EHR Pipelines
A post about creating EHR data pipelines for Deep Learning.
Assembling a single patient
- Using age span
- Other functions
ItemBase and Itemlist from fastai
- Patient
- PatientList
Side-note on Metrics
- Auroc score
- Multi-label or multi-task better than single-task
- Reason for multi-label
- Imbalanced dataset
- Separate train_pos_wts, valid_pos_wts (& test_pos_wts)
- Loss function + Last layer of models
- Moving sigmoid out of the model & into the loss function (BCEWithLogitsLoss)
Utils + Experiment class
- Eventually to use WandB or comet.ai
- Plotting
- Next steps are to use recalls or upgrade this whole thing to use fastai
- Tried to use fastai - but needed to do from scratch as a way to learn first