💻 computer science

Beyond AHI: An Interpretable Causal-Discovery-Guided Framework for Sleep Recovery in Connected Health

This paper proposes an interpretable, causal-discovery-guided framework that derives a hierarchical Sleep Recovery Score from multimodal polysomnography data, demonstrating that this new metric aligns 2.5 times more strongly with patient-reported recovery outcomes than the traditional Apnea-Hypopnea Index (AHI) and is adaptable to connected health technologies.

Original authors: Saba A. Farahani, Elahe Khatibi, Manoj Vishwanath, Amir M. Rahmani, Hung Cao

Published 2026-06-19

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Saba A. Farahani, Elahe Khatibi, Manoj Vishwanath, Amir M. Rahmani, Hung Cao

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you wake up feeling groggy, tired, and like you didn't get a good night's rest. For decades, doctors have tried to figure out why by looking at a single number: the AHI (Apnea-Hypopnea Index). Think of the AHI like a "traffic jam counter" for your breathing. It simply counts how many times your airway gets blocked or your breathing gets shallow while you sleep.

The problem, as this paper points out, is that counting traffic jams doesn't tell the whole story of why you feel tired. You might have a few traffic jams but still feel great, or you might have a smooth road but feel exhausted because your car engine is overheating or your suspension is broken. The AHI is too simple; it misses the complex, multi-part machinery of your body.

The New Idea: A "Sleep Health Report Card"

The researchers from UC Irvine propose a new way to measure sleep recovery called the Sleep Recovery Score (SRS). Instead of looking at just one thing, they built a "report card" that grades five different aspects of your sleep physiology.

Think of your sleep like a symphony orchestra.

The Old Way (AHI): The conductor only listens to the drums (breathing). If the drums are off, they say the whole concert is bad.
The New Way (SRS): The conductor listens to the entire orchestra: the drums (breathing), the strings (oxygen levels), the woodwinds (how often you wake up), the brass (brain wave patterns), and the percussion (heart rate control).

How They Built It: The "Detective" Approach

The team didn't just guess which parts of the orchestra mattered. They used a smart, step-by-step detective process to figure out which instruments actually caused the audience (the patient) to feel tired.

The Data: They looked at massive amounts of data from two large groups of people (over 2,000 people total) who wore sensors all night to record everything from brain waves to heartbeats.
The Map (Causal Discovery): They used a computer algorithm to draw a "map" of connections. Imagine trying to figure out who is talking to whom at a noisy party. The algorithm tried to draw lines connecting different body signals to how tired the person felt the next day.
The Filter (The "Two-Stage Screening"): This is the most creative part. The computer map sometimes draws silly lines (like connecting "eye color" to "tiredness"). To fix this, they used two filters:
- Filter 1 (The Biologist): A human expert check to make sure the connections make biological sense.
- Filter 2 (The AI Auditor): They used a specialized Artificial Intelligence (LLM) to act like a strict editor. The AI checked every connection and asked: "Is this a real cause, or is it just a coincidence? Is this variable overlapping with the answer we are trying to find?" If the AI said "No," that connection was cut.

The Five "Instruments" That Matter

After all the filtering, the researchers found that five specific areas consistently explained why people felt tired or rested. These are the five sections of the sleep orchestra:

Respiratory Burden: How hard your breathing works (the drums).
Hypoxic Burden: How much your oxygen drops (the strings).
Sleep Fragmentation: How often you wake up or get disturbed (the woodwinds).
Sleep Architecture: The structure of your deep sleep and dream cycles (the brass).
Autonomic Regulation: How your heart rate and nervous system behave (the percussion).

The Results: Why It's Better

When they tested this new "Report Card" (SRS) against the old "Traffic Jam Counter" (AHI), the results were clear:

The SRS was much better at predicting how tired people actually felt. In some cases, it was 2.5 times better at matching the patient's experience than the AHI.
The AHI often failed to show a connection to how tired people felt. It was like a traffic counter that couldn't explain why a driver was stressed.

The Big Picture

This paper doesn't just say "we have a better number." It says, "Sleep recovery isn't a single problem; it's a team effort."

By breaking sleep down into these five understandable parts, the new score is interpretable. This means a doctor (or a future smart watch) can look at the score and say, "Your breathing is fine, but your oxygen is dipping too low," or "Your heart rate is too high at night."

The researchers emphasize that while this score was built using hospital-grade equipment, the logic fits perfectly with the wearable devices people are starting to use today (like smartwatches that track heart rate and oxygen). It provides a blueprint for how future technology can move beyond simple counting and start understanding the complex, multi-layered story of why we feel rested or tired.

Technical Summary: Beyond AHI: An Interpretable Causal-Discovery–Guided Framework for Sleep Recovery in Connected Health

Problem Statement
Current clinical sleep assessment relies heavily on the Apnea–Hypopnea Index (AHI) and other summary indices derived from polysomnography (PSG). While these metrics are useful for diagnosing specific sleep disorders, they often fail to capture the integrated physiological burden underlying functional recovery. Consequently, there is a persistent disconnect between objective sleep physiology and patient-reported outcomes (PROs) such as daytime sleepiness, fatigue, and perceived sleep quality. Furthermore, existing machine learning approaches for PSG analysis often function as "black boxes" optimized for predictive association rather than mechanistic interpretability, limiting their utility in clinical trust and translation to connected health technologies. The challenge lies in translating continuous, multimodal biosignals into clinically interpretable recovery metrics that reflect multidomain physiology rather than isolated event counts.

Methodology
The authors propose a causal-discovery–guided framework to derive a hierarchical Sleep Recovery Score (SRS) from multimodal PSG data. The framework operates through a structured five-step pipeline applied to two independent population cohorts: MESA ( $n=1,540$ ) and MrOS ( $n=825$ ).

Feature Construction & Outcome Selection: The study utilizes a broad set of physiological features spanning five domains: respiratory burden, hypoxic burden, sleep fragmentation, sleep architecture, and autonomic regulation. These are linked to recovery-related PROs (e.g., Epworth Sleepiness Scale, fatigue, perceived sleep quality).
Causal Graph Estimation: Using the linear NOTEARS algorithm, the authors estimate a sparse Directed Acyclic Graph (DAG) to identify candidate physiological drivers for each PRO. To ensure robustness, they employ bootstrap stability selection (500 resamples), retaining only edges with a selection frequency $\ge 0.6$ .
Two-Stage Candidate Screening: To address the limitations of observational data and ensure physiological plausibility, a rigorous screening process is applied:
- Stage 1 (Physiology-based): Edges inconsistent with established sleep mechanisms (e.g., removing demographic proxies like age or race that act as structural confounders) are filtered out.
- Stage 2 (Constrained LLM-assisted Auditing): Remaining candidates undergo a structured classification protocol using a constrained Large Language Model (LLM). Features are classified as plausible mechanistic drivers, structural confounders, or construct-overlapping leakage variables (e.g., subjective measures correlated by construction with the target). Only features classified as plausible drivers by both layers advance.
Consensus Aggregation: Mechanisms are retained only if they appear in at least two different recovery outcomes, reducing outcome-specific noise.
Hierarchical Score Synthesis: Retained features are grouped into the five physiological domains. Domain scores are computed as weighted sums of normalized features, where weights are proportional to edge magnitude and stability. The final SRS is a hierarchical aggregation of these domain scores, standardized within each cohort.

Key Contributions

Interpretable Multidomain Modeling: The paper introduces a hierarchical SRS that moves beyond single respiratory indices (like AHI) by explicitly linking multimodal PSG to patient-reported recovery outcomes through a transparent, domain-structured framework.
Causal-Discovery-Guided Mechanism Identification: The authors apply DAG-based causal discovery refined by physiology-based constraints and constrained LLM-assisted auditing to extract recurrent, physiologically interpretable recovery domains.
Cross-Cohort Evidence for Connected Health: The framework demonstrates consistent convergence on five physiological domains across two distinct cohorts, showing stronger alignment with perceived recovery than AHI. This supports the translation of recovery modeling to distributed and wearable sensing systems (e.g., wearable ECG, oximetry) that capture subsets of these domains.

Results

Mechanistic Convergence: Across both MESA and MrOS, the framework consistently identified five physiological domains associated with impaired recovery: respiratory burden, hypoxic burden, sleep fragmentation, sleep architecture (including EEG dynamics), and autonomic regulation.
Feature Distillation: In MESA, the process distilled 53 initial features down to 32 consensus mechanisms. In MrOS, 174 features were reduced to 28. Structural covariates (age, sex, race) and construct-overlapping variables were successfully excluded.
Association with Outcomes: The SRS demonstrated statistically significant associations with four out of five recovery outcomes, whereas AHI reached significance for only one.
- In MESA, SRS showed significant correlations with the Epworth Sleepiness Scale ( $\rho = 0.098$ ) and daytime sleepiness ( $\rho = 0.074$ ), while AHI was non-significant for both.
- In MrOS, SRS significantly correlated with perceived sleep quality ( $\rho = 0.107$ ) and trouble staying asleep ( $\rho = 0.111$ ), while AHI was non-significant.
Performance Gain: The SRS showed up to 2.5× stronger alignment with perceived recovery outcomes compared to AHI in specific cases (e.g., perceived sleep quality in MrOS).

Significance and Claims
The paper claims that restorative sleep is better characterized as a multidomain physiological construct rather than a single respiratory index. The proposed framework does not seek to replace disease-specific metrics like AHI but rather complements them by capturing the broader physiological substrate of recovery. By explicitly linking multimodal sleep physiology to patient-centered outcomes through an interpretable, bias-aware, and domain-structured framework, the work provides a practical foundation for recovery modeling in both clinical sleep studies and emerging smart/connected health settings. The authors emphasize that the SRS allows for the decomposition of recovery drivers (e.g., distinguishing hypoxic burden from autonomic dysregulation), enabling targeted interventions that traditional AHI-based assessments cannot support.

Note on Limitations: The authors modestly acknowledge that the analysis is cross-sectional and does not establish temporal causality. They note that the linear NOTEARS formulation may miss nonlinear interactions and that the LLM-assisted auditing, while consistent, relies on constrained prompts rather than direct domain-expert review. Correlations with subjective outcomes were modest, reflecting the multifactorial nature of perceived recovery.