Challenges to an AI renaissance
Despite making strides in numerous specialties such as radiology, ophthalmology, pathology, oncology, and dermatology, AI has not yet gained mainstream usage. Why isn’t AI as ubiquitous and heavily entrenched in health care as it is in other industries? According to the National Academy of Medicine’s AI in Healthcare: The Hope, The Hype, The Promise, The Peril, there are several realities to address before we fully embrace the AI revolution (Matheny M, et al. 2019).
First, AI algorithms should be trained on quality data that are representative of the population. Interoperability between health care systems and standardization across platforms is required to access large volumes of quality data. The current framework for data gathering is limited due to regulations, patient privacy concerns, and organizational preferences. The challenges to data acquisition and standardization of information will continue to snarl progress unless there are legislative remedies.
Furthermore, datasets should be diverse enough to avoid introducing bias into the AI algorithm. If the dataset is limited and health inequities (eg, societal bias and social determinants of health) are excluded from the training set, then the outcome will perpetuate further explicit and implicit biases.
The Food and Drug Administration (FDA) reviews and authorizes AI/ML-enabled devices. Its current regulatory structure treats AI as a static process and does not allow for exercise of its intrinsic ability to continuously learn from additional data, thereby preventing it from becoming more accurate and evolving with the population over time. A more flexible approach is needed.
Lastly, recent advanced AI algorithms including deep learning and neural network methodology function like a “black box.” The models are not explainable or transparent. Without clear comprehension of its methods, acceptance in clinical practice will be guarded and further risk of inherent biases may ensue.
A path forward
But these challenges, like any, can be overcome. Research in the area of differential privacy and the adoption of recent data-sharing standards (eg, HL7 FHIR) can facilitate access to training data (Saripalle R, et al. J Biomed Inform. 2019;94:103188). Regulators are also open to incorporating feedback from the AI research community and industry in favor of innovation in this frenetic domain. The FDA developed the AI/ML Software as a Medical Device Action Plan in response to stakeholder feedback for oversight (FDA, 2021). Specifically, the “Good Machine Learning Practice” will be developed to describe AI/ML best practices (eg, data management, training, interpretability, evaluation, and documentation) to guide product development and standardization.
Sleep medicine has significantly progressed over the last several decades. Rather than maintain the status quo, AI can help fill the existing knowledge gaps, augment clinical practice, and streamline operations by analyzing and processing data at a volume and efficiency beyond human capacity. Fallibility is inevitable in machines and humans; however, like humans, machines can improve with continued training and exposure.
We asked ChatGPT about the future of AI in sleep medicine. It states that AI could have a “significant impact” on sleep disorders diagnosis, treatment, prevention, and sleep tracking and monitoring. Only time will tell if its claims are accurate.
Dr. Tan is Clinical Associate Professor with the Division of Sleep Medicine at the Stanford University School of Medicine. Dr. Bhargava is Clinical Professor with the Division of Pediatric Pulmonary, Asthma, and Sleep Medicine at the Stanford University School of Medicine.