broadinstitute/ml4h — reverse-engineered prompt

Build me a Python toolkit for machine learning on clinical and biomedical data. I want researchers to be able to gather different files in one place, turn raw formats like medical images, lab data, genetics, notes, XML, DICOM, NIFTI, and PNG into tensor files, then describe each piece of data with reusable TensorMaps so models know how to read it.

The tool should support multimodal, multitask modeling, so someone can connect those TensorMaps to trainable neural networks, choose losses and optimization settings, run experiments, and evaluate results with useful plots for domain experts. Please include notebooks or simple examples that show the full flow from raw data to training and evaluation.

Also make it practical to run locally with Docker, run tests, and optionally launch or use a Google Cloud VM for bigger jobs. Keep it research friendly, documented, and organized so new contributors can understand how to ingest, tensorize, model, and evaluate data.