Risk groups derived from a polysomnography foundational model predicted all-cause mortality (RG5 vs RG1 HR 1.58; 95% CI 1.07-2.32) and incident heart failure in an independent external cohort.
Cohort (n=10,000)
Yes
Does a RoBERTa-based foundational model applied to raw polysomnography signals predict all-cause mortality and incident heart failure?
A foundational AI model applied to raw polysomnography signals successfully clustered patients into risk groups that robustly predicted all-cause mortality and incident heart failure across different acquisition protocols.
Effect estimate: HR 1.58 (95% CI 1.07-2.32)
p-value: p=<0.001
Abstract Introduction Polysomnography (PSG) reports rely on a handful of summary indices, such as the apnea-hypopnea index (AHI), which incompletely capture the multidimensional burden of sleep-disordered breathing and its downstream health risks. We developed a risk stratification pipeline that uses RoBERTa-based foundational model on raw PSG signals to generate patient-level embeddings and cluster individuals into risk-groups (RG1-RG5). Here, we perform external, independent validation of these clusters for predicting long-term cardiovascular outcomes using the publicly available community-based Sleep Heart Health Study (SHHS) dataset. Methods We fine-tuned a RoBERTa-based model on 10,000 clinical PSG recordings from Cleveland Clinic (Jan-2012 to Dec-2022) to match PSG scoring data (e.g, sleep stages and respiratory events). Latent embeddings from the model were aggregated at the patient-level and clustered using k-means with energy-distance criteria to define five risk-groups. The trained model and fixed cluster centroids were then applied without retraining to the raw PSG data in SHHS. Primary outcomes were all-cause mortality and incident heart failure. Kaplan-Meier analyses and Cox proportional hazards models, adjusted for age and sex, were performed in both datasets. Results In the clinical cohort, higher risk-groups showed significantly reduced survival for all-cause mortality (log-rank p 0.001 for RG3-RG5 vs. RG1), with age/gender-adjusted hazards demonstrating a dose-response gradient: RG3 HR=1.51 (95% CI 1.17-1.96), RG4 HR=1.64 (95% CI 1.26-2.15), and RG5 HR=2.71 (95% CI 1.93-3.81), all versus RG1. For incident heart failure, RG5 had a significantly elevated risk (HR=1.67, 95% CI 1.12-2.48, p=0.011 vs. RG1). External validation in SHHS replicated these patterns despite differing PSG setups. For all-cause mortality, RG4 and RG5 remained associated with worse survival (log-rank p 0.001; HR=1.31, 95% CI 1.07-1.62, and HR=1.58, 95% CI 1.07-2.32, respectively, vs. RG1). For incident heart failure, RG5 again emerged as a high-risk group (log-rank p=0.003; HR=2.13, 95% CI 1.11-4.09, p=0.023 vs. RG1). Conclusion Patient clusters derived from a fine-tuned PSG foundational model showed robust external validity for predicting all-cause mortality and incident heart failure in an independent cohort with substantially different PSG acquisition protocols. These findings support the generalizability and clinical utility of model-based PSG phenotyping for risk stratification. Support (if any) This project was partially supported by NIH 1R21HL170206-01.
Araujo et al. (Fri,) conducted a cohort in Sleep-disordered breathing (n=10,000). RoBERTa-based foundational model risk groups (RG1-RG5) vs. Risk Group 1 (RG1) was evaluated on All-cause mortality (HR 1.58, 95% CI 1.07-2.32, p=<0.001). Risk groups derived from a polysomnography foundational model predicted all-cause mortality (RG5 vs RG1 HR 1.58; 95% CI 1.07-2.32) and incident heart failure in an independent external cohort.