Tai Dinh

Sleep-disordered breathing (SDB), particularly obstructive sleep apnea (OSA), is a highly prevalent group of disorders associated with disrupted sleep, intermittent hypoxemia, excessive daytime sleepiness, neurocognitive impairment, and cardiometabolic morbidity. Although laboratory polysomnography remains central to diagnosis, conventional sleep testing can be costly, labor-intensive, and difficult to scale. Recent advances in data mining, machine learning, deep learning, wearable sensing, mobile health, multimodal analytics, and large language models are creating new opportunities for more accessible, personalized, and efficient SDB care. This review examines recent advances in artificial intelligence for SDB diagnosis and management, with emphasis on polysomnography and home sleep apnea testing, wearable and contactless screening, audio and physiological-signal analysis, imaging and anatomical risk prediction, phenotyping and endotyping, treatment monitoring, continuous positive airway pressure adherence, and clinically deployable explainable AI. We also discuss the emerging role of large language models in patient education, sleep-report summarization, clinical documentation, and decision-support workflows. Finally, we identify major challenges related to data heterogeneity, external validation, interpretability, privacy, fairness, reproducibility, regulation, and real-world integration into sleep medicine. This review aims to provide an evidence-based framework for researchers and clinicians to understand current AI applications and identify future directions for precision diagnosis and management of SDB.

Rapid advancements in artificial intelligence (AI), in combination with increased availability of rich large-scale clinical data, has paved the way for promising implementations in both diagnosing/subtyping as well as managing of sleep disordered breathing (SDB). A central strength of AI in this regard is how it facilitates analysis of complex multidimensional/modal data, for example pertaining to comorbidities and so-called treatable traits. However, the utility of such applications remains somewhat limited, as most AI models are so-called “black boxes”, for which it is not intuitive to assess how an output was arrived at. This poses challenges for ensuring patient trust and controlling bias. Explainable AI (XAI), which constitutes techniques to either distill “black box” models to simpler ones or to elucidate influential patterns by surrogate “white box” models, may provide such insights. Adoption of XAI within this field, however, is still at a nascent stage, and overall, the optimal role of AI alongside clinicians continues to be unclear. This review provides a clinically oriented overview of state-of-the-art implementations of AI and XAI in SDB. In addition, we discuss limitations and trade-offs with XAI and propose a general framework for personalized management of obstructive sleep apnea using XAI and treatable traits.

Large language models (LLMs) have rapidly evolved from general-purpose text generators into widely deployed foundations for reasoning, summarization, and decision support across many disciplines. This paper presents a systematic review of 655 LLM-related studies published between 2022 and 2024, organized into ten subject areas. Across domains, LLMs frequently serve as an "intelligence core" enabling conversational systems, personalization and recommendation, sentiment and review analytics, thematic and transcript analysis, knowledge extraction, and professional decision-making. We also discuss key challenges and outline emerging trends and opportunities for LLM-based research.

Sequential decisions in volatile, high-stakes settings require more than maximizing expected return; they require principled uncertainty management. This paper presents the Uncertainty-Aware Markov Decision Process (UAMDP), a unified framework that couples Bayesian forecasting, posterior-sampling reinforcement learning, and planning under a conditional value-at-risk (CVaR) constraint. In a closed loop, the agent updates its beliefs over latent dynamics, samples plausible futures via Thompson sampling, and optimizes policies subject to preset risk tolerances. We establish regret bounds that converge to the Bayes-optimal benchmark under standard regularity conditions. We evaluate UAMDP in two domains including high-frequency equity trading and retail inventory control, both marked by structural uncertainty and economic volatility. Relative to strong deep learning baselines, UAMDP improves long-horizon forecasting accuracy (RMSE decreases by up to 25% and sMAPE by 32%), and these gains translate into economic performance: the trading Sharpe ratio rises from 1.54 to 1.74 while maximum drawdown is roughly halved. These results show that integrating calibrated probabilistic modeling, exploration aligned with posterior uncertainty, and risk-aware control yields a robust, generalizable approach to safer and more profitable sequential decision-making. From an AI perspective, we formalize a Bayes-adaptive, CVaR-constrained sequential decision framework and provide regret guarantees for posterior-sampling control under latent dynamics. From an engineering standpoint, we show how this risk-aware framework translates into improved operational decisions in volatile trading and inventory systems, yielding better risk-adjusted performance in both domains.