Multimodal Early Fusion for Obstructive Sleep Apnea Phenotyping and Risk Stratification

Obstructive sleep apnea (OSA) is a sleep-related breathing disorder caused by upper airway blockages, resulting in disrupted sleep and reduced blood oxygen levels. Despite OSA affecting many adults in the US and worldwide, it remains difficult to diagnose and treat. OSA is also associated with other serious diseases such as Heart Failure (HF). There is a lot of untapped data available, especially across different data modalities, that we want to use to improve personalized care. We combined the electronic health records of OSA and HF patients in the Medical Information Mart for Intensive Care-IV (MIMICIV) with the BERTopic embeddings taken from the discharge notes. The combined data was run through an autoencoder with a classifier head to improve the understanding of OSA so we could find patient subgroups with different risk levels. We found there are benefits to combining the data modalities, including revealing new subgroups. We also found OSA subgroups with meaningful differences. This information could aid healthcare professionals in caring for and treating these patients