Table 1.
Number of tokens by tokenizers.
| Model and metric | Breast cancer | Colorectal cancer | |||
|
|
Training and validation (n=23,789) | Test (n=4486) | Training and validation (n=23,977) | Test (n=6045) | |
| Bio+ClinicalBERT | |||||
|
|
Mean (SD) | 486 (323) | 548 (358) | 488 (316) | 547 (357) |
|
|
Maximum | 2821 | 2717 | 2987 | 3619 |
|
|
EHRsa >512, n (%) | 8860 (37.24) | 1955 (43.58) | 9186 (38.31) | 2744 (42.84) |
| Clinical-BigBird | |||||
|
|
Mean (SD) | 451 (300) | 506 (331) | 451 (292) | 505 (331) |
|
|
Maximum | 2569 | 2523 | 2784 | 3342 |
|
|
EHRs >512, n (%) | 7719 (32.45) | 1724 (38.43) | 7866 (32.81) | 2368 (36.97) |
| Clinical-Longformer | |||||
|
|
Mean | 453 (301) | 508 (332) | 454 (293) | 508 (333) |
|
|
Maximum | 2584 | 2528 | 2796 | 3353 |
|
|
EHRs >512, n (%) | 7775 (32.68) | 1737 (38.72) | 7966 (33.22) | 2401 (37.49) |
aEHR: electronic health record.