A recent study on obstructive sleep apnoea (OSA) found that machine learning could improve diagnosis but raises critical concerns among clinicians regarding its implementation. OSA affects 5% to 15% of the population, yet many cases remain undiagnosed mainly due to the complexity and accessibility issues related to polysomnography (PSG), the current gold-standard diagnostic method.
Read More
In the study, researchers conducted interviews with ten clinicians to gather their insights on the integration of machine learning into OSA diagnosis. The analysis revealed several themes, highlighting the necessity for regulatory guidelines and the involvement of medical professionals when applying these technologies. Clinicians expressed that while machine learning could offer valuable support, it should not replace human expertise in diagnosis. The interviews emphasized that the history-taking process is paramount in the diagnosis of OSA, as clinicians rely on comprehensive patient information to inform their evaluations.
While machine learning demonstrates the potential for high accuracy—some studies report up to 85.46% accuracy in OSA detection—clinicians remain skeptical about the reliability of using fewer physiological signals for diagnosis. They caution that reduced datasets could be inadequate for patients with complex medical histories and overlapping symptoms with other conditions. Participants articulated the importance of ensuring machine learning outputs are clinically useful and aligned with individualized patient assessments.
Furthermore, clinicians are concerned about the "black box" nature of many AI systems, which often lack transparency in their decision-making processes. They stressed the need for these systems to provide clear, interpretable results that can enhance trust and facilitate integration into clinical workflows. They also pointed out the importance of adhering to established clinical guidelines to mitigate legal and ethical concerns surrounding AI in medical diagnostics.
Overall, the findings point to a keen interest among clinicians in adopting machine learning tools, provided that there is clarity around their use and an emphasis on maintaining human oversight in the healthcare provision. Recommendations from the study suggest that future developments in machine learning for OSA should focus on harmonizing AI outputs with traditional diagnostic practices while prioritizing the critical role of clinician expertise.