Detect Assertion Status (assertion_dl_en)

Description

Deep learning named entity recognition model for assertions. The SparkNLP deep learning model (NerDL) is inspired by a former state of the art model for NER: Chiu & Nicols, Named Entity Recognition with Bidirectional LSTM-CNN.

Included Assertions

hypothetical, present, absent, possible, conditional, associated_with_someone_else.

Live Demo Open in Colab Download

How to use

Use as part of an nlp pipeline with the following stages: DocumentAssembler, SentenceDetector, Tokenizer, WordEmbeddingsModel, NerDLModel, NerConverter, AssertionDLModel.

...
clinical_assertion = AssertionDLModel.pretrained("assertion_dl", "en", "clinical/models") \
    .setInputCols(["sentence", "ner_chunk", "embeddings"]) \
    .setOutputCol("assertion")
    
nlpPipeline = Pipeline(stages=[documentAssembler, sentenceDetector, tokenizer, word_embeddings, clinical_ner, ner_converter, clinical_assertion])
model = nlpPipeline.fit(spark.createDataFrame([[""]]).toDF("text"))

light_result = LightPipeline(model).fullAnnotate('Patient has a headache for the last 2 weeks and appears anxious when she walks fast. No alopecia noted. She denies pain')[0]

...
val clinical_assertion = AssertionDLModel.pretrained("assertion_dl", "en", "clinical/models")
    .setInputCols(Array("sentence", "ner_chunk", "embeddings"))
    .setOutputCol("assertion")

val pipeline = new Pipeline().setStages(Array(documentAssembler, sentenceDetector, tokenizer, word_embeddings, clinical_ner, ner_converter, clinical_assertion))

val result = pipeline.fit(Seq.empty["Patient has a headache for the last 2 weeks and appears anxious when she walks fast. No alopecia noted. She denies pain"].toDS.toDF("text")).transform(data)

Results

The output is a dataframe with a sentence per row and an "assertion" column containing all of the assertion labels in the sentence. The assertion column also contains assertion character indices, and other metadata. To get only the entity chunks and assertion labels, without the metadata, select "ner_chunk.result" and "assertion.result" from your output dataframe.

|   | chunks     | entities | assertion   |
|---|------------|----------|-------------|
| 0 | a headache | PROBLEM  | present     |
| 1 | anxious    | PROBLEM  | conditional |
| 2 | alopecia   | PROBLEM  | absent      |
| 3 | pain       | PROBLEM  | absent      |

Model Information

Data Source

Trained on 2010 i2b2/VA challenge on concepts, assertions, and relations in clinical text with ‘embeddings_clinical’. https://portal.dbmi.hms.harvard.edu/projects/n2c2-nlp/

Benchmarking

|    | label                        | prec | rec  | f1   |
|---:|-----------------------------:|-----:|-----:|-----:|
| 0 | absent                        | 0.94 | 0.87 | 0.91 |
| 1 | associated_with_someone_else  | 0.81 | 0.73 | 0.76 |
| 2 | conditional                   | 0.78 | 0.24 | 0.37 |
| 3 | hypothetical                  | 0.89 | 0.75 | 0.81 |
| 4 | possible                      | 0.70 | 0.52 | 0.60 |
| 5 | present                       | 0.91 | 0.97 | 0.94 |
| 6 | Macro-average                 | 0.84 | 0.68 | 0.73 |
| 7 | Micro-average                 | 0.91 | 0.91 | 0.91 |