AI and large language models in electronic health records

In last week’s newsletter, we mentioned GatorTronGPT, a large language model (LLM) trained on medical notes. We’d like to share more studies related to AI and LLM in electronic health records (EHR) in this newsletter.

• A study published in Lancet Digital Health showcases the use of LLM for parsing clinical notes in musculoskeletal pain disorders. By fine-tuning the LLaMA-7B model on clinical notes, researchers were able to effectively extract pain characteristics such as location and acuity from clinical notes.

• In the same issue of Lancet Digital Health, a study used LLM to predict seizure recurrence from physicians' clinical notes of initial seizure-like events in children. The LLM was based on Clinical-Longformer, a transformer model with a 4,096 token size trained on long clinical notes. The study highlighted that domain-specific and location-specific pre-training of clinical LLMs can substantially boost their performance. It is understandable because the original LLM was trained on MIMIC dataset, which only has ICU patients from Beth Israel Deaconess Medical Center.

• A machine learning model was developed to rapidly identify mental health crises (e.g. suicidal/homicidal ideation, domestic violence, or non-suicidal self-injury) from electronic medical record chat messages. The model is a linear model with TF-IDF features. I wonder whether a fine-tuned LLM model could improve the performance and simplify the model development process, just like the above two studies.

• A systematic review in npj Digital Medicine explores the significant advancements and challenges in applying AI to neonatal ICU. The authors evaluated many AI tools and proposed future directions for integrating AI into neonatal ICU, signaling a significant shift in clinical practices and patient outcomes with AI. Another study in the same npj Digital Medicine issue introduced an ML model that predicts in-hospital cardiac arrest in ICU patients using electrocardiogram (ECG)-derived heart rate variability (HRV).

Other news worth mentioning

• Roche’s Genentech partners with Nvidia in AI drug deal (BioPharma Dive)

• RoseTTAFoldNA, using machine learning to predict the structure of proteins that bind to DNA and RNA (nature)

##########

If you find the newsletter helpful, please consider:

• 🔊 Sharing the newsletter with other people

• 👍 Upvoting on Product Hunt

• 📧 Sending any feedback, suggestions, and questions by directly replying to this email or writing reviews on Product Hunt

• 🙏 Supporting us with a cup of coffee.

Thanks, and see you next time!