When talking about artificial intelligence (AI) in radiology, one immediately thinks of the classic clinical use case of computer vision and powerful algorithms identifying abnormalities on imaging studies. Operational applications of AI, however, will likely be the first to be widely adopted.
Many companies and researchers today are indeed focused on the diagnostic element of the radiology pathway, and the first wave of computer-vision algorithms can undoubtedly support and augment radiologists. But AI doesn't end there. The operational applications of AI in radiology should not be overlooked or underestimated, as the imaging pathway starts well before the diagnostic process and ends a long time afterward.
There is considerable opportunity in providing AI-driven solutions in the pre- and postdiagnostic pathway, including referrals, scan setup and acquisition, reporting, follow-up management, practice management, and other "noninterpretive" workflows. Furthermore, operational applications are easier to develop and adopt, as they require less rigorous medical device regulation and much lighter integration.
Structuring the unstructured
Radiologic imaging utilization has seen substantial growth over recent years, and health systems and private radiology practices can -- and should -- make better use of the data generated through the regular course of business. This can yield operational efficiencies, lower costs, revenue growth, higher satisfaction, and better patient care and experience.
Radiologic textual reports contain important data, but such data can be lost, miscommunicated, or misinterpreted due to the unstructured nature of such reports. Attempts at structuring reports to make better use of the data have been made for several decades now, initially using distinct headers and sections for radiologists to complete.
For instance, the RSNA has compiled RadReport, a library of reporting templates that are free to download by any department wishing to provide more structure to its reports. Such templates typically include sections such as indication, technique, findings, and impression, with structured fields within the sections. Systems utilizing such templates often include pick lists in an attempt to make data more homogeneous and easier to report.
However, every case is slightly different and "one size fits all" is, more likely than not, untrue in that context and fails to offer an optimal solution. Some structured fields may be necessary, while others are redundant. Or, such fields may not adequately capture nuances or important pieces of information.
The uptake of structured reporting also remains low due to the difficulty of compliance. Asking radiologists to check boxes or to remember to use very specific terminology for certain cases can be ineffective and burdensome. When reading hundreds of exams a day, it is simply not humanly possible to manually complete the tasks and remember each and every box to check.
This is where AI comes in. Automatic mining of the data makes the inefficient attempt of structuring reports by using templates unnecessary. It allows care and operational teams to keep their workflow while letting the AI mine the most important and actionable information behind the scenes.
Natural language processing
Natural language processing and understanding (NLP/NLU) -- a branch of AI designed to read, decipher, understand, and make sense of the human languages -- is poised to dramatically improve the potential for data mining of textual reports. There are many operational applications of AI, particularly NLP, in the radiology space.
AI and NLP can help monitor and predict performance at the provider level, such as over- or underutilization of certain scans based on false positives or false negatives, and relative to a provider cohort baseline. The ability to recognize key words and automate coding for billing can also make the process more efficient and more thorough, as well as minimize claims denials.
NLP can also be utilized to determine if a follow-up scan has been recommended, and if so, when and for what purpose, while determining appropriateness and compliance. This can help ensure follow-ups occur and downstream care is delivered in a timely manner.
In addition, NLP can help radiologists anticipate required resources and plan accordingly. In certain cancer cases, for instance, patients under surveillance regularly and consistently use imaging resources. When there is new cancer detection, needs surge. NLP can then be used to more readily extract pertinent data from textual reports and help predict resource utilization.
The best deployments of AI are as augmented intelligence. This approach helps medical and administrative teams perform at their best by providing them with recommendations that they can accept, reject, or change based on their human judgment.
Providing this information in a nondisruptive, easily accessible and understandable way, such as a front-end dashboard, is equally important. With data turned into information at their fingertips and within their workflow, healthcare teams can gain insights, make decisions, and respond quickly to real-time information.
The AI revolution is here, and many industries are already benefiting from these advances. Healthcare executives should take full advantage of these capabilities and applications to promote better patient care while increasing top-line revenue, improving efficiencies, and mitigating liability risk.
While the use of AI in clinical applications may face higher adoption hurdles, there are operational and business process efficiencies to be gained right now from the technology. These operational applications also foster better patient care and experience, providing a win-win proposition for all stakeholders in the healthcare system.
Michal Meiri is the co-founder and CEO of Agamon, a healthcare AI startup focused on using deep-language understanding to close care gaps in radiology. She has built multiple products and led research and development teams in the U.K. and Israel, developing AI applications using large volumes of unstructured data. Michal holds a Bachelor of Science in chemistry and computer science from Tel Aviv University and a Master of Philosophy in nanotechnology from the University of Cambridge.
The comments and observations expressed are those of the author and do not necessarily reflect the opinions of AuntMinnie.com.