The following are key points to remember from a state-of-the-art review on artificial intelligence (AI) and the revolutionizing of cardiology with large language models (LLMs):

1. Natural language processing (NLP) techniques aim to provide a computer with an understanding of human language, either spoken or written. The most state-of-the art NLP methods are based on LLMs, which most often generate text by predicting the most probable next sequence of words based on the prior words written. These LLMs are informed by a very large collection of sample texts.
2. In addition to generating new text based on a prompt, LLMs can also be used to derive meaning, understand, and analyze free text by recognizing specific concepts and their relationships in order to summarize, translate, answers to questions, and/or provide guidance.
3. NLP techniques are increasingly impacting clinical care for patients, clinicians, administrators, and researchers. Uses include automated generation of clinical notes and discharge letters, identifying and/or selecting medical term codes for billing, medical chatbots for patient and/or clinician guidance, cohort selection for clinical trials, and auditing purposes.
4. General LLMs (e.g., ChatGPT, Bard) are pretrained on publicly available data that contains few medical documents. Therefore, these models have limited understanding of the domain knowledge and are most likely to fail to generate comprehensive and/or medically accurate responses to specialized medication questions (e.g., those related to cardiovascular medicine).
5. To ensure that LLMs are clinically relevant, they need pretraining and fine-tuning using medical data with specific human feedback to reinforce domain-specific learning. Once this has been accomplished, then they can be used for a wide range of patient-, clinician-, and administrator-centric activities (e.g., providing medical summarization for patients, providing risk-assessment for clinicians, generating discharge letters and consultative notes for administrative purposes).
6. It is critical to protect patient privacy when training and deploying LLMs. This is particularly relevant as LLMs can inadvertently reveal their training data during use.
7. Users of LLMs need to be educated about the inherent biases and potential for false results based on the training materials used. For example, if all training materials are generated from North America and Europe, then the answers provided may be biased against individuals from Africa and Asia, including not appropriately recognizing diseases that may be highly prevalent in those regions but rare or nonexistent in North America and Europe.
8. Ideally, any LLM used in cardiovascular medicine and research would have three key principles: 1) clinicians and patients should trust the model, 2) use of the models should provide benefit, and 3) the models should be safe to use. While several trials are ongoing to assess each of these elements, transparency during model design, development, validation, and deployment will be essential.
9. Promising uses of LLMs in cardiovascular medicine include for patient cohort phenotyping and identification of adverse events. This can be useful when trying to identify previously unrecognized drug-related side effects or helping to identify uncommon diagnoses or diagnoses with unusual presentations.
10. Other uses of LLMs in cardiovascular medicine are to better predict patient trajectories based on wearable data, in-hospital measurement data (e.g., telemetry), laboratory values, and medical record documentation. In combination with advanced assessment of imaging tests, these LLMs can help clinicians better identify patients at risk for decompensation who may benefit from early intervention.
11. Among the most common current uses of LLMs are for patient and clinician interaction to enhance care. This includes the use of medical chatbots who can provide answers to specific questions. For patients, this may be a way to provide answers to key questions without burdening health care providers. However, it is also possible that use of these chatbots could increase health care contact by raising new concerns for patients. It is important that answers provided by LLM chatbots be fact checked on validity by professionals.
12. A final use of LLMs in cardiovascular care is for the administrative purpose of summarizing medical records, generating new medical records, or selecting billing codes.

Keywords: Artificial Intelligence, Cardiology, Natural Language Processing

< Back to Listings