NAM framework moves beyond traditional group averages. It provides a purely data-driven way to pinpoint individual physiological vulnerabilities, giving hospitals the exact evidence needed to design safer, more sustainable shift-work environments.