About
Foundation models are quickly emerging as powerful tools to solve a variety of biomedical challenges, such as clinical text generation and summarization, radiograph analysis, disease prediction, etc. [1, 2, 3, 4, 5, 6]. Foundation models are characterized by their ability to solve multiple prediction tasks across diverse domains. These models are first pre-trained on vast quantities of unlabeled data, and then fine-tuned using task and domain specific examples. This stands in contrast to the mechanistic approach to machine learning in healthcare, which explicitly incorporates prior knowledge of physics/biology/chemistry into the models that are then refined by training from data. While both these approaches have the potential to streamline healthcare administration, reduce costs, enhance accessibility, and ultimately improve the quality of patient care, many questions remain unanswered:
- What are the key differences between the two modeling paradigms?
- What qualifies as a clinical foundation model?
- Which healthcare challenges can be effectively addressed by each of these approaches, and which remain beyond their respective scope?
- What are the challenges associated with training and applying these models in a healthcare context?
- What can these models learn and from what sources of information?
- How can such models be best integrated with the routines of healthcare professionals?
- How about challenges limiting their prospective adoption (usability, maintainability, trustworthiness, robustness, fairness, etc.)?
- How should these technologies be regulated?
References
- Zhang, Kai, et al. "BiomedGPT: A Unified and Generalist Biomedical Generative Pre-trained Transformer for Vision, Language, and Multimodal Tasks." arXiv preprint arXiv:2305.17100 (2023).
- Park, Sangjoon, et al. "AI can evolve without labels: self-evolving vision transformer for chest X-ray diagnosis through knowledge distillation" Nature communications 13.1 (2022): 3848.
- Li, Chunyuan, et al. "LLaVA-Med: Training a Large Language-and-Vision Assistant for Biomedicine in One Day." arXiv preprint arXiv:2306.00890 (2023).
- Zhang, Sheng, et al. "Large-Scale Domain-Specific Pretraining for Biomedical Vision-Language Processing." arXiv preprint arXiv:2303.00915 (2023).
- Moor, Michael, et al. "Foundation models for generalist medical artificial intelligence." Nature 616.7956 (2023): 259-265.
- Steinberg, Ethan, et al. "Language models are an effective representation learning technique for electronic health record data." Journal of biomedical informatics 113 (2021): 103637.