Evaluation of a Large Language Model to Identify Confidential Content in Adolescent Encounter Notes

Abstract

AbstractIntroductionIn adolescent care, information sharing through patient portals can lead to unintentional disclosures to patients’ guardians around protected health topics such as mental health, sexual health, and substance use. A persistent challenge facing pediatric health systems is configuring systems to withhold confidential information recorded as free text in encounter notes. This study evaluates the accuracy of a proprietary large language model (LLM) in identifying content relating to adolescent confidentiality in such notes.MethodsA random sample of 300 notes were selected from outpatient adolescent encounters performed at an academic pediatric health system. The notes were manually reviewed by a group of pediatricians to identify confidential content. A proprietary LLM, GPT-3.5 (OpenAI, San Francisco, CA), was prompted using a “few-shot learning” method to identify the confidential content within these notes. Two primary outcomes were considered: (1) the ability of the LLM to determine whether a progress note contains confidential content and (2) its ability to identify the specific confidential content within the note.ResultsOf the 300 sampled notes, 91 (30%) contained confidential content. The LLM was able to classify whether an adolescent progress note contained confidential content with a sensitivity of 97% (88/91), specificity of 18% (37/209), and positive predictive value of 34% (88/260). Only 40 of the 306 manually reviewed excerpts (13%) were accurately derived from the original note (ie. contained no hallucinations), 22 (7%) of which represented the note’s actual confidential content.DiscussionA proprietary LLM achieved a high sensitivity in classifying whether adolescent encounter notes contain confidential content. However, its low specificity and poor positive predictive value limit its usefulness. Furthermore, an alarmingly high fraction of confidential note excerpts proposed by the model contained hallucinations. In its current form, GPT-3.5 cannot reliably identify confidential content in free-text adolescent progress notes.