Dynamic Structural Causal Modeling for Sleep

Ranveer Singh, Saurabh Mathur, Pranuthi Tenali, Arun Badi, Sriraam NatarajanarXiv 2026 · arXiv

TL;DR

This paper applies dynamic structural causal modeling to Home Sleep Apnea Test (HSAT) data to map the hidden physiological mechanisms of sleep-disordered breathing. By stratifying data by age and sex, the authors reveal that while core respiratory relationships remain universal, other causal links vary significantly—even completely reversing direction between younger and older cohorts.

What Problem Does This Paper Address?

Sleep-disordered breathing (SDB), such as sleep apnea, is a complex and dynamic condition linked to cardiovascular and metabolic risks. Because SDB manifests differently from person to person and changes throughout the night, researchers want to build Dynamic Data-Driven Application Systems (DDDAS) that can adaptively support clinical decisions in real time.

However, building an effective DDDAS requires an accurate causal model of how respiratory events interact. The current gold standard for sleep tracking, polysomnography (PSG), is too expensive and intrusive for large-scale causal modeling. While Home Sleep Apnea Tests (HSATs) offer a cheaper, more accessible alternative, they are traditionally used only for basic diagnostic scoring rather than uncovering underlying physiological mechanisms. Furthermore, learning stable causal graphs from small, noisy medical datasets is notoriously difficult, often resulting in ambiguous or incorrect relationships.

Methods

The authors introduce a framework that combines time-series causal discovery with domain expertise and robust aggregation techniques:

  1. Dynamic Structural Causal Models (DSCMs): Used to model temporal causal processes where a variable's current state depends on past and contemporaneous factors.
  2. PCMCI+ Algorithm: A state-of-the-art time-series causal discovery method used with partial correlation tests to identify both lagged and simultaneous dependencies.
  3. Domain Knowledge Constraints: To resolve ambiguities (Markov equivalence classes), the authors integrated expert clinical knowledge to "blacklist" physically impossible or forbidden causal edges.
  4. Higher-Order Bootstrap Aggregation: Instead of simply averaging individual edges across bootstrapped samples—which can be highly unstable in low-data regimes—the authors employed General Model Averaging (GMA). This technique aggregates stable local subgraphs (like chains, forks, and colliders) to construct a more reliable final graph.

Datasets

The study utilized a clinical dataset consisting of 105 HSAT recordings, each lasting at least two hours.

  • Demographics: The cohort included 58 males and 47 females. Patients were stratified into "young" (males < 40, females < 50) and "old" subcohorts.
  • Features: Continuous signals for Snoring, Pulse, Oxygen Saturation, Effort, and Flow were processed into 10-second windows to calculate fractional variables (e.g., the fraction of time spent in an apnea event or desaturation episode).
Fig. 2: Age distribution for the patients based on sex
Constructed Windowed Features as a fraction of the analysis window

Key Insights and Ideas

  • Universal vs. Subcohort-Specific Dynamics: Temporal self-dependencies (a feature predicting its own future state) and the direct causal link from Apnea Hypopnea Events ($A$) to Oxygen Desaturation ($D$) are robust and persist across all patient groups.
  • The Age Reversal: Most surprisingly, the causal direction between Apnea ($A$) and Snoring ($S$) flips with age. In older adults, apnea events causally drive snoring ($A_t \rightarrow S_t$). In younger adults, the relationship reverses ($S_t \rightarrow A_t$), suggesting that the mechanical triggers of sleep apnea evolve as patients age.
  • Sex-Based Discrepancies: A distinct causal relationship between Snoring ($S$) and Inspiratory Flow Limitation ($I$) was discovered in the female subcohort but was entirely absent in males.

Key Results

The framework successfully mapped distinct, unambiguous causal graphs for the entire population as well as individual subcohorts.

The Average Time Series Causal Graph across the different subcohorts and the entire cohort
  • For the female subcohort, the local structure $S_{t-1} \rightarrow S_t \leftarrow I_{t-1}$ was confidently admitted with a posterior value of 0.10 (compared to an unstable 0.04 in males).
  • For the older subcohort, the structure $A_t \rightarrow S_t \leftarrow S_{t-1}$ achieved a posterior value of 0.35, whereas the younger subcohort favored the reversed structure $A_{t-1} \rightarrow A_t \leftarrow S_t$ with a posterior of 0.24.
  • These stark structural differences highlight why a single, population-wide model is insufficient for personalized clinical decision support systems.

Limitations and Future Work

While promising, the authors note a few key limitations:

  • Data Constraints: HSATs do not capture all relevant physiological variables, notably missing sleep stages (REM vs. non-REM) and cortical arousals, which are standard in full PSG tests.
  • Stationarity Assumption: The current model assumes that causal relationships remain constant throughout the night. In reality, sleep dynamics shift as patients cycle through different sleep stages.

Future Directions: The team plans to expand the feature set and explore non-stationary causal discovery methods (such as RPCMCI) to better capture the evolving nature of sleep architecture across the night.

Summary updated

Abstract

The causal dynamics of sleep-disordered breathing are complex and vary across patient populations, hindering the development of targeted interventions. We learn dynamic causal graphs of sleep-disordered breathing from Home Sleep Apnea Test (HSAT) recordings, revealing systematic differences in causal structure across sex and age subcohorts. We do so using the PCMCI+ algorithm on windowed fractional variables derived from 105 HSAT recordings, exploiting domain knowledge via edge blacklisting and employing bootstrap aggregation to address small subcohort sizes. The learned graphs show that temporal self-dependencies and the apnea-desaturation relationship persist across all cohorts, while other relationships vary substantially.
BibTeX
@misc{singh2026dynamic,
  author = {Ranveer Singh and Saurabh Mathur and Pranuthi Tenali and Arun Badi and Sriraam Natarajan},
  title = {{Dynamic Structural Causal Modeling for Sleep}},
  year = {2026},
  eprint = {2608.20285},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2608.20285},
}
Figure