Model Card for Labrador

Labrador is a pre-trained continuous Transformer model for masked lab modeling.

Model Details

Model Description

Laboratory data are a rich source of information about a patient's health. They are often used to diagnose and monitor disease, and to guide treatment. However, lab values are continuous, often missing and therefore difficult to model with the Transformer architecture. Labrador solves this problem by jointly embedding lab values with a token for the lab test identifier so that the quantitative and qualitative information from each test is combined into a single representation.

Labrador is pre-trained on a large corpus of 100 million lab tests from over 260,000 patients. We rigorously evaluate Labrador on intrinsic and extrinsic tasks, including four real-world problems: cancer diagnosis, COVID-19 diagnosis, predicting elevated alcohol consumption and ICU mortality due to sepsis. We find that Labrador is superior to BERT across all evaluations but both are outperformed by XGBoost indicating that transfer learning from continuous EHR data is still an open problem.

We discuss the limitations of our approach and suggest future directions for research in the corresponding paper, Labrador: Exploring the Limits of Masked Language Modeling for Laboratory Data.

Developed by: Anonymous Submitters
Model type: BERT-style transformer
License: MIT

Uses

Direct Use

The base models can be used directly to impute lab values and/or MIMIC lab codes conditional on a set of lab values and lab codes.

Downstream Use

The associated codebase includes a fine-tuning wrapper class that can be used to repurpose these base models for downstream regression or classification tasks.

Bias, Risks, and Limitations

These models were solely pre-trained on patient data from MIMIC-IV. This population is not representative of all patients and therefore the statistical patterns that these models learned will not apply equally well to all individuals.

Recommendations

Caution should be used when applying these models to downstream prediction tasks. Be sure to include a fairness assessment in your evaluations in order to assess model bias.

How to Get Started with the Model

See the Get Started instructions in the README of the associated codebase.

Training & Evaluation Details

See the associated publication and codebase.

Environmental Impact

Hardware Type: A100 PCIe 40GB
Hours used: 240 (for pre-training, not counting fine-tuning evaluations)
Cloud Provider: Private infrastructure
Compute Region: N/A
Carbon Emitted: 50 kg CO2 eq.

Carbon emissions were estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).

Citation

If you use Labrador in your work, please cite:

BibTeX:

APA:

Model Card Contact

This is an anonymous model card for a conference submission.