Table 4.

Performance of SVM classifier at the encounter level^a

Feature generation methods	AUC-ROC	F1-score
Sample: N = 46 patients, N = 46 nurses
Performance of Individual feature generation method (no feature selection method was used)
TF-IDF	97.45 ± 2.36	93.13 ± 4.87
POS-tagging	31.48 ± 15.27	64.17 ± 12.91
N-grams	89.45 ± 10.79	85.21 ± 6.48
LIWC	95.67 ± 4.55	82.59 ± 11.15
Word2Vec	92.51 ± 4.05	83.02 ± 8.52
UMLS	77.75 ± 5.48	77.36 ± 3.16
Performance of combination of feature generation methods after selecting the most informative features using JMIM method
TF-IDF (JMIM)+Word2Vec (JMIM)	99.01 ± 1.97	93.66 ± 5.16
TF-IDF (JMIM)+Word2Vec (JMIM)+LIWC (JMIM)	98.54 ± 1.81	91.07 ± 5.84
TF-IDF (JMIM)+Word2Vec (JMIM)+LIWC (JMIM)+Unigram (JMIM)+UMLS (JMIM)	99.28 ± 0.98	96.82 ± 4.1

Note: The feature generation method(s) that demonstrated the highest performance, as indicated by the AUC-ROC and F1-score values, was highlighted in bold.

^aUtterances for each identified speaker (speaker 1 and speaker 2) were aggregated at encounter level.