Description
Assign assertion status to clinical entities extracted by Radiology NER based on their context in the text.
Predicted Entities
Confirmed, Suspected, Negative
How to use
Extract radiology entities using the radiology NER model in the pipeline and assign assertion status for them with assertion_dl_radiology pretrained model. Note: Example for demo purpose taken from: https://www.mtsamples.com/site/pages/sample.asp?Type=95-Radiology&Sample=1391-Chest%20PA%20&%20Lateral
documentAssembler = DocumentAssembler()\
.setInputCol("text")\
.setOutputCol("document")
sentenceDetector = SentenceDetector() \
.setInputCols(["document"]) \
.setOutputCol("sentence")
tokenizer = Tokenizer() \
.setInputCols(["sentence"]) \
.setOutputCol("token")
word_embeddings = WordEmbeddingsModel.pretrained("embeddings_clinical", "en", "clinical/models")\
.setInputCols(["sentence", "token"])\
.setOutputCol("embeddings")
radiology_ner = MedicalNerModel.pretrained("ner_radiology", "en", "clinical/models") \
.setInputCols(["sentence", "token", "embeddings"]) \
.setOutputCol("ner")
ner_converter = NerConverter() \
.setInputCols(["sentence", "token", "ner"]) \
.setOutputCol("ner_chunk")\
.setWhiteList(["ImagingFindings"])
radiology_assertion = AssertionDLModel.pretrained("assertion_dl_radiology", "en", "clinical/models") \
.setInputCols(["sentence", "ner_chunk", "embeddings"]) \
.setOutputCol("assertion")
nlpPipeline = Pipeline(stages=[documentAssembler, sentenceDetector, tokenizer, word_embeddings, radiology_ner, ner_converter, radiology_assertion])
empty_data = spark.createDataFrame([[""]]).toDF("text")
model = LightPipeline(nlpPipeline.fit(empty_data))
text = """
INTERPRETATION: There has been interval development of a moderate left-sided pneumothorax with near complete collapse of the left upper lobe. The lower lobe appears aerated. There is stable, diffuse, bilateral interstitial thickening with no definite acute air space consolidation. The heart and pulmonary vascularity are within normal limits. Left-sided port is seen with Groshong tip at the SVC/RA junction. No evidence for acute fracture, malalignment, or dislocation."""
result = model.fullAnnotate(text)
val documentAssembler = new DocumentAssembler()
.setInputCol("text")
.setOutputCol("document")
val sentenceDetector = new SentenceDetector()
.setInputCols("document")
.setOutputCol("sentence")
val tokenizer = new Tokenizer()
.setInputCols("sentence")
.setOutputCol("token")
val word_embeddings = WordEmbeddingsModel.pretrained("embeddings_clinical", "en", "clinical/models")
.setInputCols(Array("sentence", "token"))
.setOutputCol("embeddings")
val radiology_ner = MedicalNerModel.pretrained("ner_radiology", "en", "clinical/models")
.setInputCols(Array("sentence", "token", "embeddings"))
.setOutputCol("ner")
val ner_converter = NerConverter()
.setInputCols(Array("sentence", "token", "ner"))
.setOutputCol("ner_chunk")
.setWhiteList(Array("ImagingFindings"))
val radiology_assertion = AssertionDLModel.pretrained("assertion_dl_radiology", "en", "clinical/models")
.setInputCols(Array("sentence", "ner_chunk", "embeddings"))
.setOutputCol("assertion")
val nlpPipeline = new Pipeline().setStages(Array(documentAssembler, sentenceDetector, tokenizer, word_embeddings, radiology_ner, ner_converter, radiology_assertion))
text = """
INTERPRETATION: There has been interval development of a moderate left-sided pneumothorax with near complete collapse of the left upper lobe. The lower lobe appears aerated. There is stable, diffuse, bilateral interstitial thickening with no definite acute air space consolidation. The heart and pulmonary vascularity are within normal limits. Left-sided port is seen with Groshong tip at the SVC/RA junction. No evidence for acute fracture, malalignment, or dislocation."""
val data = Seq("text").toDS.toDF("text")
val result = pipeline.fit(data).transform(data)
Results
| | ner_chunk | begin | end | ner_label | assertion | assertion confidence |
|---:|:-------------------------|:------|:----|:----------------|:----------|:---------------------|
| 0 | left-sided | 66 | 75 | Direction | Confirmed | 0.9964 |
| 1 | pneumothorax | 77 | 88 | ImagingFindings | Confirmed | 0.9963 |
| 2 | complete collapse | 100 | 116 | ImagingFindings | Confirmed | 0.9977 |
| 3 | left upper | 125 | 134 | Direction | Confirmed | 0.9962 |
| 4 | lobe | 136 | 139 | BodyPart | Confirmed | 0.9913 |
| 5 | lower | 146 | 150 | Direction | Confirmed | 0.7678 |
| 6 | lobe | 152 | 155 | BodyPart | Confirmed | 0.8673 |
| 7 | aerated | 165 | 171 | ImagingFindings | Confirmed | 0.5755 |
| 8 | bilateral | 200 | 208 | Direction | Confirmed | 0.9966 |
| 9 | interstitial | 210 | 221 | BodyPart | Confirmed | 0.9944 |
| 10 | thickening | 223 | 232 | ImagingFindings | Confirmed | 0.9954 |
| 11 | air space consolidation | 257 | 279 | ImagingFindings | Negative | 0.9434 |
| 12 | heart | 286 | 290 | BodyPart | Confirmed | 0.9941 |
| 13 | pulmonary vascularity | 296 | 316 | BodyPart | Confirmed | 0.9986 |
| 14 | within normal limits | 322 | 341 | ImagingFindings | Confirmed | 0.9999 |
| 15 | Left-sided | 344 | 353 | Direction | Confirmed | 0.9782 |
| 16 | port | 355 | 358 | Medical_Device | Confirmed | 0.9838 |
| 17 | SVC/RA junction | 393 | 407 | BodyPart | Confirmed | 0.9998 |
| 18 | acute fracture | 426 | 439 | ImagingFindings | Negative | 0.9995 |
| 19 | malalignment | 442 | 453 | ImagingFindings | Negative | 0.9964 |
| 20 | dislocation | 459 | 469 | ImagingFindings | Negative | 0.9864 |
Model Information
| Model Name: | assertion_dl_radiology |
| Compatibility: | Healthcare NLP 5.3.3+ |
| License: | Licensed |
| Edition: | Official |
| Input Labels: | [document, ner_chunk, embeddings] |
| Output Labels: | [assertion_pred] |
| Language: | en |
| Size: | 2.5 MB |
References
Custom internal labeled radiology dataset.
Benchmarking
label precision recall f1-score support
Confirmed 0.95 0.92 0.94 3511
Negative 0.95 0.95 0.95 615
Suspected 0.79 0.87 0.82 1106
accuracy - - 0.91 5232
macro-avg 0.90 0.91 0.90 5232
weighted-avg 0.92 0.91 0.92 5232