By Drs. Mattina Davenport, Scott Ryals, Trung Le and Vidya Krishnan, on behalf of the AASM Artificial Intelligence in Sleep Medicine Committee
Artificial intelligence (AI) has been increasingly highlighted as a potential tool to enhance population-level sleep health and address gaps in sleep care. There is growing interest in the potential for AI to raise community awareness about sleep health, support the prevention of sleep disorders, improve sleep surveillance and expand access to sleep care. However, the empirical evidence supporting these possibilities remains limited. As AI technologies continue to evolve, careful attention will be needed to ensure equitable access and minimize potential disparities in digital health.
A historical digital divide
The coronavirus pandemic highlighted the importance of AI health care solutions, telehealth platforms and access to real-time data for self-monitoring. There have been growing expectations that these digitalized processes will help reduce health care costs, facilitate patients’ access, improve the quality of care, pave the way for precision medicine to promote better diagnostics and personalized treatments, and reduce population-level health disparities. Despite the earnestness of these hopes, COVID-19 also shed light on the historical digital divide, which has widened during our most recent digital shift. McAuley defined the digital divide as “a societal division between those who have the means and capability to make full use of digital technology and those who lack those means for reasons relating to income, education or age.” The National Institute on Minority Health and Health Disparities designates that the following populations experience health disparities: Minoritized racial/ethnic groups (American Indian or Alaska Native, Asian, Black or African American, Hispanic or Latino American, Middle Eastern or North African, Native Hawaiian or Pacific Islander), people with lower socioeconomic status, underserved rural communities, sexual and gender minority groups and individuals with disabilities.
In sleep medicine, while there is significant enthusiasm surrounding AI advancements, funding agencies and researchers have yet to fully address key questions such as, “What populations currently benefit from our AI innovations?” and “What populations may face challenges in accessing these advancements?” Given the existing digital divide, addressing these questions remains an important area of focus.
Applying a framework for digital health equity in sleep medicine
As AI implementation expands in sleep medicine, it is important to account for both social determinants of health and digital determinants of health. Digital determinants of health may function independently as barriers and facilitators of the digital divide, interacting with other social determinants of health to impact sleep health and data disparities. The NIMHD Research Framework was recently expanded for digital health equity. This framework outlines how digital environment determinants influence individual health, family/organizational health, community health, and population health level disparities. Some notable digital determinants to consider in sleep medicine include technology access, implicit tech bias, health care infrastructure, community tech norms, data standards, and algorithmic bias. In sleep medicine, more work is needed to formally assess and address how digital determinants of health interact with social determinants of health to shape access to AI innovations and reduce sleep health data disparities across NIMHD-designated health disparate populations.
Sleep data disparities and the manifestation of bias in AI
More recent frameworks for AI and sleep medicine are highlighting the importance of leveraging multiple sources of data, such as omics, electronic medical records, objective sleep assessments, environmental data, epigenetics, and additional sleep metadata (e.g., geospatial, insurance claims). With these multiple data sources being leveraged and harmonized, the field of sleep medicine has endless possibilities to study a broader range of sleep disorders, better predict patients’ risk, understand how sleep disorders occur and progress, and identify improved strategies to enhance detection, screening and treatment. However, not all populations have the sleep health data that is necessary to equitably pursue this endeavor. In fact, some populations may have systematic differences in the quantity and/or quality of their sleep health data. These sleep health data disparities may cause certain populations to be unable to benefit from the AI discoveries or innovations emerging in sleep medicine. To avoid manifesting biased AI (Table 1), it is becoming increasingly clear that the field of sleep medicine needs to prioritize data collection in real-world settings to increase the representativeness of sleep health data sources.
Table 1 outlines the type of biases that manifest from data generation to AI implementation.
| Type of bias | Description |
|---|---|
| Historical bias | Bias due to data reflecting historical disparities or biases that existed during the data collection process in which minoritized and/or underserved populations may be underrepresented due to systematic inequalities. |
| Sampling bias | Bias resulting from certain members of a population who are systematically more likely to be selected in a sample than others. |
| Representation bias | Bias resulting from making decisions based on particular stereotypes of persons or situations. |
| Measurement bias | Bias resulting from inaccurate measurements for information collected. |
| Algorithmic bias | Systematic errors in an algorithm produce unfair or discriminatory conclusions. |
| Confirmation bias | Bias resulting from a propensity to seek out information that confirms existing beliefs while ignoring information that contradicts them. |
| Interaction bias | Bias due to a tendency for individuals to interact more frequently than by chance with others of a certain characteristic or behavior. |
| Generative bias | Bias resulting from systematic errors in generated data. |
| Deployment bias | Bias that arises during the deployment process. |
| Data augmentation bias | Although generated, synthetic data can augment the bias from the real-world data that was used, due to data scarcity and/or historical bias in the training dataset or input data. |
Final thoughts
In summary, there is incredible potential for AI to benefit all areas of sleep health. These benefits, however, will only be as good as the algorithms (and data) used to develop that AI. Particular care to address and mitigate biases and collect data from diverse populations is key to preventing further widening of the health-related equity gap. We encourage the adopters of emerging AI innovations in sleep medicine to consider this as we move toward improving sleep health for all.
