Figure 1.

Validation of keywords-based sentiment and other common classifiers on labeled note excerpts for Multiparameter Intelligent Monitoring of Intensive Care III cohort. A, Spearman correlation (Corr) of each measure with ordinal, five-value sentiment labels. “Positive” represents the count of positive keywords in a note. Negative represents the count of negative keywords in a note. B, A correlogram whereby measures are sorted for display based on similarity from hierarchical clustering. For the keywords sentiment measures, observations without any sentiment keywords were given a neutral value of 0.5. The keywords sentiment score is inversely correlated with the DeBERTa-v3 sentiment score because a higher DeBERTa-v3 score represents a more positive sentiment, whereas for the keywords score, a higher score represents a more negative sentiment. Spearman Corr was used to validate our clinical sentiment measures with the ordinal labels of clinical sentiment recorded on the note excerpts. Each of the sentiment measures is a continuous score. DeBERTa = Decoding-enhanced Bidirectional Encoder Representations from Transformers with disentangled attention.