Novel digital markers of sleep dynamics: causal inference approach revealing age and gender phenotypes in obstructive sleep apnea

Abstract

Despite evidence that sleep-disorders alter sleep-stage dynamics, only a limited amount of these parameters are included and interpreted in clinical practice, mainly due to unintuitive methodologies or lacking normative values. Leveraging the matrix of sleep-stage transition proportions, we propose (i) a general framework to quantify sleep-dynamics, (ii) several novel markers of their alterations, and (iii) demonstrate our approach using obstructive sleep apnea (OSA), one of the most prevalent sleep-disorder and a significant risk factor. Using causal inference techniques, we address confounding in an observational clinical database and estimate markers personalized by age, gender, and OSA-severity. Importantly, our approach adjusts for five categories of sleep-wake-related comorbidities, a factor overlooked in existing research but present in 48.6% of OSA-subjects in our high-quality dataset. Key markers, such as NREM-REM-oscillations and sleep-stage-specific fragmentations, were increased across all OSA-severities and demographic groups. Additionally, we identified distinct gender-phenotypes, suggesting that females may be more vulnerable to awakenings and REM-sleep-disruptions. External validation of the transition markers on the SHHS database confirmed their robustness in detecting sleep-disordered-breathing (average AUROC = 66.4%). With advancements in automated sleep-scoring and wearable devices, our approach holds promise for developing low-cost screening tools for sleep-, neurodegenerative-, and psychiatric-disorders exhibiting altered sleep patterns.

Keywords: Causal inference; Digital markers; Dirichlet regression; Obstructive sleep apnea; Polysomnography; Sleep disorders; Sleep dynamics.

Fig. 1

Graphical overview of the implemented approach for quantifying sleep-stage dynamics. Part (a): The study utilized observational data, including hypnograms of subjects with a conclusive diagnosis of either obstructive sleep apnea (OSA) or healthy status. The illustration highlights differences in the overall prevalence of OSA (OSA-affected > healthy) concerning gender (male predominance in OSA), age (higher OSA prevalence in older subjects), and comorbidities (not present in healthy subjects). Part (b): Inverse Probability Weighting (IPW) is applied to balance the data for the primary confounders of age and gender, having distributional overlap between OSA and healthy subjects. Part (c): A sleep fingerprint matrix of sleep-stage transition proportions is modelled using Dirichlet regression within a causal S-Learner framework to capture the effects of OSA, its severity (Apnea-Hypopnea Index, AHI), age, gender, and comorbidities. Part (d): the framework quantifies digital markers of OSA (raw , as the normalized Markovian , and derived quantities such as sleep fragmentation), personalized for subjects’ demographics, OSA severity, and comorbidities, and presented in terms of conditional average treatment effect (CATE) and risk-ratio CATE (RR-CATE).

Fig. 2

Heatmap of risk-ratio conditional-average-treatment-effects (RR-CATE) of OSA (compared to a matched healthy subject) on individual dimensions of sleep-fingerprint matrix of sleep-stage transition proportions, per gender (F, M), age (A1, A2, A3), and OSA-severity (O1, O2, O3). Decreased (i.e., RR < 100%) and increased (i.e., RR > 100%) risk-ratios are depicted with red and green shaded backgrounds, respectively. Significant effects are in bold and highlighted with a star (*).

Fig. 3

Heatmap of risk-ratio conditional-average-treatment-effects (RR-CATE) of OSA (compared to a matched healthy subject) on PSG-markers derived from matrix of sleep-stage transition proportions, per gender (F, M), age (A1, A2, A3), and OSA-severity (O1, O2, O3). Decreased (i.e., RR < 100%) and increased (i.e., RR > 100%) risk-ratios are depicted with red and green shaded backgrounds, respectively. Significant effects are in bold and highlighted with a star (*).

Fig. 4

Effects of age and OSA-severities on NREM-REM oscillations, , in females. The left plots (1a, 2a) depict expected probabilities for varying age with fixed AHI = 30, and for varying AHI with fixed age = 30. Based on that, the central (1b, 2b) and right (1c, 2c) plots depict age- and AHI-related CATE and RR-CATE.

Fig. 5

Heatmap of risk-ratio conditional-average-treatment-effects (RR-CATE) of OSA (compared to a matched healthy subject) on individual dimensions of row-normalized Markovian transition matrix , per gender (F, M), age (A1, A2, A3), and OSA-severity (O1, O2, O3). Decreased (i.e., RR < 100%) and increased (i.e., RR > 100%) risk-ratios are depicted with red and green shaded backgrounds, respectively. Significant effects are in bold and highlighted with a star (*).