Network structure of burnout, perceived stress, social support, and insomnia among Chinese healthcare workers: a cross-sectional study integrating Gaussian graphical model and Bayesian network analysis

Healthcare worker burnout is a global concern, yet its interrelationships with comorbid psychological factors remain poorly described at the sub-dimension level. Most network studies use undirected models alone, which estimate which variables are uniquely associated but assign no ordering. A cross-sectional survey of 362 healthcare workers at a Chinese tertiary hospital assessed nine sub-dimensional nodes covering burnout, perceived stress, social support and insomnia. A Gaussian Graphical Model (GGM) with graphical LASSO and a Spearman input matrix estimated the undirected network; Bayesian Network (BN) structure learning used Hill-Climbing with the BIC-g score, a cross-validated parent limit and bootstrap model averaging (1000 resamples, threshold 0.50). Robustness was examined against normalization, covariate adjustment, tenure restriction, alternative algorithms and sparsity levels. The GGM retained 33 of 36 edges; the strongest were emotional exhaustion–depersonalization (EE–DP, 0.549) and personal accomplishment–out of control (PA–OutCtrl, 0.536), but only 13 edges had bootstrap confidence intervals excluding zero. EE showed the highest strength centrality (1.301), expected influence (0.973) and predictability (R² = 0.677), most of which was shared with DP. Case-dropping stability coefficients were 0.517 for strength and 0.751 for expected influence, above the preferred threshold of 0.50. OutCtrl was the leading bridge node under the theoretical partition (bridge expected influence = 0.541), but the bridge ranking as a whole depended on how the communities were defined. Model averaging yielded 15 directed arcs, all corresponding to retained GGM edges. In the selected DAG, EE was classified as a source and OutCtrl as a sink, although arc orientations were substantially less stable than the underlying connections (directed-arc agreement across algorithms 0.22–0.79). Which variables were connected was reproducible across preprocessing choices and algorithm families. EE is the most strongly connected and most predictable node in this network, and OutCtrl links the stress and burnout domains under the theoretical partition. In the selected DAG, EE was classified as a source node, but arc orientations were the least stable part of the model. These are structural descriptions of cross-sectional data; they do not identify causal relations or intervention targets.