Perspective: A resident’s role in promoting safe machine-learning tools in sleep medicine

INTRODUCTION

Machine-learning advancements offer promising transformations within the field of sleep medicine, and in-training physicians have the unique privilege and responsibility of shaping this rapidly evolving domain. Integrating these tools into the technologies relied upon in sleep medicine has great potential to support clinicians in the diagnosis and monitoring of sleep disorders.

Current research highlights the potential for machine-learning tools to enhance diagnostic availability and accuracy in sleep medicine. For example, Kelly et al¹ demonstrated that machine-learning–based analysis of mandibular movements can help diagnose obstructive sleep apnea with similar accuracy compared with traditional in-home polysomnography. With findings that suggest combining oximetry and airflow data for machine-learning analysis has the capacity to yield superior results compared with single-channel methods for obstructive sleep apnea diagnostic testing. The research by Alvarez et al² further supports the present value of technological developments in this field. As the application and breadth of these technologies continue to expand and evolve, understanding their implications and what we are integrating into our medical practices becomes indispensable.

GROUND TRUTHS AND KEY VARIABLES IN SLEEP MEDICINE

Establishing ground truths in sleep medicine is crucial for developing reliable tools that assist clinicians in patient care. Ground truths, influenced by expert opinions and carefully collected data, help machine-learning applications yield precise and clinically meaningful results.³

Polysomnograms currently enable the collection of extensive data, much of which are not routinely incorporated into diagnostic and therapeutic criteria for generating meaningful clinical outcomes. For instance, while sleep-disordered breathing is diagnosed and quantified by the number of apneas and hypopneas per hour of sleep, additional variables, such as the length of apneas and hypopneas, associated oxygen desaturations, and related arousals, may offer valuable insights into the clinical significance of the observed apneas and hypopneas.^4,5 Coupled with demographic and clinical information, such as the severity of daytime sleepiness, these data can be used tactically to help determine information related to the impact of sleep disturbances on individual patients. Additionally, growing interest focuses on the utility for machine learning to enhance access to medical studies like obstructive sleep apnea screening, aiming for reliable at-home detection by training models on smartphone audio and patients’ polysomnographic data.⁶ Analyzing data through these tools emphasizes the increasing potential for personalizing each patient’s health care experience.

Key variables to consider include sleep stages, sleep-related events, diagnostic and treatment response markers, physiological signals, demographic information, sleep questionnaires, and individualized sleep monitoring.⁷ By integrating these factors and comprehending their interrelationships, we can contribute to the development of comprehensive ground truths. This, in turn, leads to the creation of more accurate and reliable tools that assist in diagnosing and monitoring sleep disorders.

Electroencephalography and electromyography data, which are currently incorporated into polysomnography studies, can be reliably used for developing electroencephalography and electromyography standards that are currently lacking. For instance, limited information is available regarding electromyography changes in patients with rapid eye movement behavior disorder. However, the diagnosis of rapid eye movement behavior disorder depends strictly on the identification of electromyography changes during polysomnography, which is often related to the individual scorer and interpreter’s experience. By openly providing comprehensive information about data sources, methodologies, and evaluation metrics, researchers and clinicians can gain a better understanding of the basis for generated results, ultimately aiding in the assessment of a tool’s reliability.⁸

RIGOR, TRANSPARENCY, AND ETHICS

Establishing research standards that promote transparent machine-learning practices in sleep medicine is crucial for facilitating the safe integration of these tools into clinical practice.⁹ Some standards that are helpful to consider include data collection, preprocessing, expert consensus, manual scoring, annotation, interrater reliability assessment, cross-validation, and updates.¹⁰ It is essential to advocate for standardized protocols and benchmark datasets during this pivotal time, which will promote the development of reliable, accurate, and generalizable tools across diverse patient populations and clinical settings.¹⁰ Key ethical concerns to address include patient privacy, data security, and the potential for bias in algorithms. Ensuring the use of diverse and representative datasets, along with conducting regular audits of the algorithms, can help maintain accuracy, fairness, and generalizability in machine-learning applications within sleep medicine.

BUILDING FOUNDATIONS: UNDERSTANDING, COMMUNICATION, AND COLLABORATION

Building solid foundations for residents and trainees to foster safe machine-learning tools in sleep medicine can be divided into 3 core areas: understanding, communication, and collaboration. Acquiring a foundational understanding of machine learning is essential for residents. In addition to a basic grasp of relevant algorithms and computational processes, an appreciation for the broader impacts of these technologies is valuable. As the field rapidly evolves, residents will benefit by staying informed via a growing number of resources, including specialized journals, newly formed sections within established journals, podcasts, and online courses.⁹ Future advantages offered by machine learning, such as improved diagnostics and personalized treatments, come with challenges related to biases and privacy issues. As such, weaving machine-learning education into sleep medicine and other medical curricula, using established educational platforms such as journal clubs, grand rounds, and various forms of continued medical education, could help deliver a more comprehensive understanding of these dynamic issues.

There is utility in learning to effectively communicate machine-learning uses, benefits, and limitations to patients, peers, and other health care providers. Clear dialogues about the function of machine-learning tools, their advantages, and possible risks, coupled with active listening to address concerns, can help promote trust and understanding. Advocacy also plays a significant role, with residents being uniquely positioned to advocate for resources, training programs, ethical guidelines, and equitable access to future advancement in health care.

Finally, collaboration with computer scientists, data scientists, as well as other experts in machine learning is integral for those interested in the intersection of machine learning and sleep medicine. Practical steps include seeking relevant research opportunities and active engagement within health care communities that are developing and using these tools.

DISCUSSION

As in-training physicians, we have a unique opportunity to shape the future of machine learning in sleep medicine. By advocating for transparency, addressing key variables, and establishing solid ground truths, we can contribute to the development of reliable, accurate, and ethical tools. Embracing this role, fostering collaboration between the fields of computer science and sleep medicine will help promote innovative solutions that can transform patient care and ultimately improve the overall quality of sleep health.

DISCLOSURE STATEMENT

All authors have seen and approved this manuscript submission. Work for this study was performed at Lehigh Valley Fleming Neuroscience Institute. The authors report no conflicts of interest.

REFERENCES