In the recent AIMed breakfast briefing: Experience the future of AI (artificial intelligence) in Radiology – Chicago, Dr. Paul J. Chang, Professor and Vice-Chairman of Radiology Informatics at the University of Chicago School of Medicine highlighted some of the fallacies in modern radiology.
He said while digital based radiology is certainly more efficient, it gradually kills off the collaboration which once existed among clinical colleagues. The lack of a more complete understanding of patients’ clinical contexts drives radiologists to be dependent on radiology reports. This makes their interpretations less precise and impactful.
At the same time, digitization is causing a significant growth in data size and complexity. Burnouts become prominent as radiologists are juggling between multiple tasks. Administrative responsibilities coming from electronic medical records (EMRs) and Picture Archiving and Communication System (PACS), on the other hand, do not directly contribute to provision of care to patients.
Challenges coming from technology hypes
Apart from AI, the other hype which Dr. Chang had identified is radiomics or the “generic characterization to more specific and ‘actionable’ phenotypic characterization”. This means the industry has moved on from isolate image reading to multidisciplinary synthesis. In spite of the progress, in Dr. Chang’s opinion, radiology is barely sustainable.
The existing information technology (IT) systems and human radiologists are unable to cope with the change. Most of the time, humans are still required to “remember to do all the right thing” in workflow orchestration and communication of incidental findings. Resources tend to be wasted on mundane work or performing manual search for patients’ details. In short, human-machine collaboration remains primitive.
Furthermore, radiology tends to “buy early into these hypes”, without making a thorough assess of available tools. This often resulted in early buy-in but long-time adoption. As Dr. Chang cited the Amara’s Law on his slide, “We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run”. Notably, validations of AI algorithms become a challenge in face of the evidence-based tradition in radiology or medicine, in general.
At the hidden layer, radiologists are unclear how AI is able to turn a massive amount of data into a credible solution. Currently, validation is statistical based and we are unsure of avoiding “overfitting” or a result which corresponds to a particular set of data which it was created but fails when new data is introduced or being employed away from its original setting.
Dr. Chang believes radiologists should not be trapped in their echo chambers. After all, AI applications are frequently driven by data availability and not
On top of data interoperability, there is also a need to prepare existing IT infrastructure to be able to consume near future advanced decision support agents, including those that are cloud based. More importantly, if the management is truly interested in AI, they should get hands dirty with analytics and initiate infrastructural support for both analytics and AI. To start, as Dr. Chang noted, EMRs should be re-engineered.
Most importantly, the way AI is being adopted into medicine will have to be re-examined. The industry has rejected generalized and one size fit all medicine yet most AI startups are aiming to develop this one algorithm which can be applied across different settings. Dr. Chang noted while this perspective is rather controversial, there remains a need to ponder whether generalization is possible or even desirable at the end of the day.
A science writer with data background and an interest in current affair, culture and arts; a no-med from an (almost) all-med family. Follow on Twitter.