Enhancing Cardiovascular Risk Prediction and Treatment Response in Sleep Apnea Patients Using Advanced AI-based Transformer Models in Multi-Modal Sleep Data

Obstructive sleep apnea (OSA) affects arounds 24 million Americans and increases cardiovascular (CV) morbidity and mortality. The primary treatment for OSA is continuous positive airway pressure (CPAP) therapy, but CPAP usage has not been shown to reduce CV events in clinical trials. One major limitation to our understanding of the OSA-CV disease link is the current diagnostic and prognostic marker of OSA: the apnea- hypopnea index (AHI). AHI does not well predict individual CV disease risk, and reducing AHI to normal levels does not improve CV disease outcomes. Thus, better CV risk prediction tools are needed for OSA patients. Further, being at high risk does not always translate to significant benefits from treatment. However, there is currently no way of forecasting which OSA patients will receive the most benefit, or harm, from CPAP therapy. Patients with suspected OSA undergo a comprehensive sleep study, known as a polysomnogram, either at home or in a lab. Polysomnograms collect up to 20 channels of data, including heart rate, muscle movements, and brain activity; but nearly none of this data is used in clinical practice and what data is used is often compressed into summary statistics (e.g., average heart rate over 8 hours). The goal of this proposal is to use a novel artificial intelligence technique known as transformer-based neural networks, or transformers, to analyze the multimodal, longitudinal data available from a polysomnogram in order to better predict CV risk and treatment response in OSA patients. To accomplish this, we will leverage data from existing diverse epidemiological datasets (Aim 1) and randomized clinical trials testing CPAP versus usual care on CV disease outcomes (Aim 2) to fine-tune our pre-trained sleep specific transformer model. We will compare our risk prediction (Aim 1) and treatment response (Aim 2) tools to clinical metrics such as the AHI to demonstrate their improved predictive utility. Our approach is innovative in its use of cutting-edge artificial intelligence and estimation of treatment heterogeneity techniques. Transformers have become a foundational and transformative technology in artificial intelligence due to their exceptional ability to handle sequential data and capture complex patterns. Their influence on the artificial intelligence landscape is profound, and they continue to drive innovations and improvements in artificial intelligence research and technology. Further, our proposal has high significance given the prevalence of OSA, and the lack of available tools to predict CV risk and CPAP treatment response specifically in OSA patients. In particular, a tool that could predict whether an OSA patient will actually receive benefit from CPAP would revolutionize the field. Our approach will improve clinical risk prediction, treatment guidelines, and patient outcomes, as well as possibly extract novel health-relevant features of sleep for future clinical applications and mechanistic insight into OSA, CV disease, and the link between the two, significantly impacting the field of sleep medicine. Project Number: 1R01HL175992-01A1 | Fiscal Year: 2025 | NIH Institute/Center: National Heart Lung and Blood Institute (NHLBI) | Principal Investigator: Girish Nadkarni (+3 co-PIs) | Institution: ICAHN SCHOOL OF MEDICINE AT MOUNT SINAI, NEW YORK, NY | Award Amount: $816,472 | Activity Code: R01 | Study Section: Special Emphasis Panel[ZRG1 HSS-A (90)] View on NIH RePORTER: https://reporter.nih.gov/project-details/1R01HL17599201A1