Table 3. Comparison of the accuracy of different BERT (Bidirectional Encoder Representations from Transformers) models.

Model size and model name	All codes, 95% CI		Top 80% codes, 95% CI
	Multilabel	Top-5	Multilabel	Top-5
Base
NorDeClin-BERT-base-NorICDa	0.44 (0.43-0.45)	0.81 (0.80-0.81)	0.54 (0.53-0.55)	0.87 (0.86-0.88)
SweDeClin-BERT-SweICD	0.25 (0.24-0.26)	0.59 (0.58-0.60)	0.35 (0.34-0.36)	0.65 (0.63-0.65)
SweDeClin-BERT-NorICD	0.40 (0.39-0.41)	0.78 (0.77-0.79)	0.50 (0.49-0.51)	0.85 (0.84-0.86)
ScandiBERT-NorICD	0.39 (0.38-0.40)	0.78 (0.77-0.79)	0.51 (0.50-0.52)	0.85 (0.84-0.86)
NorBERT3-base-NorICD	0.43 (0.42-0.44)	0.80 (0.79-0.81)	0.52 (0.51-0.53)	0.86 (0.86-0.87)
Large
NorDeClin-BERT-large-NorICD	0.47 (0.46-0.48)b	0.82 (0.82-0.83)b	0.56 (0.55-0.57)b	0.88 (0.88-0.89)b
NorBERT3-large-NorICD	0.42 (0.41-0.43)	0.81 (0.80-0.82)	0.53 (0.52-0.54)	0.88 (0.87-0.88)b

^aICD-10: International Statistical Classification of Diseases, Tenth Revision.

^bHighest score for each scenario.