JMIR Publications

ABSTRACT

Background:

There is a considerable risk of stigmatization and harm when sensitive topics such as psychiatric diagnoses, substance abuse, and self-harm are recorded in electronic health records, especially since federal laws like the 21st Century Cures Act now require patients to have access to their own clinical notes. Discharge summaries are particularly problematic since they combine all hospital experiences and perform concurrent administrative, legal, and patient-facing tasks, although they are still not well researched in sensitive topic studies. There is a major methodological vacuum in understanding how the documentation of certain sensitive topics changes over time because current NLP techniques have concentrated on single-topic detection tasks with little attention to prevalence measurement or longitudinal documentation patterns.

Objective:

To evaluate an NLP based framework for the automated detection of a predefined set of sensitive topics and measuring their change in prevalence over time using the MIMIC-IV database.

Methods:

Discharge summaries from the MIMIC-IV database were filtered using ICD codes to identify psychiatrically relevant admissions, resulting in 2670 notes from 2108 distinct patients. Both explicit keyword-based and implicit semantic mentions of each sensitive subject category were found using a dual NLP detection framework. Normalized mention counts were used to compute weighted prevalence, which was then examined at the note and patient levels. Fisher’s exact test, McNemar’s exact test, and Benjamin Hochberg FDR correction were used to evaluate temporal change.

Results:

The weighted prevalence varied across sensitive categories, ranging from 2.68% to 18.93%. Patient level prevalence was consistently higher than note level prevalence across all categories. No statistical significance was observed after FDR correction, suggesting stable documentation over the study period. Improving, worsening, and absent trajectories remained relatively steady independent of overall category predominance, but persistent and mixed documentation patterns showed the most variance among categories, according to trajectory analysis.

Conclusions:

This study demonstrates the feasibility of a scalable NLP framework for multi-topic sensitive topic detection and prevalence tracking in clinical discharge summaries. The findings highlight the value of combining note-level and patient-level analyzes and provide a foundation for future work examining the impact of policy changes on sensitive topic documentation practices.