One of the most significant advances in 2015 was the emergence of artificial intelligence (AI) in a variety of fields, including healthcare. While work on AI has been going on for years and still has a long way to go, 2015 saw progress made on self-driving cars, a growing ecosystem of AI startups, and open-sourcing of AI technology from Google, Facebook, and IBM.
In my company’s field, clinical medical imaging, IBM’s $1 billion acquisition of Merge Healthcare earlier this year, followed by extensive promotion, brought visibility to the use of AI in medical imaging. (Merge Healthcare had previously acquired the assets of our former company, Emageon/UltraVisual.)
In addition to electronic health records, lab results, and genetic data, medical images—such as CT and MRI scans—are certain to be valuable as data inputs to AI algorithms. Genomics has already proven to be valuable in routine clinical use. If the genotype—the blueprint for an individual—includes meaningful data, then undoubtedly a medical image that is more representative of the phenotype—the embodiment of that blueprint—includes meaningful data.
The promise for AI technology in the future is incredible. However, when it comes to applying AI to healthcare, caution must be taken in expecting too much too soon. And the clinician must be an integral part of the process.
An analogy is computer-aided diagnosis (CAD) technology, a field in which companies invested significant resources over the past two decades to develop algorithms to find cancerous lesions automatically. Regulatory clearance for these products was complex and lengthy. But CAD has become just a secondary feature of the radiologist’s imaging (PACS) system and hasn’t yet lived up to its promise. Why?
First, the algorithms must be tuned very conservatively to avoid missing a lesion, which would pose a significant hazard. But this generates a large number of false positives. The radiologist must already review all of the images and is now correcting mistakes rather than focusing on the diagnostic findings. Secondly, radiologists aren’t keen to adopt a technology that purports to replace a major part of their job.
If an AI system pulls in all of the patient data and spits out a diagnosis or a treatment plan, a clinician is still going to need to review the data and the result. And it’s likely that to avoid hazards and obtain regulatory clearance, the algorithms will be conservative and produce false positives. The FDA path for AI will be daunting, and while Silicon Valley is fond of non-medical founders “disrupting” things, it will take deep regulatory and medical experience to be successful. Using the diagnostic expertise of clinicians alongside AI technology for clinical decision support and efficiency is a more promising path to adoption of AI.
With an incremental approach building on the skills of clinicians, the future is very bright for AI in healthcare.
