Artificial Intelligence (AI) has the potential to revolutionize healthcare by increasing efficiency and improving patient outcomes, however, most medical AI technologies do not ultimately reach clinical patient care. Many algorithms, despite their technical appeal, are not clinically relevant or are not designed to integrate with the clinical setting. Even AI with validation results that approach or surpass human performance are not adopted into routine clinical practice due to logistical challenges in translation from the ideal, simulated research environment into clinical workflows. This session will discuss about the challenges and opportunities for the clinical translation of AI.
In the first talk, Dr. Veeraraghavan will share their experience in deploying AI auto-segmentation for clinical use. She will introduce AI auto-segmentation applications currently used for radiotherapy at Memorial Sloan Kettering Cancer Center and discuss issues with deployment and ongoing quality assessment, followed by the new developments on multi-tasked learning solutions for clinical use, and ended with a discussion on challenges and potential solutions.
In the second talk, Dr. Purdie will discuss about a major barrier for the clinical adoption of AI which occurs when AI and the clinician inevitably disagree. In this case, the AI is often discounted and fails to be adopted beyond the pilot phase as healthcare providers not only expect high accuracy from AI, but also expect AI to agree with their own views (including biases), even when in practice, large inter-clinician variation exists. He will look at sources of automation bias limiting the clinical deployment of AI and the effectiveness of AI technologies for impacting patient care.
In the third talk, Dr. Kang will discuss the translation of minimally-supervised natural language processing (NLP) to interpretable clinical AI. Natural language processing combined with subject matter expertise can extract knowledge that previously required significant time and resources to uncover. Due to these methods being pre-trained on vast sums of raw text, they inherently support transfer learning and are scalable. He will discuss developments in biomedical NLP relevant to radiation oncology that have leveraged deep learning and subject matter expertise, and contrast them with knowledge representation models.
In the fourth talk, Dr. Jiang will discuss the individualization of AI for the successful implementation of AI in clinical routine. He will start with the model generalizability issue. For most of the clinical applications, one cannot develop a universal model that works anytime anywhere for anybody. He will also talk about the fact that medicine is still an art in many cases so that part of the variability among physicians’ clinical practice is inherent, cannot be removed and should be respected when implementing AI for clinical use. A potential solution to these problems is to individualize AI for each individual institution, clinician, and even patient.
In the last part of the symposium, we will have a Q&A session for discussions between session organizers, speakers, and the audience.
1. Understand the issues with deployment and ongoing quality assessment of AI-based auto-segmentation;
2. Understand sources of automation bias limiting the clinical deployment of AI and the effectiveness of AI technologies for impacting patient care;
3. Gain knowledge on the translation of minimally-supervised natural language processing to interpretable clinical AI;
4. Understand why AI models should be adapted to local institutions, clinicians, and even patients.
Not Applicable / None Entered.
Not Applicable / None Entered.