How Is AI Used in Healthcare?

Start with the problem to be solved

Some projects examine images; others organize information or assist with drafting and searching. WHO's guidance on large multi-modal models discusses possible applications across care, research and administration while emphasizing risks from unreliable outputs. ^[1]

Before adopting a tool, define the user, the task and the consequence of an error. This gives the team a practical basis for choosing an evaluation method.

Training and retrieval are different

Image-model training adjusts model parameters using examples. Retrieval-augmented generation, or RAG, retrieves relevant material from a corpus to help a language model compose an answer. Adding documents to an index does not, by itself, retrain the language model. ^[2]

For a research team, both processes need records. Training should identify the data and model version. Retrieval should identify which sources were found and how they support the response. This technical distinction is important when describing what a product actually does.

Evaluate the output people use

Our recommendation is to check usefulness at the point of work. Can a reader open the source behind an answer? Can a reviewer recognize an unsupported statement? Does the system make uncertainty visible? Does it preserve enough context to investigate an error?

Internal testing can assess whether workflows function as designed. It should not be described as proof of clinical effectiveness, regulatory authorization or guaranteed savings.

CellSight's application

CellSight's team reports that its working first version implements the designed functionality and backend, including RAG, and has undergone internal testing. Its research platform connects image workflows, batch analysis, evidence and AI assistance.

This does not support an unqualified claim that CellSight diagnoses leukemia. The product remains subject to its stated use boundary: not for diagnosis or clinical decision-making.

Useful communication about healthcare AI should explain the task, the evidence and the limitation together.