Addressing Staffing Shortages, Burnout, and Beyond: The Role of AI in Supporting Healthcare

How can we address the complex challenges in?#healthcare?? Staffing shortages,?#burnout, worsening patient experiences, and rising costs with stagnant outcomes are issues we can't ignore. Technology is the only solution I know that can simultaneously improve?#quality?and?#efficiency?while reducing?#costs. Specifically, how can?#AI?help? While the path ahead isn’t easy, as the saying goes, nothing worthwhile ever is.

#Quality:

We’ve evaluated multiple AI computer vision models for accuracy and found they can improve radiologists' ability to detect pathologic findings. We refer to this improvement as the relative Enhanced Detection Rate (#EDR), which, in our validations, ranged from 0% to over 300%. Only one AI model had an EDR of 0%, so we decided not to roll it out. Instead, we shared our findings with the AI vendor, who subsequently improved their model. (Spoiler: the updated version now has an EDR exceeding 20%!). Importantly, AI isn't a standalone solution—radiologists also improve AI accuracy, creating a synergistic relationship. Like the collaboration between a resident and an attending physician, the highest level of accuracy is achieved when humans and AI work together. How can you determine if an AI model will enhance detection and be valuable to radiologists in your practice? We shared a summary of our five-step best practice for AI model validation in a multi-society AI paper (link to follow), and you can watch a presentation about this process (link to follow). A more comprehensive publication is currently in progress.

#Efficiency?and #Burnout:

Computer vision #AI?models demonstrate modest efficiency gains of approximately 10%. Efficiency gains are primarily driven by negative exams (exams without the pathology being reviewed by the AI model). Since studies without an individual pathology (e.g., CTPA exams without a PE) significantly outnumber those with the pathology (e.g., CTPA exams with a PE), negative exams have a greater impact on reducing the overall average read time. #Radiologists?tend to read CTPA exams faster when AI returns a negative result, as these models typically have exceptionally high Negative Predictive Values (#NPVs), often approaching 100%. Hence, a radiologist can review a CTPA more quickly when augmented by the additional confidence of a negative AI result. AI model NPV is almost always high because it is inversely proportional to disease prevalence, which is low in many imaging studies. As imaging continues to be used as a frontline diagnostic tool—sometimes even before a complete patient evaluation—apparent disease prevalence may decline further, making NPVs even higher and driving greater efficiency gains. Unfortunately, computer vision AI tools aren’t applied across all imaging types, limiting their impact on overall practice efficiency. Large language models (#LLMs), on the other hand, fine-tuned on radiology reports have made significant strides in improving radiologist efficiency and can be applied to every exam. When tailored to mimic individual radiologist’s reporting style, these tools not only improve efficiency but also reduce cognitive load and hence, reduce #burnout. With more workflow and reporting tools emerging—many of which were showcased at the Radiological Society of North America (RSNA)?2024—we’re on the cusp of even greater improvements in radiologist efficiency and burnout.

The #Takeaway:

AI has the potential to revolutionize healthcare by enhancing quality, reducing burnout, and improving efficiency. However, the true impact comes from the #synergy?between human expertise and technological innovation—an approach that depends on strong collaboration between vendors and clinical experts. I recommend partnering with a vendor who listens to your clinical insights and actively collaborates to develop solutions that meet your practice's needs and drive meaningful results.

?