Distinguishing Clinical Sentiment in Intensive Care Unit Clinical Notes

Existing sentiment analysis models have yet to succeed in recognizing the clinical sentiment polarity in intensive care unit clinical notes. Conversely, natural language processing techniques have been through significant advancement and multiple paradigm shifts recently, enhancing the feasibility of improving clinical sentiment tasks on one hand, but increasing the difficulty in selecting the most effective approach. To address this shortcoming, this research empirically investigates existing sentiment analysis models when applied to critical care clinical notes. It seeks to provide insights on which approach has the potential to improve the quality of clinical sentiment recognition within this domain-specific frame. In this study, we compare nine selected models from three families: lexicon models, BERT models, and prompt models, and across the general and clinical domain. We compare the task on whole notes truncated at default limit by language models against the section of notes including assessment section onwards. Our dataset uses a proxy for sentiment, based on timing of notes relative to positive or negative outcomes such as discharge or death. The analysis concludes that the best performing models are Scifive, and clinicalBERT with accuracy ranges (0.80 - 0.93) depending on the type of notes, and that the worst performing models were Textblob and AFINN lexicons with accuracy ranges (0.48 - 0.54) depending on the type of notes. Based on our experiments, our proxy for sentiment is sufficiently accurate to support this line of experimentation, and segments of notes with assessments and plans provide comparable results, with slight improvement, to default truncation of whole notes.