Translating Algorithms Into Action: Assessing AI Readiness in the Sleep Medicine Community
December 4, 2025
By: Ashesha Mechineni, MD, and Anita Rajagopal, MD, FCCP
Nonrespiratory Sleep Section, Sleep Medicine Network
The integration of artificial intelligence (AI) into modern medicine is accelerating, and sleep medicine is no exception. With abundant physiologic and clinical data, this field is particularly well positioned for the development of decision-support tools and predictive models. Yet, before adopting these innovations, it is critical to examine the readiness of clinicians and care teams to both embrace their potential and understand their limitations.
Vocabulary and conceptual literacy
The first determinant of readiness is literacy in AI terminology. At present, AI literacy within medicine remains in its infancy. Core terms form the foundation of understanding: an algorithm may be viewed as analogous to a clinical protocol; a model represents the application of an algorithm to data; training refers to iterative exposure to data that enables the model to “learn”; inference denotes application of the model to new, unseen data; features represent input variables; and labels are the corresponding outcomes.
Models are typically constructed with specific objectives. Traditional approaches such as regression (risk calculators) and decision trees (clinical flowcharts) are familiar to most clinicians. More contemporary models include random forests, which consist of decision trees that reduce bias through aggregation; support vector machines, which optimize separation between groups; and large language models, which are trained on vast corpora and generate contextually coherent responses.
Equally important are the learning paradigms underpinning these models: supervised learning, in which models are trained on labeled datasets; unsupervised learning, where hidden patterns are uncovered without labels; reinforcement learning, which optimizes performance through trial and error; and deep learning, a multilayer neural network approach applicable across paradigms. These distinctions are crucial in evaluating applications such as polysomnography scoring, OSA phenotyping, or risk prediction for insomnia and circadian rhythm disorders.
Perceived benefits and concerns
The second factor shaping readiness is the perceived clinical utility of AI. AI excels at recognizing complex patterns, processing large volumes of data, and extracting features imperceptible to human analysis. However, limitations remain, particularly the absence of empathy, context sensitivity, and accountability. These gaps underscore the need to integrate human judgment with machine-driven insights.
Measuring these perceptions requires validated instruments. The Medical Artificial Intelligence Readiness Scale for Medical Students evaluates self-perceived competence across cognition, ability, vision, and ethics. Broader instruments, including Likert scale surveys, capture attitudes and behavioral responses toward AI adoption. Of note, the Threats of Artificial Intelligence scale quantifies perceived risks along three dimensions: performance threat, security and privacy threat, and societal or existential threat. Together, these tools illuminate both enthusiasm for and resistance to AI integration in health care.
Adaptive capacity and system readiness
The third determinant of readiness is the adaptive capacity of individuals and institutions. AI systems evolve rapidly, with new tools and applications emerging at a pace that often outstrips implementation capacity. The medical community has previously experienced this challenge, particularly with electronic medical record adoption. Future success will depend on the robustness of infrastructure, system-wide resources, and the willingness of clinicians to adapt their workflows to novel technologies.
Implications for sleep medicine
AI has already demonstrated tangible benefits in sleep medicine. Applications include automated scribe technologies for clinical documentation, population-based risk screening through questionnaires, data-driven disease phenotyping using polysomnography, objective monitoring of treatment response, and patient-facing educational tools.
Nonetheless, as Dr. Richard M. Schwartzstein of Harvard Medical School aptly observed, AI remains a tool, not a team member.1 The absence of accountability and human judgment limits its role in medico-legal decision-making and scientific authorship. These realities highlight unresolved questions: Can AI ever serve as a legitimate reference in scholarly work or as admissible evidence in legal proceedings? Such debates will shape the trajectory of AI adoption in medicine.
Wrapping it up
Readiness for AI in sleep medicine is multifactorial, hinging on vocabulary literacy, perception of clinical benefit, and adaptive capacity. A balanced approach—embracing AI’s strengths while acknowledging its limitations—will be essential for meaningful integration. As the field advances, structured readiness assessments and ongoing dialogue will ensure that AI augments, rather than supplants, the art and science of sleep medicine.
Reference
1American Thoracic Society. ATS Breathe Easy – Lessons Learned from AI in Medical Education. Published September 2, 2025. YouTube. https://www.youtube.com/watch?v=m5Pkxcd97ho