LGLillian Gelberg

Lillian Gelberg

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article

Previous Organizations / Institutes

The organizations / institutes from the scholar's other articles in the past ten years


Lillian Gelberg has a total of 139 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Compensatory Integration Work: How people experiencing homelessness hold fragmented care systems together

    by Olivia S Jung, Juan R Delgado, Clarissa Alvarez, Michael Hsu, Kyle Mueller, Chris Rubeo, Beth A Glenn, Lillian Gelberg

    Introduction: People experiencing homelessness face interlocking needs, requiring services spanning health care, housing, employment, and food assistance that are separately funded, separately governed, and rarely organized around the person. Methods: To understand how people navigate such fragmented care ecosystems, we conducted semi-structured interviews with 26 adults experiencing homelessness at a low-barrier community meal program in Los Angeles. Results: Participants described fragmentation that was structural (separate institutions, waitlists, referrals), relational (no provider who followed them over time), and informational (outdated resource listings, unknown application status). In response, they performed what we term compensatory integration work—building resource knowledge, directing their own care, and advocating for providers’ attention and action—while lacking the basic resources for such work and dealing with the daily demands of homelessness. Relationships with community-based organization staff, volunteers, and peers eased this work by offering practical and emotional support. Conclusion: Fragmented ecosystems push the work of integration onto people with the least capacity to carry it out. Reducing this burden requires shared responsibility for care integration, financing for cross-sector collaboration, and concrete steps such as funding relational continuity and maintaining accurate resource information. Otherwise, care integration will continue to depend on the resourcefulness of those navigating fragmented ecosystems.

  • Open Access

    Development, Feasibility, Acceptability, and Usability of an Artificial Intelligence–Powered Chatbot (Suzy) to Support Patients in Substance Use Disorder Recovery: Multiphase Study

    by Warren Scott Comulada, Dallas Swendeman, Y. Xian Ho, Joanna M Streck, Delta-Marie Lewis, Roxana Rezai, David Warren, Gladys Pachas, Maria Chandler, Jonathan L Jackson, et al.

    Background: Substance use disorder (SUD) remains a major public health crisis in the United States, with significant challenges in treatment access, retention, and workforce capacity. SUD care teams, including addiction medicine physicians and peer recovery coaches (PRCs), support patients receiving SUD treatment but face heavy workloads and burnout. Artificial intelligence (AI) innovations, particularly large language model (LLM)–based chatbots, may extend PRC support and provide patients with on-demand recovery support between clinic visits and PRC contacts. However, evidence on their development, feasibility, acceptability, and usability in addiction services remains limited. Objective: This study describes the development, feasibility, acceptability, and usability of an AI-powered health coaching chatbot (Suzy) designed to support patients in SUD recovery. Methods: A total of 2 clinicians, 5 researchers, and 2 technology developers led a small, multiphase pilot study. In the formative phase, they conducted focus groups and qualitative in-depth interviews with 12 health care professionals and 8 patients with substance use histories to specify chatbot functions and develop a rule-based chatbot. In phase 2, they conducted usability testing of the rule-based chatbot with 8 patients who reported substance use and completed standardized tasks, surveys, and qualitative interviews. Measures included the System Usability Scale (SUS), Net Promoter Score (NPS), and Single Ease of Use Question (SEQ). In phase 3, they developed an LLM-based chatbot co-designed and fine-tuned with PRCs and other SUD experts. Results: Rule-based chatbot functions included craving management, appointment reminders, resource referrals, care team contacts, and goal setting. Usability task testing supported feasibility. In this small pilot sample, quantitative and qualitative feedback indicated acceptability and usability, with an average SUS score of 93 (benchmark 68), an NPS of 63 (benchmark 35), and a mean SEQ score of 6.5/7. Patients valued Suzy’s approachable, nonjudgmental language and features that promoted accountability, self-monitoring, and 24/7 availability, while emphasizing that chatbots should supplement but not replace human support. The LLM-based chatbot development emphasized information accuracy, safety escalation protocols to mitigate risks of inappropriate chatbot responses, human-in-the-loop features, and expanded conversational flexibility and personal tailoring. Conclusions: In this pilot study, a rule-based chatbot designed to support SUD care demonstrated feasibility, usability, and acceptability. LLM-based chatbot development required more robust safety and emergency reporting features, while offering more patient-responsive conversational functions. By providing on-demand coaching, referrals, and reminders, Suzy may extend the reach of care teams, alleviate provider burden, and enhance patient engagement. Additional work is needed to understand how to best integrate Suzy into patients’ recovery journeys to ensure human support remains accessible and prioritized. LLM evaluation was based on expert testing and safety review. Clinical effectiveness, including the impact on substance use, was not evaluated. Next steps include evaluating the LLM chatbot in real-world settings with larger samples and assessing its efficacy in reducing substance use.


  • Open Access

    What are the true medical concerns facing unsheltered homeless adults?

    by Randall Kuhn, Jessie Chien, Evan Michael Shannon, Sonali Saluja, Giovanni Righi, Mohammad Damra, Coley M King, Lillian Gelberg, Benjamin F Henwood

    Introduction: Estimates of behavioral and physical health conditions among unsheltered populations are often used to guide policy initiatives for healthcare and housing, yet robust, disease-specific prevalence estimates remain limited. We address this gap by leveraging multiple data sources to generate more reliable estimates of specific disease conditions for unsheltered adults. Methods: We triangulate data from three sources in Los Angeles County, which has the largest unsheltered population in the US: a probability-based survey of adults experiencing homelessness (n = 898), and electronic health records from two field-based medical providers (n = 1683; n = 527). We report univariate estimates by source and summarize across sources. Results: Across sources, median prevalence was 40% for mental health conditions (range: 40%-61%) and 33% for substance use conditions (range: 31%-43%). Post-traumatic stress disorder and major depression were the most common mental health conditions (median 22%), and stimulant use disorder was the most common substance use condition (median 17%). Physical health conditions were more common than both mental health and substance use conditions overall (median 57%; range: 49%-73%). Conclusion: These findings underscore that housing and health care are complementary components of care and support, integrating field-based medical services with permanent housing investments.


  • Open Access

    Association of Antidepressant and/or Benzodiazepine Use With Cannabis Frequency and Differences by Sex in Primary Care Patients With Depression and/or Anxiety

    by Naira Setrakian, Lillian Gelberg, Pamina M. Gorbach, Marissa J. Seamans, Steven Shoptaw, Marjan Javanbakht

    Background: Cannabis use is increasingly prevalent among adults with depression and anxiety, raising concerns about potential interactions with prescribed antidepressant medication and/or benzodiazepines. Using data from a large, university health system in Los Angeles, this cross‐sectional study examined the association between having an active prescription of antidepressant medication and/or benzodiazepines and frequency of cannabis use by sex in adult primary care patients with depression and/or anxiety. Methods: This analysis included 45,693 adult male and female patients with a current diagnosis of depression and/or anxiety (International Classification of Diseases, Tenth Revision [ICD‐10] codes: F41, F33) listed in their electronic health record (EHR). Among patients with a diagnosis, those with an active prescription for antidepressant medication and/or benzodiazepines were identified using EHR data. Cannabis use was assessed as part of routine patient self‐administered screening via the EHR‐based Tobacco and Cannabis Questionnaire (TCQ) from January 2021 to June 2023. Sex‐stratified multivariable ordinal logistic regression models assessed whether active antidepressant medication and/or benzodiazepine prescriptions were associated with increased cannabis use frequency. Models were adjusted for sexual identity, age, race/ethnicity, employment, and Charlson Comorbidity Index (CCI). Results: Females comprised 67.5% of the sample of patients with depression and/or anxiety; 16.3% reported cannabis use for mental health symptoms. Cannabis use for mental health symptoms among males was 21.2%.Among those with depression and/or anxiety, 46.2% of females and 42.8% of males were prescribed antidepressant medication and/or benzodiazepines. Females with an active prescription for antidepressant medication and/or benzodiazepines had 1.32 times the adjusted odds of more frequent cannabis use (95% CI: 1.24–1.40) compared to those without a prescription. Among males, the adjusted odds ratio was lower at 1.14 (95% CI: 1.06–1.23), suggesting a significant difference by sex. Conclusions: Findings indicate that cannabis use commonly occurs alongside prescribed mental health treatment, underscoring the need for primary care providers to routinely address cannabis and medication co‐use during clinical care.

  • Access to medications for opioid use disorder among primary care patients with homeless experience in the Department of Veterans Affairs

    by Audrey L. Jones, Haojia Li, Jacob D. Baylis, Sophia Huebler, Ying Suo, Yue Zhang, Lillian Gelberg, Susan Zickmund, Tom H. Greene, Stefan G. Kertesz, et al.


  • Open Access

    Development, Feasibility, Acceptability, and Usability of an Artificial Intelligence–Powered Chatbot (Suzy) to Support Patients in Substance Use Disorder Recovery: Multiphase Study (Preprint)

    by Warren Scott Comulada, Dallas Swendeman, Y. Xian Ho, Joanna M Streck, Delta-Marie Lewis, Roxana Rezai, David Warren, Gladys Pachas, Maria Chandler, Jonathan L Jackson, et al.

    BACKGROUND Substance use disorder (SUD) remains a major public health crisis in the United States, with significant challenges in treatment access, retention, and workforce capacity. SUD care teams, including addiction medicine physicians and peer recovery coaches (PRCs), support patients receiving SUD treatment but face heavy workloads and burnout. Artificial intelligence (AI) innovations, particularly large language model (LLM)–based chatbots, may extend PRC support and provide patients with on-demand recovery support between clinic visits and PRC contacts. However, evidence on their development, feasibility, acceptability, and usability in addiction services remains limited. OBJECTIVE This study describes the development, feasibility, acceptability, and usability of an AI-powered health coaching chatbot (Suzy) designed to support patients in SUD recovery. METHODS A total of 2 clinicians, 5 researchers, and 2 technology developers led a small, multiphase pilot study. In the formative phase, they conducted focus groups and qualitative in-depth interviews with 12 health care professionals and 8 patients with substance use histories to specify chatbot functions and develop a rule-based chatbot. In phase 2, they conducted usability testing of the rule-based chatbot with 8 patients who reported substance use and completed standardized tasks, surveys, and qualitative interviews. Measures included the System Usability Scale (SUS), Net Promoter Score (NPS), and Single Ease of Use Question (SEQ). In phase 3, they developed an LLM-based chatbot co-designed and fine-tuned with PRCs and other SUD experts. RESULTS Rule-based chatbot functions included craving management, appointment reminders, resource referrals, care team contacts, and goal setting. Usability task testing supported feasibility. In this small pilot sample, quantitative and qualitative feedback indicated acceptability and usability, with an average SUS score of 93 (benchmark 68), an NPS of 63 (benchmark 35), and a mean SEQ score of 6.5/7. Patients valued Suzy’s approachable, nonjudgmental language and features that promoted accountability, self-monitoring, and 24/7 availability, while emphasizing that chatbots should supplement but not replace human support. The LLM-based chatbot development emphasized information accuracy, safety escalation protocols to mitigate risks of inappropriate chatbot responses, human-in-the-loop features, and expanded conversational flexibility and personal tailoring. CONCLUSIONS In this pilot study, a rule-based chatbot designed to support SUD care demonstrated feasibility, usability, and acceptability. LLM-based chatbot development required more robust safety and emergency reporting features, while offering more patient-responsive conversational functions. By providing on-demand coaching, referrals, and reminders, Suzy may extend the reach of care teams, alleviate provider burden, and enhance patient engagement. Additional work is needed to understand how to best integrate Suzy into patients’ recovery journeys to ensure human support remains accessible and prioritized. LLM evaluation was based on expert testing and safety review. Clinical effectiveness, including the impact on substance use, was not evaluated. Next steps include evaluating the LLM chatbot in real-world settings with larger samples and assessing its efficacy in reducing substance use.

  • Top Cited Papers
    Open Access

    Development of a GPT-4–Powered Virtual Simulated Patient and Communication Training Platform for Medical Students to Practice Discussing Abnormal Mammogram Results With Patients: Multiphase Study

    by Dan Weisman, Alanna Sugarman, Yue Ming Huang, Lillian Gelberg, Patricia A Ganz, Warren Scott Comulada

    Background: Standardized patients (SPs) prepare medical students for difficult conversations with patients. Despite their value, SP-based simulation training is constrained by available resources and competing clinical demands. Researchers are turning to artificial intelligence and large language models, such as generative pretrained transformers, to create communication training that incorporates virtual simulated patients (VSPs). GPT-4 is a large language model advance allowing developers to design virtual simulation scenarios using text-based prompts instead of relying on branching path simulations with prescripted dialogue. These nascent developmental practices have not taken root in the literature to guide other researchers in developing their own simulations. Objective: This study aims to describe our developmental process and lessons learned for creating a GPT-4–driven VSP. We designed the VSP to help medical student learners rehearse discussing abnormal mammography results with a patient as a primary care physician (PCP). We aimed to assess GPT-4’s ability to generate appropriate VSP responses to learners during spoken conversations and provide appropriate feedback on learner performance. Methods: A research team comprised of physicians, a medical student, an educator, an SP program director, a learning experience designer, and a health care researcher conducted the study. A formative phase with in-depth knowledge user interviews informed development, followed by a development phase to create the virtual training module. The team conducted interviews with 5 medical students, 5 PCPs, and 5 breast cancer survivors. They then developed a VSP using simulation authoring software and provided the GPT-4–enabled VSP with an initial prompt consisting of a scenario description, emotional state, and expectations for learner dialogue. It was iteratively refined through an agile design process involving repeated cycles of testing, documenting issues, and revising the prompt. As an exploratory feature, the simulation used GPT-4 to provide written feedback to learners about their performance communicating with the VSP and their adherence to guidelines for difficult conversations. Results: In-depth interviews helped establish the appropriate timing, mode of communication, and protocol for conversations between PCPs and patients during the breast cancer screening process. The scenario simulated a telephone call between a physician and patient to discuss the abnormal results of a diagnostic mammogram that that indicated a need for a biopsy. Preliminary testing was promising. The VSP asked sensible questions about their mammography results and responded to learner inquiries using a voice replete with appropriate emotional inflections. GPT-4 generated performance feedback that successfully identified strengths and areas for improvement using relevant quotes from the learner-VSP conversation, but it occasionally misidentified learner adherence to communication protocols. Conclusions: GPT-4 streamlined development and facilitated more dynamic, humanlike interactions between learners and the VSP compared to branching path simulations. For the next steps, we will pilot-test the VSP with medical students to evaluate its feasibility and acceptability.


  • Open Access

    Development of a GPT-4–Powered Virtual Simulated Patient and Communication Training Platform for Medical Students to Practice Discussing Abnormal Mammogram Results With Patients: Multiphase Study (Preprint)

    by Dan Weisman, Alanna Sugarman, Yue Ming Huang, Lillian Gelberg, Patricia A Ganz, Warren Scott Comulada

    BACKGROUND Standardized patients (SPs) prepare medical students for difficult conversations with patients. Despite their value, SP-based simulation training is constrained by available resources and competing clinical demands. Researchers are turning to artificial intelligence and large language models, such as generative pretrained transformers, to create communication training that incorporates virtual simulated patients (VSPs). GPT-4 is a large language model advance allowing developers to design virtual simulation scenarios using text-based prompts instead of relying on branching path simulations with prescripted dialogue. These nascent developmental practices have not taken root in the literature to guide other researchers in developing their own simulations. OBJECTIVE This study aims to describe our developmental process and lessons learned for creating a GPT-4–driven VSP. We designed the VSP to help medical student learners rehearse discussing abnormal mammography results with a patient as a primary care physician (PCP). We aimed to assess GPT-4’s ability to generate appropriate VSP responses to learners during spoken conversations and provide appropriate feedback on learner performance. METHODS A research team comprised of physicians, a medical student, an educator, an SP program director, a learning experience designer, and a health care researcher conducted the study. A formative phase with in-depth knowledge user interviews informed development, followed by a development phase to create the virtual training module. The team conducted interviews with 5 medical students, 5 PCPs, and 5 breast cancer survivors. They then developed a VSP using simulation authoring software and provided the GPT-4–enabled VSP with an initial prompt consisting of a scenario description, emotional state, and expectations for learner dialogue. It was iteratively refined through an agile design process involving repeated cycles of testing, documenting issues, and revising the prompt. As an exploratory feature, the simulation used GPT-4 to provide written feedback to learners about their performance communicating with the VSP and their adherence to guidelines for difficult conversations. RESULTS In-depth interviews helped establish the appropriate timing, mode of communication, and protocol for conversations between PCPs and patients during the breast cancer screening process. The scenario simulated a telephone call between a physician and patient to discuss the abnormal results of a diagnostic mammogram that that indicated a need for a biopsy. Preliminary testing was promising. The VSP asked sensible questions about their mammography results and responded to learner inquiries using a voice replete with appropriate emotional inflections. GPT-4 generated performance feedback that successfully identified strengths and areas for improvement using relevant quotes from the learner-VSP conversation, but it occasionally misidentified learner adherence to communication protocols. CONCLUSIONS GPT-4 streamlined development and facilitated more dynamic, humanlike interactions between learners and the VSP compared to branching path simulations. For the next steps, we will pilot-test the VSP with medical students to evaluate its feasibility and acceptability.

  • Open Access

    Organizational and patient factors associated with positive primary care experiences for veterans with current or recent homelessness

    by Stefan G. Kertesz, Aerin J. Derussy, April E. Hoge, Allyson L. Varley, Sally K. Holmes, Kevin R. Riggs, Erika L. Austin, Adam J. Gordon, Sonya E. Gabrielian, David E. Pollio, et al.

    Objective: To identify organizational service features associated with positive patient ratings of primary care within primary care clinics tailored to accommodate persons with ongoing and recent experiences of homelessness (PEH).Data Sources and Study Setting: PEH receiving primary care in 29 United States Veterans Health Administration homeless‐tailored clinics were surveyed about their primary care experience using the validated Primary Care Quality‐Homeless (PCQ‐H) survey. Characteristics of the clinics were assessed through surveys of clinic staff using a new organizational survey developed through literature review, site visits, statistical analysis, and consensus deliberation.Study Design: Cross‐sectional examination of patients' ratings of care based on surveys of patients, and of clinic characteristics, analyzed with Classification and Regression Tree (CART) analysis, a form of machine learning.Data Collection Methods: Patient surveys (n = 3394) were obtained from a random sample of enrolled patients by both mail and telephone by an external survey contractor. Staff (n = 52 from 29 clinics) were interviewed by telephone.Principal Findings: This analysis identified service features that impact patient experience favorably, including aspects of patient‐centeredness, team identity, strong external leadership support, and service that reach beyond traditional primary care clinic confines. Results varied according to the patient experience scale analyzed. Individual characteristics of PEH, such as degree of social support, general health, and unsheltered status, were also correlated with how they rate care.Conclusions: Organizational characteristics correlate with ratings of primary care from patients with recent and ongoing homelessness. Primary care programs serving homeless individuals can assure better care based on who they hire, how they foster team identity, what services they provide, and the strength of leadership support to protect a homeless‐focused mission.


  • Open Access

    Pilot randomized controlled trial of biofeedback on reducing psychological and physiological stress among persons experiencing homelessness

    by Adeline M. Nyamathi, Benissa E. Salem, Lillian Gelberg, Dana Rose Garfin, Kate Wolitsky‐Taylor, Sanghyuk S. Shin, Zhaoxia Yu, Angela Hudson, Kartik Yadav, Richard Clarke, et al.

    People experiencing homelessness report increased exposure to traumatic life events and higher rates of depression, anxiety, and post‐traumatic stress disorder as compared with the general population. Heart rate variability‐biofeedback (HRV‐BF) has been shown to decrease symptoms of stress, anxiety, depression, and PTSD. However, HRV‐BF has not been tested with the most vulnerable of populations, homeless adults. The purpose of this randomized controlled trial was to compare the effectiveness of an HRV‐BF intervention versus a Health Promotion (HP) active control intervention focused on improving mental health symptoms among homeless adults. Guided by a community advisory board, homeless adults residing in Skid Row, Los Angeles (n = 40) were randomized to either the HRV‐BF or an active HP control group and received eight weekly, 30‐min sessions over two months, delivered by a nurse‐led community health worker team. Dependent variables of HRV, mental health, anxiety, depression, and PTSD were measured at baseline, the 8‐week session, and/or 2‐month follow‐up. All intervention sessions were completed by 90% (36/40) of participants. Both the HRV‐BF and HP interventions showed significant increases in HRV from baseline to 2‐month follow‐up, with no significant difference between the intervention groups. The HRV‐BF programme revealed a somewhat greater, although non‐significant, improvement in anxiety, depression, and PTSD symptoms than the HP programme. The usefulness of both interventions, focused on emotional and physical health, warrants future studies to examine the value of a combined HRV‐BF and HP intervention.