Sensitive detection of unopposed estrogen using estrogen receptor functionalized nanoprobes for cancer diagnosis.

Sleep is fundamental to health and well-being. Its disruption by common disorders diminishes quality of life and increases the risk of cardiovascular, neurologic, and psychiatric disease. Sleep apnea alone affects nearly one billion people worldwide, underscoring the scope of this public health burden. Yet, assessment of sleep integrity remains limited because clinical interpretation of polysomnography (PSG), the gold-standard diagnostic test, is often reduced to the apnea-hypopnea index (AHI). While AHI is central to current diagnostic criteria for sleep-disordered breathing as defined by the American Academy of Sleep Medicine (AASM ICSD-3-TR), it captures only a narrow slice of sleep physiology and carries limited prognostic value. Identifying effective ways to incorporate the full richness of PSG time-series data, together with electronic medical records (EMRs), holds the potential to move beyond these constraints and deliver scalable innovations in precision sleep medicine. Attempts to overcome these limitations have included cluster analysis of PSG-derived indices, such as AHI, arousal index, and sleep stage proportions. While informative, these approaches rely on technologist-scored aggregate measures that obscure the dynamic complexity of sleep physiology. Other efforts have focused on specific physiologic burdens, such as hypoxic burden or heart rate arousal response, which may capture particular aspects of sleep-disordered breathing but remain narrowly focused and fail to capture the broader physiologic landscape of sleep. Artificial intelligence (AI) has shown promise in revealing hidden disease phenotypes in other areas of medicine, such as metabolic subtypes in diabetes and clinical subgroups in Alzheimer’s disease. These traditional machine learning models often depend on hand-crafted features and task-specific designs, limiting their ability to generalize across populations or adapt to new datasets. Foundation models, a class of AI architectures trained on large, diverse datasets, offer a scalable alternative. By learning broad, general-purpose physiologic representations, they can be adapted with minimal additional training to downstream tasks. Extending this methodology to PSG data, however, presents unique challenges, since sleep recordings consist of multimodal time-series signals that vary in length and sampling frequency. Developing effective methods to represent and tokenize such data remains an open area of research.