This paper is in the following e-collection/theme issue:

Generative Language Models Including ChatGPT (309) Clinical Informatics (1018) Ontologies, Classifications, and Coding (239) Natural Language Processing (690) Clinical Information and Decision Making (1404) Artificial Intelligence (1369) Homelessness (31) Social Determinants of Health (32) Ethical, Legal, and Social Issues in AI (32) Personal Health Records, Patient-Accessible Electronic Health Records, Patient Portals (571) Electronic Health Records (1051)

Published on in Vol 26 (2024)

Preprints (earlier versions) of this paper are available at https://preprints.jmir.org/preprint/63445, first published .
Using Large Language Models to Abstract Complex Social Determinants of Health From Original and Deidentified Medical Notes: Development and Validation Study

Using Large Language Models to Abstract Complex Social Determinants of Health From Original and Deidentified Medical Notes: Development and Validation Study

Using Large Language Models to Abstract Complex Social Determinants of Health From Original and Deidentified Medical Notes: Development and Validation Study

Authors of this article:

Alexandra Ralevski1, 2 Author Orcid Image ;   Nadaa Taiyab3 Author Orcid Image ;   Michael Nossal2 Author Orcid Image ;   Lindsay Mico2 Author Orcid Image ;   Samantha Piekos1 Author Orcid Image ;   Jennifer Hadlock1, 4 Author Orcid Image
Article Authors Cited by Tweetations Metrics

    Original Paper

    1Institute for Systems Biology, Seattle, WA, United States

    2Providence Health & Services, Renton, WA, United States

    3Tegria, LLC, Madison, WI, United States

    4Department of Biomedical Informatics and Medical Education, University of Washington, Seattle, WA, United States

    Corresponding Author:

    Jennifer Hadlock, MD

    Institute for Systems Biology

    401 Terry Ave N

    Seattle, WA, 98121

    United States

    Phone: 1 732 1359

    Email: jhadlock@isbscience.org


    Background: Social determinants of health (SDoH) such as housing insecurity are known to be intricately linked to patients’ health status. More efficient methods for abstracting structured data on SDoH can help accelerate the inclusion of exposome variables in biomedical research and support health care systems in identifying patients who could benefit from proactive outreach. Large language models (LLMs) developed from Generative Pre-trained Transformers (GPTs) have shown potential for performing complex abstraction tasks on unstructured clinical notes.

    Objective: Here, we assess the performance of GPTs on identifying temporal aspects of housing insecurity and compare results between both original and deidentified notes.

    Methods: We compared the ability of GPT-3.5 and GPT-4 to identify instances of both current and past housing instability, as well as general housing status, from 25,217 notes from 795 pregnant women. Results were compared with manual abstraction, a named entity recognition model, and regular expressions.

    Results: Compared with GPT-3.5 and the named entity recognition model, GPT-4 had the highest performance and had a much higher recall (0.924) than human abstractors (0.702) in identifying patients experiencing current or past housing instability, although precision was lower (0.850) compared with human abstractors (0.971). GPT-4’s precision improved slightly (0.936 original, 0.939 deidentified) on deidentified versions of the same notes, while recall dropped (0.781 original, 0.704 deidentified).

    Conclusions: This work demonstrates that while manual abstraction is likely to yield slightly more accurate results overall, LLMs can provide a scalable, cost-effective solution with the advantage of greater recall. This could support semiautomated abstraction, but given the potential risk for harm, human review would be essential before using results for any patient engagement or care decisions. Furthermore, recall was lower when notes were deidentified prior to LLM abstraction.

    J Med Internet Res 2024;26:e63445

    doi:10.2196/63445

    Keywords

    housing instability; housing insecurity; housing; machine learning; artificial intelligence; AI; large language model; LLM; natural language processing; NLP; electronic health record; EHR; electronic medical record; EMR; social determinants of health; exposome; pregnancy; obstetric; deidentification 


    The overwhelming majority of patients in the United States have their data stored in electronic health records (EHRs). Information regarding a patient’s exposure to social determinants of health (SDoH), such as housing status, employment status, education, and quality of domestic life, provides relevant information that informs patient care and provides valuable avenues for intervention and treatment [1]. It has been estimated that SDoH can affect almost 50% of country-level variation in health outcomes, while clinical care impacts as little as 20% [2]. Housing data, in particular, including a patient’s recent housing status, are known to be intricately linked to their health status [3-5]. Therefore, gaining insight into a patient’s current and past living situation is essential to providing more complete and equitable care. It is also important for research, where capturing longitudinal exposome data is essential for analysis of health outcomes.

    Housing stability is known to exist on a continuum, from complete stability (access to housing of reasonable quality in the absence of threats) to complete instability (no access to housing of reasonable quality) [6]. It is well known that people who are experiencing housing instability are at greater risk for other health issues, including substance use, comorbidities, and mental illness [3,7,8]. People facing housing instability are also at an increased risk for homelessness [5,9,10], which is associated with increased risk of morbidity and mortality [7]. Patients experiencing homelessness are also more likely to end up in the emergency department (ED), have longer hospital stays than low-income housed persons, and are less likely to use preventive services [3,10]. Women who are experiencing housing instability while pregnant face additional challenges, as they usually require consistent access to care throughout their pregnancy. Adverse exposures prior to and during pregnancy can put a child at an increased risk of both short- and long-term health consequences, and it is known that women who experience housing instability during pregnancy are at higher risk of adverse pregnancy outcomes, including preeclampsia, preterm birth, neonatal intensive care unit admission, and maternal morbidity [11-14].

    SDoH information is rarely well documented in structured EHR data [15-17]. This leads to access barriers for health care teams and researchers. In addition, manually identifying SDoH, for example, through chart abstraction, is time-consuming, expensive, and impractical to scale. Because structured data have often been optimized for purposes other than individual care or research, free-text descriptions capture greater breadth and complexity of a patient’s social and behavioral history. Existing projects, such as the Protocol for Responding to & Assessing Patients’ Assets, Risks & Experiences [18], and emerging national interoperability plans [19] are providing paths for clinicians to better capture SDoH data in structured fields, but widespread data standards for data harmonization are still in early development [20].

    Non–large language model (LLM)–based methods for extraction of SDoH information from free-text notes have relied heavily on identification of keywords or phrases, using either manual or semiautomated lexicon curation or rule-based methods [21-24]. However, most of these models are vulnerable to false positives (FP) and can capture only simplified concepts related to SDoH. In addition, previous research to identify housing instability from the EHR has focused primarily on homelessness or simplified housing-related concepts [24-26]. However, because housing instability is heterogeneous with many intersecting dimensions, classifying a patient experiencing housing instability can be more complex than some aspects of social history and exposures, such as smoking. By contrast, LLMs such as OpenAI’s Generative Pre-trained Transformer (GPT) models can handle large quantities of complex, unstructured data using only simple prompts. Research using LLMs on EHR data is still in its early stages, and most work has focused on either fine-tuning models for medical relevance [27-30], comparing model performance to identify the presence or absence of SDoH statements [31,32], or using LLMs for disease diagnosis or phenotyping [33,34]. In addition, it has not been made clear whether the quality of note text flagged as relevant by GPT is similar to that of a human abstractor, and whether or not it is likely to contain hallucinatory text. LLMs such as GPT could also perpetuate health inequity if they perform differently for different patient populations. It is therefore important to test for bias in these models to inform future decisions regarding the use of LLMs in the health care setting. In addition, the possibility of using deidentified clinical notes for abstraction is appealing for further supporting patient privacy. However, deidentification processes involve obfuscation of important details, including dates and locations. This may alter the semantic underpinnings of a given text, making it difficult for an LLM to accurately identify and label SDoH within a given note.

    We examined whether LLMs were able to identify housing instability in clinical free-text notes with greater accuracy than manual abstraction, regular expressions (RegEx), and a pretrained named entity recognition model for SDoH, using EHRs for a population of pregnant women. We also examined the ability to distinguish between present and past instability, the possibility of algorithmic bias in the predictions made by GPT-4 and GPT-3.5, and LLM performance on deidentified versions of patient notes. In addition, we asked both LLMs for a justification of their housing label, helping to provide a better understanding of how these models understand and interpret data. These methods provide new insight relevant to housing security and use of LLMs for SDoH data that is valuable for retrospective research and to help identify patients who might benefit from proactive outreach.


    Ethical Considerations

    This retrospective study protocol was performed in compliance with the Health Insurance Portability and Accountability Act Privacy Rule and was approved by the institutional review board at Providence Health and Services study (2020000783). Consent was waived because disclosure of protected health information (PHI) for the study was determined to involve no more than a minimal risk to the privacy of individuals. All data remained within the secure Providence system. Manual and automated review of notes was conducted solely by the authors, using internal Providence resources.

    Study Setting and Participants

    Providence Health and Services is an integrated US community health care system that provides care in urban and rural settings across 7 states: Alaska, California, Montana, Oregon, New Mexico, Texas, and Washington. Using the Providence EHRs, we identified deliveries from June 8, 2010, through May 29, 2023 (N=595,600). We included singleton deliveries in a cohort of pregnant people aged 18-44 years at the start of pregnancy (n=557,406) as previously described [35]. We limited the deliveries to records that had associated gravida, term, preterm, abortion, and living data information, and to patients who received care in the Providence system during pregnancy. For patients with more than 1 pregnancy episode, we randomly selected a single episode (n=408,158). We limited our patient cohort to those with complete Social Vulnerability Index (n=372,208) information as previously described [35]. To identify patients from our cohort who were experiencing housing instability (“preliminary positive class”), we searched for patients who had either a Systematized Nomenclature of Medicine Clinical Terminology (SNOMED-CT) code for housing instability (Table S1 in Multimedia Appendix 1) or a matching string for the word “homeless” in 1 or more of their free-text notes (n=13,024). Patients who did not meet these criteria were considered in the “preliminary negative class” (n=359,184; Figure 1 [35]).

    ‎
    Figure 1. Cohort selection and experimental design. *Pregnant patients were selected as previously described. †For the positive class, charts were selected that included the SNOMED code for housing insecurity or had 1 or more notes containing the word “homeless.” ‡The most recent note tagged by each method was selected. GPT 3.5T: GPT 3.5 Turbo; JSL: John Snow Labs; LLM: large language model; RegEx: regular expressions.

    Models

    Data processing was accomplished using the Azure AI Services Application Programming Interface for GPT-4 and GPT-3.5 within the secure Providence cloud environment. GPT-4 version 0613 had a 32K token window [36], while GPT-3.5 Turbo (hereafter, GPT-3.5) version 0613 had a 16K token window [37]. Both GPT models were run using LangChain and OpenAI libraries. The John Snow Labs (JSL) named entity recognition model (ner_sdoh_en_4.4.3_3.0_1686654976160) [36] is a state-of-the-art SDoH model designed to detect and label SDoH entities within text data. The housing-specific label includes entities related to the conditions of the patient’s living spaces, for example, homeless, housing, small apartment, and so forth. JSL was run using sparknlp_jsl.version 5.0.0. To determine whether a RegEx search would identify relevant patient notes related to housing, we generated a preliminary list of keywords and phrases related to housing instability. The final list was reviewed by a registered nurse and clinical informaticist with 3 decades of experience (Table S2 in Multimedia Appendix 1).

    Prompt Engineering

    Our prompt was developed using chain-of-thought prompting, where the problem or question description is initially stated and the LLM is asked to identify relevant evidence first and then provide an answer. This method has been shown to be more accurate than asking the LLM to provide only an answer [37]. GPT-4 and GPT-3.5 were asked to first identify chunks of evidence verbatim from the text (evidence). The model was then asked to go through each of the 4 labels: housing noted, housing instability current, housing stability current, and housing instability history and provide an answer for each. The model was then asked to provide a justification explaining why it chose a specific label and the LLMs were explicitly asked not to make up any information. GPT-3.5 was not used in the prompt engineering phase, and the prompt developed for GPT-4 was also used for GPT-3.5. The final prompt can be found in Multimedia Appendix 1.

    Task Definition and Data Labeling

    We defined different instances of housing stability and instability by first carrying out interviews with various subject matter experts, including clinicians, social workers, and resource specialists. All abstraction was conducted by researchers AR, NT, and MN. We generated preliminary abstraction guidelines that were then iteratively refined and finalized with additional input from subject matter experts. The final abstraction guidelines can be found in Multimedia Appendix 1. These guidelines distinguish between stable and unstable housing versus an unknown housing status, with examples for each pulled from EHR notes. An explicit definition of history of housing instability was also created. If a note contained any information on housing it had to be labeled as stably housed, current housing instability, or history of/past housing instability. If the note contained no information on housing, it was labeled as unknown. These guidelines were then used to create the prompt used by the LLM. The abstraction was divided into 2 rounds, with each reviewer abstracting their own set of notes for the first round (original label). For the second round, 25% of the notes underwent dual abstraction. If there was any disagreement for a given note between 2 abstractors, a third abstractor was asked to make a final decision. All the notes were then manually compared with results from GPT-4 to identify any notes that were obviously missed by reviewers. After the second round, a final label was assigned to each note. Performance metrics for manual abstraction were calculated by comparing the original and final labels for each note. After manual abstraction from reviewers, interrater reliability was calculated using Cohen k on the notes that underwent dual abstraction during the second round of abstraction. The average time spent manually abstracting a single note was calculated by each reviewer timing himself or herself for the time spent to abstract 10 notes and then taking the average. The 2 averages for each reviewer were then averaged.

    For prompt engineering, we randomly selected 100 patients from the preliminary positive class and extracted all patient notes the year prior to the patient’s conception date. Of the 100 patients, 51 had at least 1 note within the year prior to conception date, for a total of 1569 notes. Of those 51 patients, we used the JSL model to identify 70 notes from 16 patients related to housing. Each of the 70 notes were then manually abstracted by 2 different abstractors and the answers were compared. Any disagreement between the abstractors was discussed and a final decision was made. From our abstraction guidelines we developed an initial prompt for GPT-4 that contained definitions of housing instability as stated in the abstraction guidelines, as well as examples of housing instability from samples found in the patient notes. This prompt was then run through GPT-4 on the 70 notes and the results were compared with the manually abstracted results. All of the results were then compared, and the prompt was updated again based on results from GPT-4. For example, GPT-4 initially misclassified several cases of past housing instability as current housing instability. We then updated the prompt to specify that “a patient can only experience a ‘history’ of housing instability if they had housing instability in the past, then were stably housed, then experienced housing instability again. If the note refers to past housing instability, for example, ‘the patient was homeless in the past,’ then this can be treated as a ‘history of housing instability.’”

    Patient and Note Selection for Abstraction

    Overall, the 4 methods flagged 25,217 notes from 795 patients from the preliminary positive class as being related to housing or housing instability past or present. If a given method flagged more than 1 note for a specific patient, only the most recent note in relation to the conception date was used for abstraction. For example, if GPT-4 flagged 2 notes dated January 1, 2019, and May 1, 2019, reviewers abstracted only the note dated May 1, 2019. This ensured that a maximum of 4 notes per patient were used for abstraction. Model performance was measured by examining accuracy, recall, precision, and F1-score using the SciKit-Learn library.

    Because it was uncertain how many notes would be flagged by the models for a given number of patients, we initially selected 500 patients from the preliminary positive class. Out of those 500 patients, 295 patients had 1 or more notes within a 12-month period prior to pregnancy, for a total of 9451 notes. A full breakdown of the number of patients and notes used in each round of model tagging and abstraction can be found in Table 1 and Table S3 in Multimedia Appendix 1.

    Table 1. Breakdown of the number of notes used in each round of model tagging and abstraction from each model for the positive and negative classes.
    Class and type of notesNotes tagged by models for housing instability annotation, n


    		JSLaGPT-3.5GPT-4RegExbAny
    Positive class: manual annotation, round 1

    		Most recent notes114737482204

    		Any tagged notes300181284226511
    Positive class: manual annotation, round 2

    		Most recent notes169115119133335

    		Any tagged notes552440546434900
    Negative class

    		Most recent notes2711121459

    		Any tagged notes3312152779

    aJSL: John Snow Labs.

    bRegEx: regular expressions.

    GPT Bias Evaluation

    FPR (false-positive rate) and FNR (false-negative rate) were calculated using FP, FN (false negatives), TP (true positives), and TN (true negatives), derived from the confusion matrix:

    FPR = FP/FP + TN
    FNR = FN/TP + FN

    Moreover, 95% CIs for a population proportion were calculated using the following formula: CI = p̂ ± z× SE, where z=1.96 for a 95% CL. SE was calculated using SE = √(p̂ × (1 – p̂) / n), where p̂ = FPR or FNR, and n = sample size.

    Deidentification of Patient Notes

    Providence has an existing corpus of deidentified notes that was created using a sequence of operations performed on text data to remove PHI [38]. These operations included multiple pretrained machine learning models and RegEx. Two versions of deidentified notes were used: complete de-id, in which all PHI was obfuscated and all dates were shifted or masked if shifting was not possible, and de-id except date, in which all PHI was obfuscated but the dates were not shifted. All notes remained within the secure Providence cloud environment.


    Manual Abstraction of EHR Notes

    From the 25,217 notes from the 795 patients, the 4 automated methods flagged a total of 1411 notes (Table S3 in Multimedia Appendix 1). Given how the models were designed, JSL was able to flag only those notes related to housing in general (housing noted), while RegEx could not distinguish between current and past housing instability. Both GPT-4 and GPT-3.5 were able to flag notes related to general housing status, housing instability current, and housing instability past. After the models were run on the 25,217 notes, we selected notes that were flagged as housing noted for JSL, housing instability current or past for RegEx, and currenthousing instability or pasthousing instability for GPT-4 and GPT-3.5.

    The most commonly flagged note types were assessments, plan of care, obstetric triage, history and physical, consults, discharge summary, and ED notes, indicating that these types of notes are most likely to provide relevant information related to housing or housing instability (Table S4 in Multimedia Appendix 1). We selected the 539 most recent notes from 355 patients for manual abstraction. Demographic characteristics of the 355 patients can be found in Table S5 in Multimedia Appendix 1. Of the 539 manually abstracted notes, the most common type of notes were progress notes (n=216) followed by ED provider notes (n=102) and telephone encounters (n=38; Table S6 in Multimedia Appendix 1).

    Of the 182 patients who were identified as having current or past housing instability, only 18% (33/182) of patients had a structured SNOMED-CT code related to housing instability in their chart (Table 2). Although the percentage of patients with structured codes for housing instability is low, it is higher than what has been previously reported, and it is well known that structured fields do not adequately capture a patient’s housing status [31,39]. These results could reflect that these notes were selected from patients flagged for housing instability, or that health care teams are more likely to ask about and document housing instability when a patient is pregnant. A total of 10% (11/109) of patients who were labeled as stably housed and 11% (7/64) of patients labeled as unknown had a SNOMED-CT code related to housing instability (Table 2). This is likely because researchers analyzed notes only within 1 year of pregnancy, while the SNOMED-CT code could have been added to a patient’s chart at any time before pregnancy.

    Table 2. Number of patients in each housing category who either did or did not have a SNOMED-CTa code related to housing instability in their chartb.
    Housing labelPatients with a SNOMED-CT code for homelessness, n

    		YesNo
    Stably housed1198
    Current housing instability32131
    Past housing instability118
    Unknown757

    aSNOMED-CT: Systematized Nomenclature of Medicine Clinical Terms.

    bSNOMED-CT codes used to identify housing instability can be found in Table S1 in Multimedia Appendix 1.

    Two abstractors manually abstracted the 539 notes and flagged 415 as related to housing. Of those 415, a total of 164 (30.4%) were labeled as stably housed, 223 (41.4%) were labeled as current housing instability, 28 (5%) were labeled as a history of housing instability, and 124 (23%) were labeled unknown (Table 3). Moreover, 25% (103/415) of the notes underwent dual abstraction. Before adjudication, dually abstracted notes had a Cohen k coefficient of 0.589, which reflects moderate agreement [40] and highlights the ambiguity and subjectivity of abstracting complex concepts such as housing instability compared with other SDOH mentions such as employment or parental status, despite the abstractors having been trained on the same guidelines (see Supplementary Methods in Multimedia Appendix 1).

    Table 3. Number and Percentage of manually abstracted notes by housing label.
    Housing labelNotes (n=539), n (%)
    Current housing instability223 (41.4)
    Stably housed164 (30.4)
    Unknown124 (23)
    History of housing instability28 (5)

    Identification of Current and Past Housing Instability

    Because housing status can change over time, it is important to be able to distinguish between current and past housing instability. This may be especially important for certain groups of patients, such as pregnant women, where housing instability during pregnancy may have different implications than prior housing instability. Tables 4 and 5 and Tables S7 and S8 in Multimedia Appendix 1 illustrate the differences in performance metrics between GPT-4, GPT-3.5, RegEx, and manual abstraction in identifying notes related to current or past housing instability, measured against final adjudicated labels. For current and past housing instability, the recall of GPT-4 was higher (0.924) than that of GPT-3.5 (0.717), RegEx (0.649), and manual abstraction (0.702). However, manual abstraction had the highest precision among the 4 methods (0.971), compared with GPT-4 (0.850), GPT-3.5 (0.759), and RegEx (0.632). The low performance of RegEx was due, in part, to acronyms related to housing, such as supportive living services and recreational vehicle, that also serve as equivalent medical shorthand for terms such as single limb support and review. This highlights the shortcomings of using a RegEx-based approach when attempting to identify a complex concept such as housing instability. For identifying current housing instability, GPT-4 still had higher recall than both GPT-3.5 and manual abstraction, but both LLMs had lower precision than manual abstraction. The recall for GPT-3.5 was higher for current housing instability alone, indicating that this model struggled to identify notes where past housing instability was mentioned, further evidenced in Table 6. This demonstrates that while LLMs can identify past or current instances of an event such as housing instability, specific models should be tested for their performance in each category individually.

    Table 4. Comparison of recall and precision for RegExa, GPT-3.5, GPT-4, and manual abstraction in identifying notes with current or past housing instability, measured on 539 manually abstracted notes.
    ModelRecallPrecision
    RegEx0.6490.632
    GPT-3.50.7170.759
    GPT-40.9240.85
    Manual annotation0.7020.971

    aRegEx: regular expressions.

    Table 5. Comparison of recall and precision for GPT-3.5, GPT-4, and manual abstraction in identifying notes with current housing instability, measured on 539 manually abstracted notes.
    ModelRecallPrecision
    GPT-3.50.7530.734
    GPT-40.9060.831
    Manual annotation0.7180.933
    Table 6. Recall and precision metrics for GPT-4 and GPT-3.5 for each housing label measured on 539 manually abstracted notes
    Housing labelGPT-4GPT-3.5

    		RecallPrecisionRecallPrecision
    Stable housing0.390.880.340.67
    Current housing instability0.910.830.750.73
    Past housing instability0.680.630.180.62
    Unknown0.820.530.680.38

    The drop in recall for manual abstraction highlights the high level of ambiguity associated with abstracting this type of information, as well as the energy needed for humans to conduct manual abstraction for extended periods of time. Examples of relevant notes, along with the housing labels and justifications from GPT-4 and GPT-3.5, can be found in Table 7. There were several ambiguous cases related to housing in patient notes. For example, “Was in domestic violence. Daughters are [redacted ages] and living with parents for now.” In this example, it is not explicitly stated where the patient is staying, making her current housing status unknown. However, this was flagged as current instability by GPT-4 and stably housed by GPT-3.5. In another example, a patient was living with her family and their landlord was a meth dealer. However, the patient’s boyfriend was actively trying to find them a new living situation, which made the patient happy. While both abstractors agreed that the housing situation was not stable, this type of instability was not explicitly mentioned in the note prompt for the LLMs. The reviewers labeled the note as current instability, but the note was tagged as stably housed by GPT-4 and current instability by GPT-3.5.

    Table 7. Examples of relevant note text and manual abstraction labels compared with GPT-4 and GPT-3.5 and relevant justificationa.
    Manual abstractionGPT-4 labelGPT-3.5 labelText selected from longer note by human abstractorJustification GPT-4Justification GPT-3.5
    Current instabilityCurrent instabilityCurrent instabilityShe has been under a great deal of stress over the past year and recently became homeless. She and her boyfriend have moved to this area where she has family where she is now staying.The patient note mentions that the patient recently became homeless, indicating current housing instability. The patient is now staying with family, but this is likely a temporary arrangement given the recent homelessness and the reference to a halfway house, which is a form of transitional housing. There is no evidence in the note to suggest a history of housing instability prior to the current situation.The evidence shows that the patient recently became homeless and is currently facing housing instability. The patient’s history of frequent kidney stones and recent surgeries for kidney stones further suggests a lack of stable housing.
    Current instabilityPast instabilityCurrent instabilityShe was having many stressors at that time which included being kicked out of her mother’s house and staying with ex BF.The patient note mentions that the patient was kicked out of her mother’s house and was staying with her ex-boyfriend. This indicates a past instance of housing instability. However, the note does not provide current information about the patient’s housing situation, so it is unknown whether the patient is currently stably housed or not.The patient note mentions that the patient was kicked out of her mother’s house and was staying with her ex-boyfriend, which indicates current housing instability. In addition, the patient’s active problem list includes depression, which can be associated with housing instability. However, there is no explicit evidence or mention of past housing instability, so it is unknown whether the patient has a history of housing instability.
    Current instabilityCurrent instabilityNo housing noted, current instability unknownShe said she is staying at a motel.The patient note mentions that the patient is currently staying at a motel, which indicates a form of housing instability. There is no evidence in the note to suggest that the patient has faced housing instability in the past.There is no explicit evidence or mention of the patient’s housing status in the note. The patient mentions staying at a motel, but it is unclear whether this is a temporary situation or a stable housing arrangement. Therefore, the current housing stability is unknown.
    Current instabilityStably housedCurrent instabilityPt reports she is living w/[redacted]; they are renting a house from [redacted]; he is a meth dealer. [Bf] has made arrangements to get them into a new living situation. She is very happy about this.The patient is currently living with [redacted], in a house they are renting. There is no evidence of housing instability at the moment. The boyfriend has made arrangements for a new living situation, which the patient is happy about. There is no mention of past housing instability.The evidence from the patient note indicates that [the patient] is currently facing housing instability. She is living in a rented house with [redacted], but the house is owned by [redacted] who is a meth dealer. In addition, the note mentions that [the patient] is in the process of moving to [redacted] and has been applying for jobs. This indicates that her current housing situation is unstable. There is no evidence of past housing instability mentioned in the note.
    UnknownCurrent instabilityStably housedStates lots of changes lately. Was in domestic violence. Daughters are [redacted] and living with parents for now.The patient note mentions that the patient’s daughters are currently living with her parents, which suggests that the patient is temporarily staying with family members. This is an indication of current housing instability. There is no evidence in the note to suggest a history of housing instability.The patient’s housing status is noted in the evidence. The patient is currently stably housed as she is living with her parents. There is no evidence of current or past housing instability.

    aCertain sections of note text and evidence from GPT-4 and GPT-3.5 have been removed to protect patient privacy.

    A minority of notes contained specific references to housing, such as “patient lives in the woods of [redacted location] with her boyfriend.” This phrase signified likely homelessness to most abstractors, but one abstractor assumed that this meant a cabin in the woods. This case was caught by GPT-4 but missed by RegEx and GPT-3.5. There were additional cases where housing instability was explicitly mentioned and was missed by 1 or more reviewers. This most often occurred in longer notes that contained a significant amount of information, and only 1-2 sentences related to housing, for example “section 8 voucher,” which refers to a US program for assisting very low-income families. This sentence was missed by GPT-3.5 and RegEx but was correctly identified by GPT-4. In addition, there were several notes that contained no information (“blank” notes). In several cases, GPT-3.5 used sentences from the prompt text as evidence and justification and flagged the note as current or past instability. This did not occur with GPT-4. This is likely because the LLM was asked to provide text evidence verbatim, and GPT-3.5 used the prompt because no relevant note text was available. However, the researchers found no instances of hallucinated evidence in any of the GPT-4 responses that were reviewed, suggesting that requiring verbatim evidence from LLMs can be a solution to hallucinated responses.

    Although GPT-3.5 struggled to identify several cases of housing instability, the researchers could not identify a consistent trend in the type or content of the FP or FN notes. However, there were several cases where GPT-3.5 listed known risk factors mentioned elsewhere in the note as evidence of housing instability. For example, it noted a patient’s frequent kidney surgeries or a depression diagnosis as justification for housing instability, although there was clear mention of housing instability elsewhere in the note (Table 7). This was not the case with GPT-4, which used only direct evidence from the note text that mentioned housing-related terms as justification. Although kidney disease and depression are associated with housing instability [41-44], a human abstractor would not use this as a justification for housing instability. Having an LLM use this as justification could be a potential concern but could also be an opportunity for a different use case, where researchers ask LLMs to identify potential risk factors observed in a set of records. Those results might show bias in LLMs or highlight patterns that humans overlook. These examples also demonstrate significant differences between the 2 GPT releases and highlight the value of providing evidence and justification for every note when using LLMs. In most cases, the evidence gathered by the reviewers was similar or identical but the interpretation differed. Similarly, the evidence gathered by GPT-4 was similar to the reviewers in most cases (although this was not always the case with GPT-3.5). This indicates that, while manual abstraction is likely to yield more accurate results, GPT-4 could be used to rapidly gather relevant note text for computer-assisted manual review, helping save time without losing important or relevant patient information.

    Identification of General Housing Status

    Table 8 and Table S9 in Multimedia Appendix 1 show the performance of GPT-3.5 and GPT-4 compared with JSL and manual abstraction in identifying notes where housing was mentioned. GPT-4 outperformed both GPT-3.5 and JSL across all 4 metrics but had a slightly worse precision compared with manual abstraction (0.936 compared with 0.952), although recall was higher (0.781 compared with 0.720). Interestingly, the majority of cases that were missed by GPT-4 were instances where housing was stable, for example, “she lives at home with her children” or “patient was requesting to go home.” A possible underlying cause was that the prompt had been heavily focused on identifying cases of housing instability, with little guidance provided on identifying housing status overall. Prompt engineering focused on different proportions of relevant information might yield different and more accurate results.

    Table 8. Comparison of recall and precision for JSLa, GPT-3.5, GPT-4, and manual abstraction in identifying notes where housing was noted, measured on 539 manually abstracted notes.
    ModelRecallPrecision
    JSL0.7330.864
    GPT-3.50.6750.875
    GPT-40.7810.936
    Manual annotation0.720.952

    aJSL: John Snow Labs.

    LLM Performance by Housing Category

    Table 6 and Tables S10 and S11 in Multimedia Appendix 1 show the differences in performance for GPT-4 and GPT-3.5 in identifying notes across the different housing categories: stable housing, current housing instability, past housing instability, or unknown. GPT-3.5 performed worse than GPT-4 across all categories and had particularly low recall for notes labeled as past instability compared with GPT-4, which had a higher recall than precision in this category. Both GPT-4 and GPT-3.5 demonstrated poor recall for stable housing notes. A possible underlying cause was that the prompt focused more heavily on housing instability and did not provide training focused on stable housing. These data indicate a performance improvement for the GPT-4 release and demonstrate the effects of prompt engineering on the model outcome.

    LLM Bias Evaluation

    To test for bias in the LLMs, we used the fairness criteria of separation [45] to examine the FPR and FNR between GPT-4 and GPT-3.5 across 3 different housing labels (housing noted, housing instability past or current, and housing instability current) across the age (18-30 years and 31-44 years), race (American Indian and Alaska Native, Asian, Black, Native Hawaiian and other Pacific Islander, Unknown or Declined, White, and Other), and ethnic (Hispanic or Latino, not Hispanic or Latino, and Unknown or Declined) demographic groupings. We then examined the 95% CIs of the FPR and FNR for each group (Tables S12-S17 in Multimedia Appendix 1 and Figures S1-S3 in Multimedia Appendix 1). According to the fairness criteria of separation, any difference in FPR and FNR between groups suggests potential algorithmic bias. We did observe differences in FPR and FNR between all of the groups within the 3 demographic categories. However, when we examined the overlap of the 95% CIs between groups, we found that in all cases, except in cases where the sample sizes were extremely small (n<5), there was overlap between CIs for all the groups, suggesting that the differences between groups are not significant. However, further work with a larger sample population is needed.

    Cost Breakdown of LLMs Compared with Manual Abstraction

    There was a substantial difference in cost between GPT-4 and GPT-3.5, as shown in Tables 9 and 10, due to the increase in cost per 1000 tokens for notes for GPT-4 (US $0.06) compared with GPT-3.5 (US $0.003). The output cost also increased from US $0.004 per 1000 tokens for GPT-3.5 to US $0.12 per 1000 tokens for GPT-4. Interestingly, the prompt (US $1778.39 for GPT-4) cost more than the total for all the notes (US $701.72 for GPT-4), and this was the case for GPT-3.5 as well. This is because the prompt had to be included as part of each note. Because the prompt was long (1182 tokens), this increased the cost substantially. Future work comparing model performance in relation to prompt length would provide valuable insight into this trade-off.

    The cost of manual abstraction varies by location, but in the United States, it can be estimated to be the minimum wage per hour for that state. As of January 2024, in the state of Washington, the minimum wage was US $16.28 per hour [46]. To analyze 25,217 notes would have cost approximately US $9442, substantially higher than either LLM.

    Table 9. Cost analysis of GPT-4 compared with GPT-3.5 for 25,217 notes from 795 patientsa.
    ModelCosts (US $)

    		PromptNotesOutputTotal
    GPT-41788.39701.72181.352671.46
    GPT-3.589.4235.098.89133.40

    aGPT-4 prompt: 1182 tokens × 25217 notes × 0.06 per 1000 tokens. GPT-4 notes: 11,695,270 tokens × 0.06 per 1000 tokens. GPT-4 output: 15,11,244 tokens × 0.12 per 1000 tokens. GPT-3.5 prompt: 1182 tokens × 25,217 notes × 0.003 per 1000 tokens. GPT-3.5 notes: 11,695,270 tokens × 0.003 per 1000 tokens. GPT-3.5 output: 22,23,674 tokens × 0.004 per 1000 tokens.

    Table 10. Estimated cost of running a single note on GPT-4 and GPT-3.5 compared with manual abstraction.
    ModelCost (US $)
    Manual annotation0.374
    GPT-40.106
    GPT-3.50.005

    Analysis on Patients in the Preliminary Negative Class

    To evaluate the performance of our 4 methods compared with patients from the preliminary positive class, we selected a random sample of 5455 notes from 348 patients in the preliminary negative class. GPT-4, GPT-3.5, RegEx, and JSL were run on all 5455 notes using the same method as notes from patients in the preliminary positive class. Of the 5455 notes, all 4 methods flagged only 59 of the most recent notes from 50 patients with 1 of the 4 housing labels (GPT-4: n=12, GPT-3.5: n=11, RegEx: n=14, and JSL: n=27). A randomly selected sample of 20 notes from the preliminary negative class that were not flagged by any of the 4 methods was reviewed by a manual abstractor (AR). Of the 20 notes, 19 notes did not contain any information on housing and 1 note suggested stable housing. An additional randomly selected sample of 20 notes was taken from the 59 notes that were flagged by the 4 methods and reviewed by a manual abstractor (AR). Of the 20 notes, 8 were current housing instability, 1 was past housing instability, 5 were stably housed, and 6 were unknown.

    This demonstrates that, as expected, patients from the preliminary negative class had far fewer notes related to housing or housing instability than patients in the preliminary positive class. To generate a manually abstracted data set of ~500 notes to compare with the preliminary positive class, the 4 methods would need to analyze approximately 46,000 notes. However, JSL is not able to distinguish between notes related to general housing and housing instability, and RegEx has low precision and recall for identifying notes related to housing instability. Therefore, LLMs provide the best chance of finding relevant notes related to housing or housing instability for patients in either class. Because the number of notes flagged by GPT-4 and GPT-3.5 were very low in the preliminary negative class (12 and 11, respectively, out of 5455 notes), the 2 models would need to analyze approximately 227,300 notes to find ~500 notes related to housing or housing instability in the preliminary negative class. Due to the cost restrictions of running these LLMs, we were unable to perform this analysis. However, future work to analyze additional notes from patients in the preliminary negative class could provide insight into any differences in notes between these 2 classes.

    Evaluation of GPT-4 on Deidentified Patient Notes

    Deidentification can help mitigate privacy risks to individuals to support secondary use of data for research. In the United States, the Health Insurance Portability and Accountability Act specifies 18 categories of information that are PHI that must be removed from medical records [47,48]. However, while the process of deidentification is necessary to protect patient privacy, the information that is removed during this process, such as dates and locations, may result in the loss of important contextual clues needed for LLM analysis of housing instability. We wanted to examine whether LLMs performed similarly on two versions of deidentified patient notes compared with original notes: (1) fully deidentified notes where all PHI was obfuscated and all dates were shifted or masked if shifting was not possible (hereafter referred to as complete de-id), and (2) patient notes where all PHI was obfuscated but the dates were not shifted (hereafter referred to as de-id except date). All notes remained within the secure Providence system. All data processing was conducted in Providence’s secure cloud environment.

    We ran the 2 deidentified versions of the 539 manually abstracted notes through GPT-4 and compared the performance metrics with the original notes to identify current or past housing instability or general housing status (housing noted). We found that in all cases, recall dropped but precision increased for the deidentified notes compared with the original notes (Multimedia Appendix 2 and Table S18 in Multimedia Appendix 1). For example, for the notes labeled as current housing instability, the recall for GPT-4 on the original notes was 0.906, but this dropped to 0.812 and 0.834 for the complete de-id and de-id except date notes, respectively. By contrast, the precision increased from 0.831 for the original notes compared with 0.862 and 0.849 for the complete de-id and de-id except date notes, respectively. These minor increases in precision are likely because results from GPT-4 are slightly different each time the model is run, resulting in slight differences in performance. The drop in recall is not surprising given the nature of deidentification, in which both places and locations have been obfuscated, making it more difficult for the model to identify relevant notes. For example, there were several cases in the original notes where the patient was stated to be living in a specific location, such as a city or county, but these locations were changed to medical facilities or, in one instance, a jail, resulting in the model sometimes mislabeling the patient as unstably housed or missing the note as related to housing altogether. In other cases, when the dates were shifted, instances of past housing instability were made current, making it difficult for the LLM to properly identify and label these notes.


    Principal Findings

    Our results demonstrate the potential benefit of using LLMs to identify instances of complex SDoH concepts in the EHR, such as past or current housing instability. Although manual abstraction correctly classified more notes related to housing instability, it is more expensive than using LLMs, with a limited increase in performance. GPT-4 outperformed GPT-3.5, JSL, and RegEx in identifying patients experiencing current or past housing instability. In most cases, the evidence from GPT-4 was similar or identical to that of the manual abstractors, and no hallucinations were observed in GPT-4 output. Our work also suggests that requiring the LLM to provide verbatim evidence and justification from the original text can help reduce the risk that relevant context about housing information is omitted from LLM results. We also found that LLMs perform well in identifying notes related to housing insecurity without the need for fine-tuning, using only few-shot learning. However, given the sensitive nature of housing insecurity and complexity of individual situations, final human review would be essential for any decision regarding patient communication or patient care. Another important finding is that recall was lower on notes after they had been deidentified. Further research is needed to determine whether human annotators would also have lower performance on deidentified notes.

    It is important to note that housing instability does not exist in a vacuum; oftentimes, there are multiple compounding factors that can either contribute to or be influenced by housing instability, including domestic violence, drug abuse, and mental illness. One limitation of this study was our focus solely on housing instability and not an additional identification of these risk factors. This resulted in some cases where a patient was technically considered to have stable housing, but there were other risk factors in the patient note that would likely be important for users of abstraction results: case workers, clinicians, or researchers. Expanding the prompt might improve performance and enable labeling that separates out multiple dimensions of housing security, including uncertainty about future housing, frequency of housing transitions, and risks from unsafe housing situations. Because GPT-4 and GPT-3.5 are not deterministic models, responses, and therefore performance, may also change if rerun on the same notes. However, the newest release of GPT, GPT-4 Turbo, allows researchers to add a deterministic seed parameter to ensure that the model returns the same response every time, helping to prevent changes in performance across multiple runs.

    Because all our methods required that each note be analyzed individually, all 4 methods identified 1411 notes out of 25,217 related to housing or housing instability in the preliminary positive class, and only 64 out of 5455 notes related to housing or housing instability in the preliminary negative class, we can conclude that many notes in this study likely did not contain information on housing or housing instability. However, future work could investigate the similarities and differences in note content related to housing or housing instability between patients in both classes. In addition, because we used automated methods for the initial selection of relevant patient notes, we likely missed some patient notes related to housing or housing instability that were not captured with any of the 4 automated methods. Future work to manually abstract a larger corpus of patient notes related to housing and housing instability, as well as other SDoH categories, would be valuable for full validation. Furthermore, repeat validation over time would be important to reduce the risk of drift, and new tests would be appropriate before deploying the system in different geographic settings. . Another limitation is that the study was limited to the content documented in EHR notes, and a recent survey reported that only about 60% of patients felt comfortable sharing SDoH-related information [49]. Future studies would benefit from longitudinal confidential surveys or interviews with patients and health care teams. In addition, because the time and cost to run GPT-4 and GPT-3.5 might not be feasible across millions of patient notes, work with newly emerging open source language models may provide a similar performance for a much lower cost and run time.

    Conclusions

    Methods for abstracting SDoH data can be valuable for retrospective research, chart review for prospective trials, and population health interventions to identify those who might benefit from proactive outreach. This work demonstrates that LLMs have potential for computer-assisted abstraction of social history, improving recall and reducing costs. This includes temporal feature engineering, such as identifying whether patient experienced housing instability before or during their pregnancy. At the same time, precision was slightly lower than that of human abstractors, and given the potential risk for harm to patients, human review would be essential for any decisions on patient communication or patient care. Results also identified 2 important areas where further work is needed: separating out 3 different dimensions of housing insecurity and advancing deidentification methods that do not result in loss of information on social history.

    Acknowledgments

    Support was provided in part by the National Center for Advancing Translational Sciences (NCATS), National Institutes of Health, through the Biomedical Data Translator program (award 3OT2TR00344301S1). Any opinions expressed in this document are those of the Translator community at large and do not necessarily reflect the views of NCATS, individual Translator team members, or affiliated organizations and institutions.

    Data Availability

    Results have been aggregated and reported within this paper to the extent possible while maintaining privacy from personal health information as required by US law. All data are archived within Providence systems in a Health Insurance Portability and Accountability Act–secure audited compute environment, and those wishing to verify study conclusions can contact the chief data officer. All biomedical codes used to extract data from electronic health records have been shared. The source code used to run models and reproduce results shown for this study is available at GitHub [52].

    Authors' Contributions

    AR, NT, MN, and SNP had full access to all data in the study and take responsibility for the integrity of the data and accuracy of the data analysis. AR, NT, MN, LM, and JH led the concept and design. AR, NT, MN, and SNP contributed to acquisition, analysis, and interpretation of data. AR, NT, MN, and SNP drafted of the manuscript. All authors contributed to critical revision of the manuscript. AR and NT led the statistical analysis. JH and LM did the supervision.

    Conflicts of Interest

    None declared.

    Multimedia Appendix 1

    Supplementary methods, tables, and figures.

    DOCX File , 314 KB
    Multimedia Appendix 2

    Comparison of GPT-4 performance in identifying (A) current or past housing instability, (B) current housing instability, or (C) general housing status from 3 different versions of the same set of 539 notes. Note versions were original patient notes, completely deidentified notes (complete de-id), or deidentified notes dates unchanged (de-id except date).

    PNG File , 73 KB
    1. Social determinants of health. World Health Organization. URL: https://www.who.int/teams/social-determinants-of-health [accessed 2023-11-29]
    2. Hood CM, Gennuso KP, Swain GR, Catlin BB. County health rankings: relationships between determinant factors and health outcomes. Am J Prev Med. 2016;50(2):129-135. [CrossRef] [Medline]
    3. Liu M, Hwang SW. Health care for homeless people. Nat Rev Dis Primers. 2021;7(1):5. [CrossRef] [Medline]
    4. Kushel MB, Vittinghoff E, Haas JS. Factors associated with the health care utilization of homeless persons. JAMA. 2001;285(2):200-206. [CrossRef] [Medline]
    5. Easton C, Javadizadeh E, Plett P, Perez SN, Isard R, Hariramani S, et al. A scoping review examining the association of housing quality and psychosocial well-being following homelessness: an ecological systems perspective. J Soc Distress Homeless. 2023:1-25. [CrossRef]
    6. Frederick TJ, Chwalek M, Hughes JR, Karabanow J, Kidd S. How stable is stable? defining and measuring housing stability. J Community Psychol. 2014;42(8):964-979. [CrossRef]
    7. Fazel S, Geddes JR, Kushel M. The health of homeless people in high-income countries: descriptive epidemiology, health consequences, and clinical and policy recommendations. Lancet. 2014;384(9953):1529-1540. [FREE Full text] [CrossRef] [Medline]
    8. Fazel S, Khosla V, Doll H, Geddes J. The prevalence of mental disorders among the homeless in western countries: systematic review and meta-regression analysis. PLoS Med. 2008;5(12):e225. [FREE Full text] [CrossRef] [Medline]
    9. Pruitt AS, Barile JP. Actionable research for understanding and addressing homelessness. J Community Psychol. 2022;50(5):2051-2057. [CrossRef] [Medline]
    10. Kushel MB, Gupta R, Gee L, Haas JS. Housing instability and food insecurity as barriers to health care among low-income Americans. J Gen Intern Med. 2006;21(1):71-77. [FREE Full text] [CrossRef] [Medline]
    11. Huang K, Waken RJ, Luke AA, Carter EB, Lindley KJ, Joynt Maddox KE. Risk of delivery complications among pregnant people experiencing housing insecurity. Am J Obstet Gynecol MFM. 2023;5(2):100819. [CrossRef] [Medline]
    12. DiTosto JD, Holder K, Soyemi E, Beestrum M, Yee LM. Housing instability and adverse perinatal outcomes: a systematic review. Am J Obstet Gynecol MFM. 2021;3(6):100477. [FREE Full text] [CrossRef] [Medline]
    13. Leifheit KM, Schwartz GL, Pollack CE, Edin KJ, Black MM, Jennings JM, et al. Severe housing insecurity during pregnancy: association with adverse birth and infant outcomes. Int J Environ Res Public Health. 2020;17(22):8659. [FREE Full text] [CrossRef] [Medline]
    14. Green JM, Fabricant SP, Duval CJ, Panchal VR, Cahoon SS, Mandelbaum RS, et al. Trends, characteristics, and maternal morbidity associated with unhoused status in pregnancy. JAMA Netw Open. 2023;6(7):e2326352. [FREE Full text] [CrossRef] [Medline]
    15. Wang M, Pantell M, Gottlieb L, Adler-Milstein J. Documentation and review of social determinants of health data in the EHR: measures and associated insights. J Am Med Inform Assoc. 2021;28(12):2608-2616. [FREE Full text] [CrossRef] [Medline]
    16. Palacio AM, Suarez M, del Toro Y, Seo D, Garay D, Pareja D, et al. Integrating social determinants of health into the electronic health records of a large health system: a qualitative perspective. Perspectives in Health Information Management. 2018;(Summer 2018):1-19. [FREE Full text]
    17. Cantor MN, Thorpe L. Integrating data on social determinants of health into electronic health records. Health Aff (Millwood). 2018;37(4):585-590. [FREE Full text] [CrossRef] [Medline]
    18. Protocol for Responding to and Assessing Patients' Assets, Risks, and Experiences. PRAPARE. 2021. URL: https://prapare.org/ [accessed 2023-12-08]
    19. Social determinants of health. Office of Disease Prevention and Health Promotion. URL: https://www.healthypeople.gov/2020/topics-objectives/topic/social-determinants-of-health [accessed 2020-01-13]
    20. The Office of the National Coordinator for Health Information Technology (ONC). Social determinants of health. HealthIT.gov. URL: https://www.healthit.gov/health-equity/social-determinants-health [accessed 2023-12-07]
    21. Li I, Pan J, Goldwasser J, Verma N, Wong WP, Nuzumlalı MY, et al. Neural natural language processing for unstructured data in electronic health records: a review. Comput Sci Rev. 2022;46:100511. [CrossRef]
    22. Patra BG, Sharma MM, Vekaria V, Adekkanattu P, Patterson OV, Glicksberg B, et al. Extracting social determinants of health from electronic health records using natural language processing: a systematic review. J Am Med Inform Assoc. 2021;28(12):2716-2727. [FREE Full text] [CrossRef] [Medline]
    23. Han S, Zhang RF, Shi L, Richie R, Liu H, Tseng A, et al. Classifying social determinants of health from unstructured electronic health records using deep learning-based natural language processing. J Biomed Inform. 2022;127:103984. [FREE Full text] [CrossRef] [Medline]
    24. Lybarger K, Dobbins NJ, Long R, Singh A, Wedgeworth P, Uzuner Ö, et al. Leveraging natural language processing to augment structured social determinants of health data in the electronic health record. J Am Med Inform Assoc. 2023;30(8):1389-1397. [FREE Full text] [CrossRef] [Medline]
    25. Bejan CA, Angiolillo J, Conway D, Nash R, Shirey-Rice JK, Lipworth L, et al. Mining 100 million notes to find homelessness and adverse childhood experiences: 2 case studies of rare and severe social determinants of health in electronic health records. J Am Med Inform Assoc. 2018;25(1):61-71. [FREE Full text] [CrossRef] [Medline]
    26. Reeves RM, Christensen L, Brown JR, Conway M, Levis M, Gobbel GT, et al. Adaptation of an NLP system to a new healthcare environment to identify social determinants of health. J Biomed Inform. 2021;120:103851. [FREE Full text] [CrossRef] [Medline]
    27. Yang X, Chen A, PourNejatian N, Shin HC, Smith KE, Parisien C, et al. A large language model for electronic health records. NPJ Digit Med. 2022;5(1):194. [FREE Full text] [CrossRef] [Medline]
    28. Singhal K, Azizi S, Tu T, Mahdavi SS, Wei J, Chung HW, et al. Large language models encode clinical knowledge. Nature. 2023;620(7972):172-180. [FREE Full text] [CrossRef] [Medline]
    29. Shah-Mohammadi F, Finkelstein J. AI-powered social determinants of health extraction from patient records: a GPT-based investigation. 2024. Presented at: 2024 IEEE First International Conference on Artificial Intelligence for Medicine, Health and Care (AIMHC); February 5-7, 2024; Laguna Hills, CA. [CrossRef]
    30. Fu Y, Ramachandran GK, Dobbins NJ, Park N, Leu M, Rosenberg AR, et al. Extracting social determinants of health from pediatric patient notes using large language models: novel corpus and methods. 2024. Presented at: Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024); May 20-25, 2024; Torino, Italia. URL: https://aclanthology.org/2024.lrec-main.618
    31. Guevara M, Chen S, Thomas S, Chaunzwa TL, Franco I, Kann BH, et al. Large language models to identify social determinants of health in electronic health records. NPJ Digit Med. 2024;7(1):6. [FREE Full text] [CrossRef] [Medline]
    32. Ong JCL, Seng BJJ, Law JZF, Low LL, Kwa ALH, Giacomini KM, et al. Artificial intelligence, ChatGPT, and other large language models for social determinants of health: current state and future directions. Cell Rep Med. 2024;5(1):101356. [FREE Full text] [CrossRef] [Medline]
    33. Ahsan H, McInerney D, Kim J, Potter C, Young G, Amir S, et al. Retrieving evidence from EHRs with LLMs: possibilities and challenges. Proc Mach Learn Res. 2024;248:489-505. [Medline]
    34. Alsentzer E, Rasmussen MJ, Fontoura R, Cull AL, Beaulieu-Jones B, Gray KJ, et al. Zero-shot interpretable phenotyping of postpartum hemorrhage using large language models. NPJ Digit Med. 2023;6(1):212. [FREE Full text] [CrossRef] [Medline]
    35. Piekos SN, Hwang YM, Roper RT, Sorensen T, Price ND, Hood L, et al. Effect of COVID-19 vaccination and booster on maternal-fetal outcomes: a retrospective cohort study. Lancet Digit Health. 2023;5(9):e594-e606. [FREE Full text] [CrossRef] [Medline]
    36. GPT 4. OpenAI. URL: https://platform.openai.com/docs/models#gpt-4-turbo-and-gpt-4 [accessed 2024-11-08]
    37. GPT 3.5 Turbo. OpenAI. URL: https://platform.openai.com/docs/models#gpt-3-5-turbo [accessed 2024-11-08]
    38. Social determinants of health (SDOH) named entity recognition (NER) model. John Snow Labs. 2020. URL: https://nlp.johnsnowlabs.com/2023/06/13/ner_sdoh_en.html [accessed 2024-06-19]
    39. Kojima T, Gu S, Reid M, Matsuo Y, Iwasawa Y. Large language models are zero-shot reasoners. 2022. Presented at: 36th Conference on Neural Information Processing Systems (NeurIPS 2022); November 28 to December 9, 2022:22199-22213; New Orleans, LA. URL: https:/​/proceedings.​neurips.cc/​paper_files/​paper/​2022/​file/​8bb0d291acd4acf06ef112099c16f326-Paper-Conference.​pdf
    40. Kocaman V, Talby D. Accurate clinical and biomedical named entity recognition at scale. Softw Impacts. 2022;13:100373. [CrossRef]
    41. Hatef E, Ma X, Rouhizadeh M, Singh G, Weiner JP, Kharrazi H. Assessing the impact of social needs and social determinants of health on health care utilization: using patient- and community-level data. Popul Health Manag. 2021;24(2):222-230. [FREE Full text] [CrossRef] [Medline]
    42. McHugh ML. Interrater reliability: the kappa statistic. Biochem Med (Zagreb). 2012;22(3):276-282. [FREE Full text] [Medline]
    43. Novick TK, Baweja M, American Society of Nephrology Healthcare Justice Committee AdvocacyScholarship Work Group, American Society of Nephrology Healthcare Justice Committee AdvocacyScholarship Work Group are. Housing: a critical contributor to kidney disease disparities. J Am Soc Nephrol. 2022;33(8):1471-1473. [FREE Full text] [CrossRef] [Medline]
    44. Novick TK, Kushel M, Crews D. Unstable housing and kidney disease: a primer. Kidney Med. 2022;4(4):100443. [FREE Full text] [CrossRef] [Medline]
    45. Shelton KH, Taylor PJ, Bonner A, van den Bree M. Risk factors for homelessness: evidence from a population-based study. Psychiatr Serv. 2009;60(4):465-472. [CrossRef] [Medline]
    46. Padgett DK. Homelessness, housing instability and mental health: making the connections. BJPsych Bull. 2020;44(5):197-201. [FREE Full text] [CrossRef] [Medline]
    47. Privacy tech-know blog: when worlds collide - the possibilities and limits of algorithmic fairness (part 1). Office of the Privacy Commissioner of Canada. 2023. URL: https://www.priv.gc.ca/en/blog/20230405_01/ [accessed 2024-03-26]
    48. Minimum wage. Washington State Department of Labor & Industries. URL: https://lni.wa.gov/workers-rights/wages/minimum-wage/ [accessed 2024-03-06]
    49. Neamatullah I, Douglass MM, Lehman LH, Reisner A, Villarroel M, Long WJ, et al. Automated de-identification of free-text medical records. BMC Med Inform Decis Mak. 2008;8:32. [FREE Full text] [CrossRef] [Medline]
    50. Office for Civil Rights‚ HHS. Standards for privacy of individually identifiable health information. Final rule. Fed Regist. 2002;67(157):53181-53273. [FREE Full text] [Medline]
    51. Strawley C, Richwine C. Patient preferences with social needs information sharing. HealthIT.gov. URL: https:/​/www.​healthit.gov/​buzz-blog/​health-information-exchange-2/​patient-preferences-with-social-needs-information-sharing [accessed 2023-11-24]
    52. Hadlock-Lab/LLM_SDoH. GitHub. URL: https://github.com/Hadlock-Lab/LLM_SDoH [accessed 2024-11-07]

    ‎
    ED: emergency department
    EHR: electronic health record
    FN: false positive
    FNR: false-negative rate
    FP: false negative
    FPR: false-positive rate
    GPT: Generative Pre-trained Transformer
    JSL: John Snow Labs
    LLM: large language model
    PHI: protected health information
    RegEx: regular expressions
    SDoH: social determinants of health
    SNOMED-CT: Systematized Nomenclature of Medicine Clinical Terminology
    TN: true negatives
    TP: true positives

    Edited by G Eysenbach, X Ma; submitted 20.06.24; peer-reviewed by NJ Dobbins; comments to author 19.07.24; revised version received 31.08.24; accepted 25.09.24; published 19.11.24.

    Copyright

    ©Alexandra Ralevski, Nadaa Taiyab, Michael Nossal, Lindsay Mico, Samantha Piekos, Jennifer Hadlock. Originally published in the Journal of Medical Internet Research (https://www.jmir.org), 19.11.2024.

    This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in the Journal of Medical Internet Research (ISSN 1438-8871), is properly cited. The complete bibliographic information, a link to the original publication on https://www.jmir.org/, as well as this copyright and license information must be included.