Clinical prediction models are potentially effective, but very limited in a medical sense [Moor et al 2023]: (i) the models are often task-specific for one particular disease, intervention or clinical outcome. For example, each of the 500 FDA AI models is approved for 1 or 2 specific tasks [FDA 2022]; (ii) they are developed and applied unimodally, i.e., with a single type of data (modality), i.e., tables, images or free text; and (iii) they do not use contextual information, for example about pathophysiological processes, but are often based solely on identifying statistical associations. These limitations have major implications: (i) the models have little effective use in medical decision-making, because much more information (multimodal and contextual) is considered than in a prediction model. For example, laboratory diagnostics is involved in more than 70% of medical decisions [NVKC 2022]; and (ii) task specificity means that a multitude of models are needed (so to speak for each disease/intervention/outcome), each of which is developed, implemented and maintained almost separately. Our aim is to develop a clinical foundation model with laboratory data and contextual information. Just as ChatGPT was developed based on text data from the internet with a chatbot as an application, we propose to develop a foundation model based on laboratory data with a series of prediction models as an application.
LabGPT
