Table 6 Summary of models performing structured-longitudinal EHR analysis
Study | Data sources | Model architecture | Biomedical tasks |
|---|---|---|---|
BEHRT [107] | CPRD | Transformer-based encoder | Disease prediction |
Hi-BEHRT [108] | CRPD | Hierarchical BEHRT | Disease prediction |
G-BERT [106] | MIMIC-III | Graph neural network and BERT | Drug recommendation |
BRLTM [119] | UCLA EHR data | Transformer-based encoder | Disease prediction |
Med-BERT [109] | Cerner Health Facts®, Truven Health MarketScan® | Transformer-based encoder | Disease prediction |
ExMed-BERT [110] | IBM Explorys Therapeutic Dataset | Extended Med-BERT | Disease prediction |
CEHR-BERT [111] | CUIMC-NYP OMOP | Transformer-based encoder with additional FFN for temporal embedding | Various predictive tasks (disease, readmission, death, hospitalization) |
Med-PLM [120] | MIMIC-III | G-BERT / Med-BERT + ClinicalBERT + Cross-modal module | ICD coding, readmission, drug recommendation |
TransMED [118] | STARR OMOP | Hierarchical use of BERT | Hospital stay, ventilation risk |
T3Net [114] | KPMAS | Transformer-based encoder | Hospitalization and mortality prediction |
TAPER [113] | MIMIC-III | Transformer-based encoder, BERT | Readmission and mortality prediction |
CEHR-GPT [121] | EHR data from the Columbia University Irving Medical Center-New York Presbyterian Hospital | Transformer-based decoder, GPT | Generation of synthetic EHR data |