I first learned about artificial intelligence (AI) in healthcare during my undergraduate studies. As a biomedical engineering major planning a career in medicine, the concepts of AI in healthcare intrigued me. There was talk that these innovations could revolutionize medical practice in a way that would disrupt traditional physician roles. When I entered medical school, I expected to see these innovations translated into medical practice. The launch of Watson Health shortly after my medical school graduation instilled hope that these innovations would finally make it into my medical practice during residency training. I am now finishing my specialty training and will transition into the “real world” of medical practice as an attending. I have not seen any significant use of AI in my own clinical practice during my time in Boston, but exciting developments in the last two years hold much promise. Some data are starting to emerge that healthcare metrics can be improved with AI, as evidenced by Emergency Department visits being deceased by AI based mobile technology in the UK. The recent dissolution of a partnership between Watson Health and MD Anderson is a reminder that translating these innovations into clinical practice is fraught with challenges. Dr. Michael Forsting’s article “Machine Learning Will Change Medicine” in the Journal of Nuclear Medicine grappled with some of these issues.
Dr. Forsting is a medical doctor and researcher for the Institute of Diagnostic and Interventional Radiology and Neuroradiology at Essen University Hospital in Germany. His article focused on AI in the field of Radiology, but many of the ideas in his paper can be applied to other specialties.
One of the potential benefits when integrating AI into medical practice is improvement of clinical decision making and diagnosis. Dr. Frosting points out that rare diseases are often overlooked and biases can skew diagnostic pathways. The concept of using AI to provide clinical decision support systems for physicians has been studied in some medical specialties with varying degrees of effectiveness. The goal is to provide automation of certain physician cognitive tasks during a medical workup to improve time and accuracy to diagnosis. Some researchers envision a future where AI can provide end-to-end automation in the workup, diagnosis, and treatment of patients. There are still many challenges before this level of functionality can be achieved. The diagnostic accuracy of methods used by AI systems remains an issue. Comparisons between the paradigms of machine learning based, knowledge based, and hybrid methods have not yielded a clear model on how AI best uses clinical data to arrive at a diagnosis. Current methods have yet to achieve the level of diagnostic accuracy for a broad range of diseases necessary to gain acceptance amongst most physicians. More studies are needed to determine which method in AI will be robust enough to help providers in their clinical decision making process.
This paper also outlined some barriers facing the integration of AI systems into medicine. One interesting challenge is that some diseases do not correlate 1:1 to set patterns. Often, providers rely on gestalt gained through their depth of training to determine whether a diagnostic test is deemed necessary. There is promise that Deep Learning methodology will allow for AI to train much like a medical resident does through a large data set of disease presentations. The AI gains a capability to use “inference”, another Deep Learning concept, to apply this capability to handle novel sets of data. The problem is that there is too much data which cannot be utilized by learning systems, either due to data silos or other factors contributing to incompatibility. In the datasets that are available, it is unclear what portions are useful. Elucidating useful health data to “train” AI systems is a current area of active development. Having access to data representative of disease states is crucial in AI’s ability to develop an accurate capability to make inferences. Google’s acquisition of Deepmind has secured data to develop AI powered health app. Google is also collecting baseline physiologic biomarkers in healthy subjects. Further developments in accurate datasets are necessary before AI is able to provide accurate inferences during workup of a new patient.
Many tout last year as the year of the AI. Google and Amazon have brought AI into the consumer home. The healthcare systems have improved data sharing and access for AI systems. It has been 10 years since I first came across the concepts of AI in healthcare, and I still maintain that these promises will hold true. Mobile technology and AI will fundamentally change the way medicine is delivered to patients, but there are still several barriers for innovators to overcome. There is potential for AI to do more than just automate the processes for the diagnosis and treatment of disease. The true power comes from how computing could potentially transcend conventional ways of thinking about disease and medicine. AI can integrate and infer from a much larger dataset than any human can, discerning patterns that are difficult to appreciate from a human perspective. I am excited to see what the next few years will hold for AI and healthcare.