Technical and regulatory hurdles abound, but another challenge is finding enough providers who are comfortable and competent with artificial intelligence technology.
Artificial intelligence (AI) could play a major role in the future of obstructive sleep apnea (OSA) diagnosis, but before that can happen a number of technological, regulatory, and educational hurdles must be overcome, according to a new analysis.
Hannah L. Brennan and Simon D. Kirby, MD, both of Memorial University of Newfoundland and Labrador, in Canada, outlined the problem in a recent review published in the Journal of Otolaryngology - Head & Neck Surgery.
The investigators noted that while OSA is believed to be increasing in prevalence (with as many as 14% of men and 5% of women believed to suffer from the disease), diagnosis of OSA remains a challenge.
“Despite the high prevalence, the number of people diagnosed and seeking treatment is much less, reflecting lack of accessibility in treatment and testing and lack of correlation with the index measured on diagnostic tests,” they wrote.
The authors laid out the case for AI in OSA diagnosis. Polysomnography, the “gold standard” of OSA diagnosis, is costly, with its requirement of specialized facilities and medical personnel. Besides, many patients never make it to a sleep clinic because the tools used to screen patients for OSA are either underutilized or ineffective.
AI has the ability to leverage huge amounts of data to spot patterns and warning signs, potentially offering a better way to diagnose patients or identify those who warrant a formal sleep study.
The idea of integrating AI into sleep medicine is already gaining steam in the marketplace.
In May, the software-maker Somnoware announced it had received venture capital funding to further its respiratory care management software. The tool is designed to help better integrate data from “siloed” devices and software in order to help physicians better identify patients with sleep disorders such as OSA and respiratory disorders such as chronic obstructive pulmonary disease (COPD).
Earlier this month, EnsoData, a Madison, Wisconsin, startup, said it had closed a $20 million round of funding. The company’s platform uses AI to streamline the scoring of sleep studies, thereby speeding up diagnosis.
Brennan and Kirby said the most prominent kind of AI being studied is artificial neural networks (ANNs), in which software is used to replicate the structure and behavior of the human nervous system. ANNs can take clinical data on conditions such as body mass index, snoring status and pulse oximetry and either rule OSA in or out. Though some ANNs are more complex than others, Brennan and Kirby said studies suggest ANNs can reach levels of specificity and sensitivity in the high 80% to 90% range.
“Studies have shown the value of ANNs and the high specificity and sensitivity that is necessary for effective screening, notably much higher than based on subjective impressions of experienced sleep physicians, which detected OSA with a 60% sensitivity and 53% specificity,” they wrote.
Yet, that gap between human interpretation and big data computation creates problems. One barrier to AI implementation, Brennan and Kirbey said, is the inaccessibility of the “reasoning” of ANNs. Even if the software might be able to notice patterns humans cannot, physicians will have to be convinced to trust the software’s results even if the physician cannot fully understand how the software arrived at its conclusion.
“There is a balance between accuracy and algorithm transparency, where the most accurate models (deep learning and ANNs) are the least interpretable and the more intuitive models are less accurate,” Brennan and Kirby wrote. They said ANNs will therefore have to achieve such a high rate of accuracy that it outweighs the desire for a clear understanding of the algorithm.
Another challenge the authors identified is the persistent problem of bias within data. A data set that might lead to highly accurate diagnosis in a general population might not be as effective in smaller population subgroups that have a higher prevalence of OSA.
This also leads to regulatory hurdles, the authors noted. Regulators will need to verify that these tools are effective and also ensure that the underlying data are used in accordance with privacy and data protection regulations.
In some regards, regulation could be like trying to hit a moving target.
“Since AI systems change and ‘learn’ there is also a question of whether these systems require serial cycles of approval,” they wrote.
The final barriers, however, are decidedly human. For AI to be an effective tool, Brennan and Kirby said, providers will need to have personnel with the technical expertise to build, maintain, and improve AI systems. Healthcare providers and institutions will need to embrace the technology. Medical schools will also need to do a better job of training students to be familiar and competent with AI tools.
“Introducing AI education into the medical school curriculum is an important step, yet there are barriers including the dense curriculum, the knowledgeable personnel required to teach such skills, and the uncertainty surrounding how AI will impact the future of healthcare and the scope of healthcare workers responsibilities,” Brennan and Kirby said.
Though a number of important barriers remain that limit AI’s use in diagnosing OSA, the authors said the barriers to AI implementation in sleep medicine are similar to those in other fields. They said future research into AI should focus not only on developing the technology, but also on breaking down these barriers.