. 2021 Oct 8;9(10):e29584. doi: 10.2196/29584

Table 2.

Accuracy comparisons.

			Accuracy		F score
Coder average			0.795861		0.739396
	Coder 1: EK	0.833211		0.796272
	Coder 2: SCM	0.775165		0.710356
	Coder 3: SD and CD	0.779206		0.711559
Neural network: no embeddings			0.436697		0.436697
Neural network: GloVe^a word embeddings			0.544954		0.457813
LSTM^b: no embeddings			0.631193		0.549997
LSTM+GloVe word embeddings			0.655046		0.593942
BERT^c: default weights			0.766972^d		0.718878
BERT: domain-specific			0.818349^d		0.775830

^aGloVe: Global Vectors for Word Representation.

^bLSTM: long short-term memory.

^cBERT: Bidirectional Encoder Representations from Transformers.

^dThe final accuracy scores for the Bidirectional Encoder Representations from Transformers–based models are based on selecting the best network from the results based on the results from the development data set. The reported numbers are from evaluating the training data set.