|
--- |
|
license: mit |
|
tags: |
|
- medical |
|
--- |
|
|
|
# Clinical-T5 Models |
|
We train four different T5 variants on the union of MIMIC-III and MIMIC-IV: (1) Initialized from T5-Base, |
|
(2) Initialized from SciFive-Base, (3) T5-Base initialized from scratch, and (4) T5-Large initialized from scratch. |
|
|
|
This particular model card describes the T5-Large model trained from scratch on MIMIC notes. |
|
|
|
More information about the models and training procedure can also be found in the [preprint](https://arxiv.org/abs/2302.08091). |
|
|
|
# Model Pretraining |
|
In this section, we will describe the pretraining procedure. |
|
|
|
### Pretraining Data |
|
We train on the union of MIMIC-III and MIMIC-IV. MIMIC-III contains a wide variety of note types, whereas MIMIC-IV contains only radiology reports and discharge summaries. We remove duplicate notes. This results in ~1.2B words. |
|
|
|
### Note Preprocessing |
|
We make two important preprocessing steps: |
|
* We replace all DEID tags with special tokens. For example, `"The patient, [**First Name 123**], has a history of high blood pressure"` is replaced with `"The patient, [NAME], has a history of high blood pressure"`. |
|
* We remove any duplicate notes based on edit times. There are roughly ~300M/800M words from MIMIC-III, which are repeats of the same note, with only a few words changed! This is due to the fact that a nurse might save a note, and then edit it 10 minutes later. Both would appear. |
|
|
|
### Pretraining Procedures |
|
We train the Clinical-T5-Large model from scratch using a cased-vocab of 32,000. We train it for 780,000 steps, using a batch size of 12 per TPU pod (8 pods total), and a sequence length of 512. |
|
This results in a batch size of 49,152. Accounting for the number of steps, this equates to 38B tokens. We were aiming for 40B, but our Google Cloud instance broke! We use the same LR schedule as the original T5 paper. |
|
|
|
We train the Clinical-T5-Scratch model using an uncased vocab of 32,0000. This model is trained for 28 epochs total, with a sequence length of 512 (~40B tokens total). We use the same LR schedule as the original T5 paper. |
|
|
|
As mentioned previously, we also initialize train two models initialized from T5-base and SciFive. These are trained for ~13B tokens, using a batch size of 32 per GPU (8 GPUs), and a sequence length of 512. In an attempt to speed up training, and help the models quickly adapt, we increase the warm-up steps from 10K to 40K. This helps the model initialized from T5-Base, but not SciFive. |
|
|
|
# How to use the Model |
|
You will first need to have credentialed PhysioNet access to use model. Why? There is reasonable evidence that these models contain leakage, especially the larger ones. Releasing a model that leaks these notes would be a data-use agreement violation. To get PhysioNet access, you must pass the CITI training. |
|
Once you have PhysioNet, access the model by doing the following: |
|
``` |
|
wget -r -N -c -np --user "INSERT_USER" --ask-password https://physionet.org/files/clinical-t5/1.0.0/ |
|
``` |
|
|
|
Then, you can load the model + tokenizer: |
|
``` |
|
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM |
|
tokenizer = AutoTokenizer.from_pretrained(INSERT_PATH_TO_MODEL_FOLDER) |
|
model = AutoModelForSeq2SeqLM.from_pretrained(PATH_TO_MODEL_FOLDER) |
|
``` |
|
|
|
# Tips |
|
Use the models initialized from scratch! Based on our preliminary results, we find that these are best. |
|
|
|
# Questions? |
|
If you have any questions about using the models, please email eric@xyla.com. |
