icf-levels-ins

Regression Model for Exercise Tolerance Functioning Levels (ICF b455)

Description

A fine-tuned regression model that assigns a functioning level to Dutch sentences describing exercise tolerance functions. The model is based on a pre-trained Dutch medical language model (link to be added): a RoBERTa model, trained from scratch on clinical notes of the Amsterdam UMC. To detect sentences about exercise tolerance functions in clinical text in Dutch, use the icf-domains classification model.

Functioning levels

The predictions generated by the model might sometimes be outside of the scale (e.g. 5.2); this is normal in a regression model.

Intended uses and limitations

• The model was fine-tuned (trained, validated and tested) on medical records from the Amsterdam UMC (the two academic medical centers of Amsterdam). It might perform differently on text from a different hospital or text from non-hospital sources (e.g. GP records).
• The model was fine-tuned with the Simple Transformers library. This library is based on Transformers but the model cannot be used directly with Transformers pipeline and classes; doing so would generate incorrect outputs. For this reason, the API on this page is disabled.

How to use

To generate predictions with the model, use the Simple Transformers library:

from simpletransformers.classification import ClassificationModel

model = ClassificationModel(
    'roberta',
    'CLTL/icf-levels-ins',
    use_cuda=False,
)

example = 'kan nog goed traplopen, maar flink ingeleverd aan conditie na Corona'
_, raw_outputs = model.predict([example])
predictions = np.squeeze(raw_outputs)

The prediction on the example is:

3.13

The raw outputs look like this:

[[3.1300993]]

Training data

• The training data consists of clinical notes from medical records (in Dutch) of the Amsterdam UMC. Due to privacy constraints, the data cannot be released.
• The annotation guidelines used for the project can be found here.

Training procedure

The default training parameters of Simple Transformers were used, including:

• Optimizer: AdamW
• Learning rate: 4e-5
• Num train epochs: 1
• Train batch size: 8

Evaluation results

The evaluation is done on a sentence-level (the classification unit) and on a note-level (the aggregated unit which is meaningful for the healthcare professionals).

Authors and references

Authors

Jenia Kim, Piek Vossen

References

TBD