To view webinar recording of expert panel discussion, click here
Generative AI, a branch of artificial intelligence focused on large language models (LLMs), is rapidly evolving, and finding its way into clinical practice. Unlike traditional AI, which analyzes existing structured data, generative models can extract information from all modes of data storage including text, images, and voice, offering physicians a range of powerful tools to improve care workflows.
In this post, we explore how Generative AI can significantly benefit physicians in their clinical practice.
Document Creation
The prime capability of Generative AI is in generating content based on provided reference data. Specifically relevant to clinical practice are the capabilities in generating well-crafted textual data, summarization, and improving readability for different reader subsets.
Encounter Notes: Accurate recording of medical or surgical encounters is an essential element of a physician’s responsibility. A systematic approach to a patient also needs to be recorded systematically for recall, hand-off, and administrative purposes. Large Language Models can be trained on historical medical records to help generate accurate records of new encounters. The encounter can now be a natural conversation between the patient and the physician, which can now be automatically parsed into Subjective, Objective, Assessment, and Plan notes by such trained models.
Discharge Summary: After a clinical encounter has ended, summary of the encounter is provided to the patient with instructions for further management of the disease. In addition, the patient medical record itself may be shared with the next site of care for ensuring a continuity in patient management and avoiding duplication of tests, procedures, and medications. While interoperability of medical records has made a lot of progress over the past several years, clinical handoff still requires a lot of manual effort in generating summaries for different audiences. Generative AI can fill this gap today by generating different summaries for patient, family caregivers, as well as for clinical personnel to ensure a continuity in care.
Payer Communications: Prior Authorization submission and claims denial appeals require information extracted from siloed medical records to fill the forms required by multiple payers. This currently manual process can be largely or fully automated by using Generative AI to fill patient specific information extracted from the EMRs.
State of the art Large Language Models can be used to retrieve information from medical notes, lab and imaging reports, discharge summaries and other documents to assist physicians in clinical care. In addition, physicians will have easy access to information contained in other documents such as clinical guidelines, prior authorization requirements, reducing need to dig through information during clinical care.
Longitudinal Care Journey: Patients with multiple chronic conditions tend to have long records, myriad prescriptions, lab tests, and imaging studies, from multiple providers. Sifting through this information during a patient encounter to find a particular snippet can be time consuming and vulnerable to inaccuracies. Generative AI models can “read” such longitudinal patient records and retrieve temporal trends, disease and biomarker progression, unified prescription medication lists, and specific lab test information to assist clinical management.
Population Health Analytics: Practice-based Population Health Analytics is an important tool available to physicians to understand the cumulative impact of interventions on overall health of their member patients. However, ad hoc hypothesis testing is currently obstructed by data locked in EMR silos. Generative AI is not dependent on data being structured in a database. It can analyze free text medical records to identify similarity cohorts and provide rapid answers to hypothesis testing.
Anomaly Detection: An accurate diagnosis for a rare disease often takes a myriad of tests, multiple physicians, and often years to arrive. One of the most important causes of this avoidable patient suffering is a fragmented medical care system where different symptoms and diseases are managed by different physicians. However, integrated medical records allow Generative AI to analyze the “whole” patient and potentially identify such rare diagnoses earlier. Similar models can also be utilized to identify unintended anomalies in lab tests, medications, or other interventions.
Follow up care: Follow-up diagnostics, medical management, and specialist referrals are often suggested in medical notes, lab reports and imaging reports. Generative AI can consume all such reports for a patient and provide an appropriate scheduling and avoid missing follow ups.
Expert Clinical Companion
As Language Models are trained on the entirety of online information available at the time of training, these models can provide answers to specific clinical questions albeit with a rider of potential hallucinatory answers. “Grounding” of foundation models by Retrieval Augmented Generation (RAG) or fine-tuning on medical textbooks and peer reviewed journals can make these models far more intelligent in clinical question answering.
Second Opinion: Human body is a highly complex machine. Diagnosing a disorder and deciding on a treatment plan in this machine is equally complicated. A variety of signals arising from the body are synthesized and analyzed to arrive at such decisions. Even so there are times when a physician may want to get a second expert opinion to confirm or contradict these decisions. With Generative AI an expert is always available to the physician, which can synthesize unseen variables and identify new possibilities in diagnosis and treatment that may have been missed by the physician, while reducing the cognitive load of decision making.
Clinical Guidelines: Generative AI models can be trained on the organizational knowledge base and learn the best practices collected over years in medical notes. This knowledge can now be used to generate clinical guidelines for patient care and made available to physicians as real-time internal chat bots.
Transfer Learning: AI models trained on medical notes and clinical outcomes can identify which interventions are most suitable for a given patient profile. As expertise in medicine grows with experience, these models can help fill the experience gap by identifying effective interventions from highly experienced physicians and providing guidance to younger physicians and reducing variability in clinical care.
%20(1).webp)