Foundation Model

How is a foundation model different from a typical AI algorithm?

Most traditional narrow healthcare algorithms are trained on a labelled dataset to complete one specific task. Foundation models are trained differently. They're typically pre-trained on enormous, often unlabelled, datasets using unsupervised learning techniques, learning the underlying structure and patterns in the data without needing every example to be manually labelled. Once this pre-training is complete, the model can be adapted ("fine-tuned") to specific tasks using a much smaller amount of task-specific data.

This is part of why Large Language Models (LLMs), like the ones used in tools such as ChatGPT, are considered a type of foundation model: they're trained on vast amounts of text and can then be applied to many different tasks, from answering questions to summarising documents, without being rebuilt each time.

What kinds of foundation models exist in healthcare?

Foundation models in healthcare are being developed across several types of data:

• Language-based models — trained on medical text, clinical notes, or research literature, and used for tasks like summarisation, clinical documentation, or answering medical questions. Google's Med-PaLM 2 is one example, trained to answer medical exam-style questions.
• Imaging models — trained on large volumes of medical images (X-rays, MRIs, pathology slides), and used for tasks like classification, segmentation, or anomaly detection. For example, H0-mini, an open-source pathology foundation model, created by Owkin together with Bioptimus.
• Genomic and biological models — trained on genomic, proteomic, or other molecular data, used to support tasks like drug discovery or biomarker identification. Evo, built by the Arc Institute, is an example.
• Multimodal models — like Harvard’s CONCH, trained across multiple types of data at once (e.g., imaging plus clinical text), aiming to combine insights the way a human clinician might.

Why are foundation models considered useful for healthcare?

Foundation models are seen as a promising direction for healthcare AI for a few reasons:

• They can help address the historical challenge of limited, task-specific datasets, since a foundation model's broad pre-training can make it more data-efficient when adapted to a new, narrower task.
• They're more adaptable than earlier "narrow" AI models, potentially reducing the need to build a completely new model for every clinical task.
• They can support zero-shot or few-shot learning, allowing a model to perform reasonably well on a new task with little or no additional labelled data, similar to how LLMs can respond to entirely new prompts without retraining.

What are the risks and limitations of foundation models in healthcare?

Foundation models come with real limitations that need to be carefully managed:

• Bias and representativeness — if the underlying training data isn't representative of diverse patient populations, the model's performance may vary significantly when applied to different groups. See bias for more detail on how this happens.
• Interpretability — foundation models are often large and complex, which can make it difficult to fully explain how a specific output was generated, a particular concern in clinical settings where "explainability" matters for trust and accountability.
• Validation and regulation — because foundation models can be adapted to many different downstream tasks, validating and regulating them for safe clinical use is more complex than for a traditional, single-purpose algorithm. Each specific clinical use case may still require its own validation.
• Computational cost — training and running large foundation models requires significant computational resources, which can be a barrier to access, particularly for smaller research groups or healthcare systems.

For these reasons, foundation models in healthcare are generally treated as a starting point, a broad base that still needs careful fine-tuning, testing, and validation before being applied to any specific clinical task.