Source/disclosure information
Issuer:
Fernandez BM. Unlocking the Potential of AI in Clinical Applications: Challenges and Opportunities. Venue: American Society of Cataract and Refractive Surgery Congress. May 5-8, 2023. San Diego.
Disclosure:
Fernandez reports that he is an employee of Heidelberg Engineering.
Important points:
- Artificial intelligence models must explain decisions and quantify uncertainties in order to be safe and trustworthy.
- These elements can be implemented without compromising performance.
SAN DIEGO — To gain clinical adoption, an artificial intelligence model cannot be a black box, according to a speaker at the American Society for Cataract and Refractive Surgery’s DOS Digital Day.
Dr. Brian M. Fernandez, AI is gaining attention in the field of ophthalmology, but AI model designers need to consider safety, he said.
“The most important aspect of these tools is clinical safety,” he said. “Making a wrong diagnosis or missing a disease has serious consequences for patients. …It is important that these models exhibit some degree of uncertainty in their results.”
In addition to clinical needs such as safety, Fernandez said AI models need to be explainable and interactive.
AI models also need to adapt to new real-world data outside the framework in which they were trained.
Fernandez explained the concept of out-of-distribution detection, which is important in implementing AI in clinical settings. Out-of-distribution detection is the ability of AI models to recognize untrained data. If the model fails to recognize this data, Fernandez said, the model may make false predictions.
Finally, the model should be able to quantify its uncertainty. Including a small number of outliers in the training set, Fernandez said, protects the model from image misclassification while improving outlier exposure and uncertainty measures.
“We believe that a secure and trustworthy AI model can explain why it made a decision and quantify the degree of uncertainty,” Fernandez said. “These aspects can be incorporated into the training of the model without compromising the performance of the original task.”
