Table 5.

Experimental results from different NLP models. The test results are macroaverage classification values.

Methods (model name) and models		Precision (%)	Recall (%)	F1 score (%)
Convolution neural network (CNN)
	CNN-randoma	89.33	90.67	89.91
	CNN-fixed-Word2Vecb	88.01	93.12	90.43
	CNN-Word2Vecc	92.01	92.87	92.33
Long short-term memory
	Unidirectional long short-term memory	87.23	93.89	90.32
	Bidirectional long short-term memory	87.97	92.48	90.09
Transformer encoder
	Bidirectional encoder representations from transformers	86.44	89.69	87.99
	ELECTRAd-version 1	87.73	92.12	89.82
	ELECTRA-version 2	91.03	92.33	91.60
Data trimming
	CNN-Word2Vec (trimmede)	90.59	93.56	91.98
	Unidirectional long short-term memory (trimmed)	84.77	93.30	88.61
	ELECTRA-v2 (trimmed)	89.63	94.47	91.92
Ensemble combination
	CNN-Word2Vec + Uni-LSTM	89.53	94.24	91.76
	SCENTf-v1: CNN-Word2Vec + ELECTRA-v2 (trimmed)	91.10	94.18	92.56
	Unidirectional long short-term memory + ELECTRA-v2 (trimmed)	89.53	94.24	91.76
	CNN-Word2Vec + unidirectional long short-term memory + ELECTRA-v2 (trimmed)	91.02	94.19	92.52
Hierarchical ensemble
	CNN-Word2Vec and unidirectional long short-term memory + ELECTRA-v2 (trimmed)	91.30	92.86	91.92
	Unidirectional long short-term memory and CNN-Word2Vec + ELECTRA-v2 (trimmed)	86.83	93.88	90.09
	SCENT-v2: ELECTRA-v2 (trimmed) and CNN-Word2Vec + unidirectional long short-term memory	89.04	94.44	91.58

^aRandom: randomly initialized embedding.

^bFixed-Word2Vec: nontrainable pretrained Word2Vec embedding.

^cWord2Vec: trainable pretrained Word2Vec embedding.

^dELECTRA: efficiently learning an encoder that classifies token replacements accurately.

^eTrimmed: data sets are trimmed based on the keyword “thyroid” in the comprehensive medical examination text part.

^fSCENT: static and contextualized ensemble NLP network.