FORMED is a foundation model for medical time series classification, which first achieves generalizable adaptation across different classification tasks.
FORMED is built and used in a three stage fashion:
- First, we use a pre-trained general purpose large time series model, e.g. TimesFM, as backbone foundation model for time series pattern extraction.
- Then we repurpose the model for medical time series classification with shared decoding attention mechanism.
- Now the model can be easily applied to new datasets with limited labeled data for adapting.
