Table 1 Recent advances in the State-of-the-Art of NER applied to EHR
Model | Ref. | NER Dataset | Application | Entities | NER Performance | Entity Usage |
|---|---|---|---|---|---|---|
MC-BERT + BiLSTM + CNN + MHA + CRF | [4] | NER in clinical notes | 9 entities: Body, Treatment, Signs, Check, Disease, Lab, Medicine, Operation, Symptom | F1: 94.2%, 86.5%, 92.3% on CCKS17, CCKS19, cEHRNER | None | |
BiLSTM-CNN-Char | [8] | NER in clinical notes | 4 entities: Medical Problem, Treatment, Test, Drug | F1: 87.6%, 96.1%, 89.9% on i2b2/VA, 2014 n2c2, 2018 n2c2 | None | |
MUSA-BiLSTM-CRF | [12] | NERin clinical notes | 5 entities: Disease, Symptom, Examination, Treatment, Body part | F1: 92.0%, 91.8% on CCKS17, CCKS18 | None | |
BERT | [14] | 2018 n2c2 [11], 2009 n2c2 [15], 2010 n2c2 [9], 2012 n2c2 [16], ShARe13 [17] | NER in clinical notes | 4 entities: Drugs, Dosages, Reasons, Adverse drug events | F1: 90.0%, 80.9%, 88.4%, 87.5%, 82.6% on 2018 n2c2, 2009 n2c2, 2010 n2c2, 2012 n2c2, ShARe13 | None |
BERT-BiLSTM-CRF | [18] | NER in clinical notes | 1 entity: Disorder | F1: 79.9%, 80.7% on ShARe13, ShARe14 | None | |
BERT | [20] | NER in clinical notes | 3 entities: Medical Problem, Treatment, Tests | F1: 87.7%, 80.9% on i2b2-2010, VietBioNER | None | |
CancerBERT | [22] | Proprietary dataset (EHRs) | Breast cancer phenotypes | 8 entities: Hormone receptor type, Hormone receptor status, Tumor size, Tumor site, Cancer grade, Histological type, Tumor laterality, Cancer stage | F1: 87.6% | None |
scispaCy | [23] | MIMIC-III [24] | NER in clinical notes | 2 entities: Disease, Chemical | None | Mortality prediction |
med7 | [25] | MIMIC-III [24] | NER in clinical notes | 7 entities: Dosage, Drug, Duration, Form, Frequency, Route, Strength | None | Mortality prediction |
Rule-based | [26] | CCKS20 [27], gastroscopy text dataset, mixed dataset | Breast cancer phenotypes | 6 entities: Disease, Anatomy, Imaging, Lab, Drug, Operation | F1: 87.9%, 99.8%, 96.2% on CCKS20, gastroscopy text, mixed dataset | None |
Ensemble of CRF, multilingual Transformers (BERT, XLM RoBERTa) and LSTM | [28] | Proprietary dataset (hospital EHRs) | NER in clinical notes | 11 entities: Clinical Dept, Date, Duration, Evidential, Frequency, Occurrence, Problem, Test, Time, Treatment, Value | F1: 89.2% | None |
RoBERTa | [29] | Proprietary dataset (hospital EHRs) | Breast cancer information | 23 entities related to Breast Cancer domain | F1: 95.0% | None |