By Ameena Elahi, Associate Director of AI Initiatives, Penn Medicine
Rethinking the Value of AI Modeling: Focusing Beyond Speed
When health systems begin to explore cognitive AI, the first question is likely to be something like this: How fast will this make us? Good question. But if AI is evaluated through the lens of speed, organizations risk missing out on a deeper value proposition. Cognitive load, the mental load required to process, prioritize, and act on information, should be the primary evaluation metric, not because it is a “soft” effect, but because it directly affects physician performance, patient safety, and physician well-being. Ultimately, the value of AI facing the clinic can be seen in what it changes for the doctor who no longer has to do everything at once.
Testing Imaging AI: What Really Matters
Before contracting a product or testing AI technology, it is important to determine whether the tool is compatible with the clinical environment. It is important to understand how to integrate technology without disrupting what is already working.
Many AI visualization tools come with their own portals, dashboards, or separate learning spaces. These artificial intelligence (AI) models are trained on data, and that data has population, scanner ships, and case mix baked into it. A model highly trained in the thinking of an academic medical center may work very differently in a community radiology setting with older materials and patient populations.
Some AI functions are designed to help administrators take action based on reporting metrics; others are designed to help the doctor read the image. These factors must be evaluated and addressed with products that fit the unique need, even if the same vendor is used for the same system. Conduct reasonable use cases with your radiologists, technicians, administrators, and referring physicians before purchasing decisions are made.
It is also important to assess whether your organization is ready to use this technology not only technically, but culturally. By putting together a change management plan, adoption is more likely to be successful rather than delayed or abandoned. Clear communication is needed about what AI will and won’t do, as well as allowing time to adapt to workflows to create long-term success.
A clear, confident practitioner at the end of the transition is not a soft outcome, but a fundamental system performance metric based on the power of insight and clarity.
The Quieter Revolution: Reducing the Cognitive Burden of Radiologists
In terms of a deeper value proposition, the reduction of cognitive load is where the conversation around imaging AI really needs to come from. This is very evident in the daily work of the clinic. Radiologists participate in the practice of a double workflow: when reading, a continuous, multi-particle attitude always tells the person which findings are incidental, urgent, which need to be correlated with the previous image, and which need to be answered immediately. That brainstorming process ends in a way that metrics, like hourly readings, are not designed to show.
Reducing Cognitive Load in Image Technology
The same burden of understanding is felt by experts. The CT technician receiving the scan doesn’t just push a button; their duties include evaluating the system, selecting the appropriate protocol, adjusting the patient’s condition, ensuring contrast needs, and anticipating what the radiologist may need beyond the standard protocol. AI can support this process by recommending protocol adjustments based on indication, patient history, and previous imaging, while flagging inconsistencies, such as discrepancies between the order and the clinical situation. The scan itself may take the same amount of time, but the technician spends less mental energy second-guessing decisions and worrying if something is missed.
Setting a New Path Forward: Start with Workflow, Not Technology
Radiology burn is real. When organizations examine AI only through a throughput lens, they risk solving the wrong problem. Strategic decision-making should not start with the technology itself, but with the problem it aims to solve. A common framing question is: What problem are you trying to solve? One way to break that down is to focus first on current capabilities and the desired future state, rather than technology.
Start by mapping out where the need for understanding is highest within your graphic workflow. Where are radiologists and technologists most often reported to be mentally exhausted? Where do handoffs break? Where does uncertainty accumulate?
Augmentation Over Automation
Many strong implementations of AI thinking are those that keep the radiologist logically in the loop, not as a rubber stamp for algorithmic output, but as the main thinking agent, supported by tools that reduce unnecessary mental friction. When evaluating, measure more than performance, survey about decision fatigue, perceived workload, and confidence at the end of the change, not just volume. Create a business case that includes fatigue costs, profitability, and the number of incremental benefits of efficiency.
A doctor who finishes a shift feeling competent and clear, instead of tired, is not a soft result, it is a system requirement. An AI that protects that situation is more important than an AI that just acts quickly.
