Integrating large language models for enhanced predictive analytics in healthcare - PubMed
5 hours ago
- #healthcare AI
- #predictive analytics
- #clinical decision support
- Introduces the Hopkins LLM framework, which leverages structured EHR data to develop and deploy clinical LLMs as multi-task predictive engines.
- The model uses the LLaMA architecture with 7 billion parameters, pre-trained on a comprehensive corpus, fine-tuned on 42,160 patients, and validated across three external health systems.
- Addresses key prediction tasks: 30-day all-cause readmissions, 90-day all-cause mortality, 30-day ICU admissions, and treatment recommendations, validated on 1,329 patients.
- Achieves a mean ROC-AUC of 0.84, showing a significant 0.28 improvement over zero-shot baseline LLMs, highlighting the potential for unified, user-friendly clinical prediction systems.