AUSTIN – Before AI came along, the Society for Imaging Informatics in Medicine (SIIM) seemed to be a conference in search of itself. SIIM (and before it, SCAR) built its reputation on education and training for radiology’s shift to digital image management.
But what happens when the dog catches the truck? Radiology eventually fully adopted digital imaging, and that meant less need to teach people about technology they were already using every day.
Fast forward to the AI era, and SIIM seems to have found its new mission. Once again, radiology is faced with a transformative IT technology that few understand and even fewer know how to put into clinical practice. With its emphasis on education and networking, SIIM is a great forum to learn how to do both.
That’s exemplified by the SIIM keynote address on Wednesday, by Ziad Obermeyer, MD, a physician and researcher in machine learning at UC Berkeley who has published important research on bias in machine learning.
While not a radiologist, Obermeyer served up a fascinating talk on how AI should be designed and adopted to have maximum impact. His advice included:
- Don’t design AI to perform the same tasks humans do already. Train algorithms to perform in ways that make up for the shortcomings of humans.
- Training algorithms on medical knowledge from decades ago is likely to produce bias when today’s patient populations don’t match those of the past.
- Access to high-quality data is key to algorithm development. Data should be considered a public good, but there is too much friction in getting it.
To solve some of these challenges, Obermeyer is involved in two projects, Nightingale Open Science to connect researchers with health systems, and Dandelion Health, designed to help AI developers access clinical data they need to test their algorithms.
The rise of AI – particularly generative AI models like ChatGPT – has given SIIM a shot in the arm from a content perspective, and the return of in-person meetings plays to the conference’s strength as an intimate get-together where the networking and relationship-building is almost as important as the content. Please follow along with the proceedings of SIIM 2023 on our Twitter and LinkedIn pages.
CT continues to flex its muscles as a tool for predicting heart disease risk, in large measure due to its prowess for coronary artery calcium scoring. In JAMA, a new paper found CT-derived CAC scores to be more effective in predicting coronary heart disease than genetic scores when added to traditional risk scoring.
Traditional risk scoring – based on factors such as cholesterol levels, blood pressure, and smoking status – has done a good job of directing cholesterol-lowering statin therapy to people at risk of future cardiac events. But these scores still provide an imprecise estimate of coronary heart disease risk.
Two relatively new tools for improving CHD risk prediction are CAC scoring from CT scans and polygenic risk factors, based on genetic variants that could predispose people toward heart disease. But the impact of either of these tools (or both together) when added to traditional risk scoring hasn’t been investigated.
To answer this question, researchers analyzed the impact of both types of scoring on participants in the Multi-Ethnic Study of Atherosclerosis (1,991 people) and the Rotterdam Study (1,217 people). CHD risk was predicted based on both CAC and PRS and then compared to actual CHD events over the long term.
They also tracked how accurate both tools were in reclassifying people into different risk categories (higher than 7.5% risk calls for statins). Findings included:
- Both CAC scores and PRS were effective in predicting 10-year risk of CHD in the MESA dataset (HR=2.60 for CAC score, HR=1.43 for PRS). Scores were slightly lower but similar in the Rotterdam Study
- The C statistic was higher for CAC scoring than PRS (0.76 vs. 0.69; 0.7 indicates a “good” model and 0.8 a “strong” model)
- The improved accuracy in reclassifying patient risk was statistically significant when CAC was added to traditional factors (half of study participants moved into the high-risk group), but not when PRS was added
This study adds to the growing body of evidence supporting cardiac CT as a prognostic tool for heart disease, and reinforces CT’s prowess in the heart. The findings also support the growing chorus in favor of using CT as a screening tool in cases of intermediate or uncertain risk for future heart disease.
If you’re a radiologist, chances are at some point in your career you’ve cherry-picked the worklist. But picking easy, high-RVU imaging studies to read before your colleagues isn’t just rude – it’s bad for patients and bad for healthcare.
That’s according to a new study in Journal of Operations Management that analyzes radiology cherry-picking in the context of operational workflow and efficiency.
Based on previous research, researchers hypothesized that radiologists who are free to pick from an open worklist would choose the easier studies with the highest compensation – the classic definition of cherry-picking.
To test their theory, they analyzed a dataset of 2.2M studies acquired at 62 hospitals from 2014 to 2017 that were read by 115 different radiologists. They developed a statistical metric called “bang for the buck,” or BFB, to classify the value of an imaging study in terms of interpretation time relative to RVU level.
They then assessed the impact of BFB on turnaround time (TAT) for different types of imaging exams based on priority, classified as Stat, Expedited, and Routine. Findings included:
- High-priority Stat studies were reported quickly regardless of BFB, indicating little cherry-picking impact
- For Routine studies, those with higher BFB had much lower reductions in turnaround — a sign of cherry-picking
- Adding one high-BFB Routine study to a radiologist’s worklist resulted in a much longer increase in TAT for Expedited exams compared to low-BFB studies (increase of 17.7 minutes vs. 2 minutes)
- The above delays could result in longer patient lengths of stay that translate to $2.1M-$4.2M in extra costs across the 62 hospitals in the study.
The findings suggest that radiologists in the study prioritized high-BFB Routine studies over Expedited exams – undermining the exam prioritization system and impacting care for priority cases.
Fortunately, the researchers offer suggestions for countering the cherry-picking effect, such as through intelligent scheduling or even hiding certain studies – like high-BFB Routine exams – from radiologists when there are Expedited studies that need to be read.
The study concludes that radiology’s standard workflow of an open worklist that any radiologist can access can become an “imbalanced compensation scheme” that can lead to poorer service for high-priority tasks. On the positive side, the solutions proposed by the researchers seem tailor-made for IT-based interventions, especially ones that are rooted in AI.
In the never-ending quest to get referring physicians to follow radiologist recommendations for follow-up imaging, Massachusetts researchers in JAMA Network Open offer an IT-based solution: Structured reporting software that was found to triple the number of radiology reports judged to be complete.
A recent study found that 65% of radiologist recommendations for follow-up imaging aren’t followed by referring physicians. Authors of that study found that recommendations that were strongly worded and communicated directly to referring doctors had higher uptake.
But what if radiologists don’t follow this advice? In the new paper, researchers from Brigham and Women’s Hospital and Harvard Medical School offer a more structured solution thanks to software developed as part of their Addressing Radiologist Recommendations Collaboratively project.
The ARCC software is a closed-loop communication system that’s designed to channel radiologist recommendations into a structured format that’s clearly understood, while also tracking whether they were accepted and fulfilled. The ARCC tool runs separately from the radiologist’s dictation software, so while it asks them to include a standardized recommendation sequence in their report, it leaves the specific free-text language up to them.
Under the ARCC criteria, the main factors that make up a complete follow-up recommendation are:
- Reason for imaging study
- Timeframe when study should be completed
- Imaging modality to be used
The researchers implemented the ARCC software in October 2019 in thoracic imaging, and rolled it out to other departments through December 2020. Use of the software was “strongly encouraged but voluntary.”
In testing the ARCC software’s effectiveness, the researchers found that the number of follow-up recommendations considered to be complete – with all three key elements – rose from 14% to 46%. Even so, one-third of reports filed with ARCC “still contained ambiguous language” in the free-text section – indicating that old habits are hard to break.
Radiologists may hate it when their recommendations for follow-up imaging are ignored, but referring physicians are also frustrated with free-text radiology reports that are wishy-washy and contain vague impressions. The ARCC software could bridge the gap by steering radiologists toward recommendations that are more concrete and specific – and more likely to be followed.