Temporal Analysis of Patient-Centered Sentiment in Clinical Notes for Patients With Mental Health Conditions: Retrospective Cohort Study

The study moves clinical sentiment analysis from overall note tone to the patient's distinct perspective, labeling the same records from patient, physician, and general perspectives. Existing clinical sentiment analysis primarily evaluates overall note tone, whereas this work explicitly separates the patient's own perspective from the provider perspective and notes that they can differ. Based on 6,382 patients and 16,447 notes from MIMIC-IV, focusing on the Brief Hospital Course and Discharge Instructions sections, with a large sample and an explicit perspective design.

Sentiment trajectories change over time in a section-dependent way: patients with multiple admissions show substantial directional change, and Discharge Instructions fluctuate more than Brief Hospital Course notes. Prior work often stays with sentiment polarity in a single note, while this study quantifies longitudinal trends at the patient level using Kendall τ and reveals temporal dynamics that differ by note section. Trends are quantified at the patient level with Kendall τ across patients with multiple admissions, as a temporal description within a retrospective cohort design.

Large language models align better with patient-centered sentiment annotations than lexicon-based methods, although overall performance remains limited. The study directly compares two LLMs with three lexicon tools on the same task and highlights the strengths and weaknesses of both approaches. Model performance is evaluated on a manually annotated subset (n=165) using precision, recall, and F1-score; the subset is small, so the conclusion is framed as relatively better rather than sufficiently reliable.

Discharge-note sentiment trajectories are associated with postdischarge mortality: patients who died within 30-90 days of discharge had significantly more negative discharge-note sentiment than survivors across both LLMs. The study links patient-perspective sentiment to postdischarge mortality, suggesting that sentiment signals in clinical narratives may carry clinical relevance. Associations between sentiment patterns and mortality were assessed with independent-samples (Welch) t tests, and the result was consistent across both LLMs, representing observational association rather than causal evidence.

This work applies to hospitalized patients in the MIMIC-IV database who have ICD-10 mental health diagnoses and whose records include Brief Hospital Course and Discharge Instructions sections, with results framed around patient, physician, and general sentiment perspectives and a 30-90 day postdischarge mortality window. It provides a starting point for validating patient-perspective sentiment trajectories in larger, more diverse cohorts, for improving sentiment modeling oriented to clinical narratives, and for incorporating section-dependent temporal patterns into future clinical sentiment analysis designs.

A careful reader would still watch how well model performance evaluated on the n=165 manually annotated subset generalizes to larger annotation efforts; whether the association between patient-perspective sentiment and postdischarge mortality holds across different databases, diagnostic subgroups, and follow-up windows; and, given a fast parse that lacks figures and supplementary materials, what specific annotation criteria and error patterns remain to be clarified.