TABLE 7.

Comparison of NLP performance for classification from clinicians’ notes.

Author	Task	NLP model	Embeddings and Corpus	AUC
Xia et al. (77)	Identify patients with complex neurological disorder	cTAKES-based	Train corpus: ∼600 clinical notes Test corpus: ∼500 clinical notes	0.96
Wissel et al. (104)	Identify candidates for epilepsy surgery	Multiple linear regression	Train corpus: ∼1,100 clinical notes Test corpus: ∼8,340 clinical notes	0.9
Heo et al. (85)	Prediction of stroke outcomes	CNN + LSTM + Multilayer perceptron	Train corpus: ∼1,300 clinical notes Test corpus: ∼500 clinical notes	0.81
Lineback et al. (78)	Prediction of 30-day readmission after stroke	Logistic regression + naïve Bayes + SVM + RF + Gradient boosting + XGBoost	Train corpus: ∼2,300 clinical notes Test corpus: ∼550 clinical notes	0.64
Lin et al. (81)	Identify UAU in hospitalized patients	Logistic regression	Train corpus: ∼58 k clinical notes	0.91
Bacchi et al. (87)	Prediction of cause of TIA-like presentations	RNN + CNN	Corpus: 2,201 clinical notes (∼150 words each)	0.88

NLP, natural language processing; AUC, area under curve; cTAKES, clinical text analysis and knowledge extraction system; CNN, convolutional neural network; LSTM, long short-term memory; SVM, support vector machine; RF, random forest; UAU, unhealthy alcohol use; ISP-D, internet-based self-assessment program for depression.