In a conversation with Kang Zhang, MD, one of the pioneers in the applications of healthcare AI, he mentioned to me about the amazing potential of generative AI to be a game changers in healthcare. Dr. Zhang has done some of the most pioneering studies in this space and has been face to face with the limitations of natural language processing (NLP) on clinical notes and reports. He talked about how there’s no standard format for generating clinical, radiology, and pathology data. It’s therefore been difficult to apply natural language processing (NLP) techniques to accurately extract important clinical data out of this information and to feed it into the structured data models that deep learning applications use for decision support.
He also thinks that although NLP has made rapid progress in extracting information from unstructured EHR systems, it may take considerable time to achieve complete and accurate identification of multidimensional health data and thus to generate effective clinical decisions with AI. Dr. Zhang also mentioned that current AI clinical application scenarios are basically post-hoc analyses and the improvement of order sets, such as guiding the treatment of severe pneumonia in a hospital.
This is also true of the decision support systems that are based on large language model (LLM)-based generative AI systems. Although these systems hold great promise for clinical decision support, it’s important to remember that even if they’re trained on large amounts of historic structured and unstructured data, they’ll need complete patient data to provide accurate decision support. So, although these powerful models will be transformative in building effective decision support systems, these systems will still need timely access to complete patient data to reach their true potential.
The foundation models that ingest large amounts of information and synthesize it for clinicians can prove to be a game changer here. Clinicians are challenged with absorbing incredibly large amounts of clinical literature in the form of publications and scientific studies about their specialty. If foundation models are used to consume and summarize scientifically validated data, it could help physicians to stay up-to-date. For maximum effectiveness, these models would need to be fed the most current publications on an ongoing basis. This practical application of the tool would benefit the clinician and healthcare in general without AI making definitive clinical decisions. This is important given the ongoing issues about having access to timely and complete patient data at the point of care.
Further into the future, the technology could play a valuable role by working alongside clinicians and helping them to improve the accuracy of diagnoses and the quality of treatment plans. Only a few months after the release of ChatGPT, foundation models were already being used to create a medical version of this. I’ve been testing two different versions and found them to be accurate and impressive. The use case here is that the clinician could enter prompts and receive responses. Ideally, the model would have access to patient data and proactively provide suggestions to the clinician based on its training in valid medical literature. The bottleneck here would be the limited access to complete and accurate patient data.
Several EHR vendors are already rolling out initial versions of medically trained GPT models inside their systems to help physicians and patients. Epic’s initial rollout of this includes support for answering medical questions from patients and is being tested with customers like UC San Diego Health, UW Health in Madison, Wisconsin, and Stanford Health Care. It is important to keep in mind thatgenerative AI applications like ChatGPT aren’t currently suited to helping clinicians treat patients because they pull from existing medical and popular literature to answer clinical questions and therefore aren’t as accurate as they should be for medical use cases. As such, medical versions of Chat-GPT are being developed and as I mentioned earlier, they’re already showing good results.
