Yuting Wang, Na Ni, Hui Chen, Jiahao Zhao, Yanqing Xie, Bixia Zhao, Fengying Zhang
In this cross-sectional clinical sample, the estimated network was characterized by predominantly stronger within-modality associations and limited cross-modal conditional associations between psychometric and PSG indicators. These findings suggest that psychometric and PSG assessments may capture partly non-overlapping and complementary aspects of patients' clinical presentations. However, the observed network structure does not establish mechanistic decoupling, causal direction, or independence between psychological and physiological processes. Longitudinal studies with closer temporal alignment and repeated PSG measurements are needed to determine the sources and clinical significance of these cross-modal associations.
BACKGROUND: Subjective-objective sleep discrepancy is widely recognized in sleep medicine, but the multivariable relationships between psychometric assessments and polysomnographic (PSG) parameters remain insufficiently understood. Previous network studies in sleep populations have relied primarily on subjective scales or questionnaire items. This study estimated a Gaussian graphical model (GGM) integrating psychometric and PSG indicators to examine their cross-modal conditional association structure.
METHODS: We analyzed data from 318 sleep center patients who underwent overnight PSG and routine psychometric assessment. Eighteen indicators were included: two Hamilton Anxiety Rating Scale factors, seven 24-item Hamilton Depression Rating Scale factors, four Dysfunctional Beliefs and Attitudes about Sleep Scale factors, and five PSG parameters comprising sleep efficiency, apnea-hypopnea index, arousal index, N3 percentage, and rapid eye movement sleep percentage. Redundancy screening removed one psychometric factor before network estimation, leaving 17 network nodes. The analyses included latent profile analysis with comparison of all available covariance parameterizations, an EBICglasso-estimated GGM, network comparison tests, and moderated network analysis.
RESULTS: Latent profile analysis identified five patient profiles (entropy = 0.897), including two small profiles with opposite patterns: one combined markedly elevated anxiety and depression indicators with near-average PSG values, and the other combined clearly abnormal respiratory and arousal indicators with lower psychometric scores. The estimated GGM contained 43 nonzero edges out of 136 possible, including three small cross-modal edges, all of which were negative (mean weight = -0.027). Psychic anxiety had the highest expected influence (1.134), whereas four of the five PSG indicators occupied the four lowest ranks. The case-dropping bootstrap yielded a correlation stability coefficient of.673. Psychic anxiety (-0.038) and log(AHI + 1) (-0.038) had the largest-magnitude bridge expected influence values within the psychometric and PSG modalities, followed by medication beliefs (-0.034). Network comparison tests did not detect statistically significant differences in network structure (M = 0.187, p = .808) or global strength (S = 0.304, p = .819) between the two largest profiles. No nonzero moderation parameter involving apnea-hypopnea index was retained by the regularized moderated network model.
CONCLUSIONS: In this cross-sectional clinical sample, the estimated network was characterized by predominantly stronger within-modality associations and limited cross-modal conditional associations between psychometric and PSG indicators. These findings suggest that psychometric and PSG assessments may capture partly non-overlapping and complementary aspects of patients' clinical presentations. However, the observed network structure does not establish mechanistic decoupling, causal direction, or independence between psychological and physiological processes. Longitudinal studies with closer temporal alignment and repeated PSG measurements are needed to determine the sources and clinical significance of these cross-modal associations.