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

Introduction: In 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.

A 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.

Of 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.

A 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. All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprintthis version posted August 28, 2023. ; https://doi.org/10.1101/2023.08.25.23294372doi: medRxiv preprint

In adolescent care, information sharing through patient portals can lead to unintentional disclosures to patients’ guardians around protected health topics like mental health, sexual health, and substance use. Maintaining confidentiality in the wake of the 21st Century Cures Act is crucial to providing equitable care to this population1. Pediatric institutions have employed a variety of methods to prevent such unintentional disclosures, including configuring specific types of medical information from being released to patient portals2. A persistent challenge is withholding confidential information recorded as free text in visit notes, up to 21% of which have been shown to contain confidential information3. Electronic health records lack functionality to identify and filter such information, and while natural language processing algorithms have shown promise4, there is no widely accessible solution. The emergence of large language models (LLMs) offers a potential solution. LLMs are capable of summarizing large bodies of text and extracting relevant clinical information5. The ability of LLMs to accurately identify the presence of confidential content in clinical notes has not been studied. We evaluate the accuracy of a proprietary LLM in identifying content related to mental health, sexual health, and substance use in adolescent progress notes.

A random sample of 300 visit notes were selected from outpatient adolescent encounters performed at an academic pediatric health system between January 1, 2016 and December 31, All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprintthis version posted August 28, 2023. ; https://doi.org/10.1101/2023.08.25.23294372doi: medRxiv preprint 2019. The notes were manually reviewed by a group of pediatricians specifically trained in adolescent confidentiality laws to identify confidentiality disclosures3,6. A proprietary LLM, GPT-3.5 (OpenAI, San Francisco, CA), was prompted to identify confidential content within these notes. The LLM was accessed via a secure computing platform approved for use with protected health information. Two primary outcomes were considered: (1) the ability of the LLM to determine whether a progress note contains confidential content (note classification); and (2) its ability to identify the specific confidential content within the note (excerpt extraction). For the note classification task, sensitivity, specificity, and positive predictive value were calculated, including 95% confidence intervals assuming a binomial distribution. For excerpt extraction, the model’s proposed excerpts were compared to the original note content to verify accuracy, as LLMs are liable to “hallucinations.” Hallucinations are model outputs that–while plausible–are incorrect or unrelated to the input. Excerpts were also reviewed by a physician to evaluate whether they correctly refer to the actual confidential concepts within the note. Figure 1 illustrates the structure of the model prompt. A “few-shot learning” approach was used, where the model is given both definitions and examples of confidential content derived from real clinical notes. The prompt is available in the Supplement.

Of 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 All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprintthis version posted August 28, 2023. ; https://doi.org/10.1101/2023.08.25.23294372doi: medRxiv preprint 97% (88/91), specificity of 18% (37/209), and positive predictive value of 34% (88/260) (Table 1). The LLM proposed a total of 768 confidential excerpts, 306 of which were from a note with confidential content and were reviewed. Only 40 excerpts (13%) were accurately derived from the original note (ie. 87% of excerpts contained a hallucination), and 22 (7%) of excerpts reflected the note’s actual confidential content.

A proprietary LLM achieved a high sensitivity in classifying whether adolescent notes contain confidential content. However, its low specificity and poor positive predictive value limit its usefulness in hospital operations. Furthermore, the proportion of excerpts containing hallucinations is alarmingly high, such that the output of the model cannot be used for regulatory purposes. With this prompt structure, OpenAI’s GPT-3.5 is not able to reliably identify confidential content in free-text adolescent progress notes. All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprintthis version posted August 28, 2023. ; https://doi.org/10.1101/2023.08.25.23294372doi: medRxiv preprint Figure Legends Figure 1. Architecture of the GPT prompt. There are five portions of the prompt provided to the large language model. “Instructions” provide a general description of the problem; “Context” provides the model with further details; “Task” explicitly defines the specific task; “Output” specifies the desired formatting of the model’s response; “Sample” provides the clinical note to be interpreted by the model. The right column shows text from the actual prompt provided to the system. All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprintthis version posted August 28, 2023. ; https://doi.org/10.1101/2023.08.25.23294372doi: medRxiv preprint Tables Table 1. Performance results from GPT-3.5’s ability to identify confidential content in adolescent visit notes. Brackets report 95% confidence intervals. Confidential content note classification performance Sensitivity 97% (88/91) [91% - 93%] Specificity 18% (37/209) [13% - 24%] Positive Predictive Value 34% (88/260) [28% - 40%] Confidential content excerpt performance Excerpt is accurately derived from note (no hallucination) 13% (40/306) Excerpt correctly identifies confidential content from the note 7% (22/306) All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprintthis version posted August 28, 2023. ; https://doi.org/10.1101/2023.08.25.23294372doi: medRxiv preprint

1. Pasternak RH, Alderman EM, English A. 21st Century Cures Act ONC Rule: Implications for Adolescent Care and Confidentiality Protections. Pediatrics. 2023 Apr 1;151(Suppl 1). 2. Parsons CR, Hron JD, Bourgeois FC. Preserving privacy for pediatric patients and families: use of confidential note types in pediatric ambulatory care. J Am Med Inform Assoc. 2020 Nov 1;27(11):1705–10. 3. Bedgood M, Rabbani N, Brown C, Goldstein R, Carlson JL, Steinberg E, et al. The Prevalence of Confidential Content in Adolescent Progress Notes Prior to the 21st Century Cures Act Information Blocking Mandate. Appl Clin Inform. 2023 Mar;14(2):337–44. 4. Rabbani N, Bedgood M, Brown C, Steinberg E, Goldstein RL, Carlson JL, et al. A Natural Language Processing Model to Identify Confidential Content in Adolescent Clinical Notes. Appl Clin Inform. 2023 May;14(3):400–7. 5. Lee P, Goldberg C, Kohane I. The AI Revolution in Medicine: GPT-4 and Beyond. Pearson Education, Limited; 2023. 6. Sharko M, Jameson R, Ancker JS, Krams L, Webber EC, Rosenbloom ST. State-by-State Variability in Adolescent Privacy Laws. Pediatrics. 2022 Jun 1;149(6). All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprintthis version posted August 28, 2023. ; https://doi.org/10.1101/2023.08.25.23294372doi: medRxiv preprint Contributors Statement Page Naveed Rabbani: conceptualization, methodology, formal analysis, investigation, data curation, writing - original draft, writing - review & editing, project administration Conner Brown: methodology, software, data curation, formal analysis, writing - review & editing Michael Bedgood, Rachel Goldstein, and Jennifer Carlson: methodology, data curation, writing - review & editing Natalie Pageler: methodology, writing - review & editing Keith Morse: conceptualization, methodology, investigation, data curation, writing - original draft, writing - review & editing, supervision All authors approve the final manuscript as submitted and agree to be accountable for all aspects of the work. All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprintthis version posted August 28, 2023. ; https://doi.org/10.1101/2023.08.25.23294372doi: medRxiv preprint Instructions General instructions and

of the task to be completed by model. “The following is a multi-label classification task for confidential information found in progress notes at a large pediatric hospital…” Context Additional information to help the model complete the task. Includes definitions and real-life examples of confidential content. “Confidential information topics: mental health, sexual and reproductive health, drugs and alcohol. Mental health: words or phrases… "PHQ-9 Screen Score: 7", "mood changes, anxiety"…” Task Specific command to be completed by model. “Label the following sample as containing any of the following confidential information topics or not. Provide a list of any topics which are present.” Output Specifies formatting of the model’s answer. “Label: Excerpt: ” Sample Progress note content to be analyzed. “This is a 17-year-old male who presents for…” GPT Prompt Architecture All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprintthis version posted August 28, 2023. ; https://doi.org/10.1101/2023.08.25.23294372doi: medRxiv preprint