Foundation Model

INSPECT releases the weights and code of a 143M-parameter foundation model with Time-To-Event objective pretrained on 2.57 million deidentified EHRs from Stanford Hospital.

The foundation model is based on the MOTOR architecture (Steinberg et al. 2023), which uses a next token prediction task to learn representations for patients.

Access#

These assets will be publicly shared in coming months, under a standard research Data Usage Agreement. We will update this page once the model is available.

Additional Details#

For more information, please read the original INSPECT paper.

Questions?#

For questions and feedback, please open an Issue on Github