A-PROOF ICF-domains Classification

Description

A fine-tuned multi-label classification model that detects 9 WHO-ICF domains in clinical text in Dutch. 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.

ICF domains

The model can detect 9 domains, which were chosen due to their relevance to recovery from COVID-19:

ICF code	Domain	name in repo
b440	Respiration functions	ADM
b140	Attention functions	ATT
d840-d859	Work and employment	BER
b1300	Energy level	ENR
d550	Eating	ETN
d450	Walking	FAC
b455	Exercise tolerance functions	INS
b530	Weight maintenance functions	MBW
b152	Emotional functions	STM

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 MultiLabelClassificationModel

model = MultiLabelClassificationModel(
    'roberta',
    'CLTL/icf-domains',
    use_cuda=False,
)

example = 'Nu sinds 5-6 dagen progressieve benauwdheidsklachten (bij korte stukken lopen al kortademig), terwijl dit eerder niet zo was.'
predictions, raw_outputs = model.predict([example])

The predictions look like this:

[[1, 0, 0, 0, 0, 1, 1, 0, 0]]

The indices of the multi-label stand for:

[ADM, ATT, BER, ENR, ETN, FAC, INS, MBW, STM]

In other words, the above prediction corresponds to assigning the labels ADM, FAC and INS to the example sentence.

The raw outputs look like this:

[[0.51907885 0.00268032 0.0030862  0.03066113 0.00616694 0.64720929
  0.67348498 0.0118863  0.0046311 ]]

For this model, the threshold at which the prediction for a label flips from 0 to 1 is 0.5.

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
Threshold: 0.5

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).

Sentence-level

	ADM	ATT	BER	ENR	ETN	FAC	INS	MBW	STM
precision	0.98	0.98	0.56	0.96	0.92	0.84	0.89	0.79	0.70
recall	0.49	0.41	0.29	0.57	0.49	0.71	0.26	0.62	0.75
F1-score	0.66	0.58	0.35	0.72	0.63	0.76	0.41	0.70	0.72
support	775	39	54	160	382	253	287	125	181

Note-level

	ADM	ATT	BER	ENR	ETN	FAC	INS	MBW	STM
precision	1.0	1.0	0.66	0.96	0.95	0.84	0.95	0.87	0.80
recall	0.89	0.56	0.44	0.70	0.72	0.89	0.46	0.87	0.87
F1-score	0.94	0.71	0.50	0.81	0.82	0.86	0.61	0.87	0.84
support	231	27	34	92	165	95	116	64	94

CLTL
/

icf-domains