KHKarim Hanna

Karim Hanna

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

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Karim Hanna has a total of 90 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Generating Question Prompt Lists From Electronic Health Record Data Using Large Language Models: Iterative Evaluation Study

    by Zhe He, Balu Bhasuran, Mia Liza A Lustria, Karim Hanna, Michael Killian, Cindy Shavor, Mandy Dailey, Sai Sidharth Manikandan, Xiao Luo

    Background: Patients frequently access laboratory results through patient portals, but many struggle to interpret these values and formulate relevant questions for their clinicians. Question prompt lists (QPLs) can enhance communication but are rarely tailored to individual clinical contexts. Objective: This study evaluated the feasibility of using large language models (LLMs) to generate patient-friendly, clinically relevant questions grounded in electronic health record (EHR) laboratory data. Methods: We extracted deidentified clinical profiles, including laboratory results, diagnoses, and medications, from patients with chronic conditions (eg, diabetes and chronic kidney disease). Using 9 deidentified clinical profiles from the OneFlorida Data Trust, we generated 486 questions across all rounds: 126 with GPT-4o in round 1, 120 with GPT-4o in round 2, 180 with GPT-4o in round 3, and 60 with LLaMA 3.2 in round 3. Prompt refinements were informed by clinician ratings consisting of 2 binary questions (ie, clear phrasing and clinical validity) and 3 Likert-scale questions (ie, significance for the patient’s health; clinical appropriateness, that is, the likelihood of being asked in primary care setting; and willingness to answer). Refinements were incorporated after each round. Patient participants then evaluated selected questions for understandability, perceived usefulness, and intention to use. Readability was assessed with standard indices. Results: Iterative clinician feedback improved question clarity and reduced clinically irrelevant suggestions. Across rounds, GPT-4o consistently produced more coherent and patient-friendly questions, while LLaMA 3.2 demonstrated competitive performance on Likert-scale metrics. It exhibited greater variability in clinical appropriateness as noted by clinicians. In round 3, the binary metric “clear phrasing” reached a ceiling effect for both models, while clinical validity ratings showed greater variability, particularly from one clinician. Likert-scale evaluations tended to favor LLaMA 3.2 across all 3 clinicians for clinical appropriateness (3.37‐4.82 vs 3.02‐4.67), significance for the patient’s health (3.38‐4.28 vs 2.97‐3.83), and willingness to answer (3.17‐4.70 vs 2.82‐4.47), with multiple comparisons reaching statistical significance after Bonferroni correction. Patient evaluation (N=134) of GPT-4o–generated questions showed that 25 of 30 questions had moderate to high understandability (average Likert score ≥3.5), and 19 of 30 questions had moderate to high usefulness (average Likert score ≥3.5). Conclusions: This study supports the feasibility of using LLMs with structured EHR-derived laboratory data to generate contextualized QPLs, but model outputs varied in clinical appropriateness and readability. Clinician-in-the-loop review remains necessary before patient-facing use.


  • Prevalence of statin utilization in cancer survivors in a family medicine clinic

    by Raechel T. White, Cyrille K. Cornelio, Athanasios Tsalatsanis, Karim Hanna, Jay Sharad Maru, Emily Vial, Nicholas W. Carris

    Purpose As people live longer after a cancer diagnosis, there is an increasing need to focus on chronic disease state management in the primary care setting. Particular attention to appropriate statin prescribing in cancer survivors is warranted given a heightened risk of atherosclerotic cardiovascular disease (ASCVD). The purpose of this study was to further understand cardiovascular care in cancer survivors by identifying the proportion of cancer survivors with a prescription for appropriate statin therapy and to identify associated factors. Methods A retrospective cohort study was conducted to identify cancer survivors aged 40–75 years with a diagnosis of breast, lung, colorectal, or lymphatic and hematopoietic cancers cared for within family medicine clinics in a large academic health system. Statin prescriptions were assessed to identify the percentage of statin candidates prescribed a statin based on a diagnosis of ASCVD or type 2 diabetes. Results Data from 8763 cancer survivors were included in the final analysis. Among patients with ASCVD, 66% were prescribed a statin. In patients with type 2 diabetes without ASCVD, 58% were prescribed a statin. Female sex was negatively associated with statin prescribing (OR 0.88, 95% CI 0.80–0.97). Conclusions Statins are underutilized in cancer survivors, as they are in the general population. This may present an area of health disparity given the heightened risk of ASCVD in cancer survivors. Cancer survivors suffer from a heightened risk of death from CVD. As survivors enjoy longer lifespans after diagnosis, there is a need to optimize cardiometabolic disease state management in this high-risk population.

  • Open Access

    Comparative Performance of Large Language Models in Muscle Histology Classification Highlights Enhanced Accuracy of ChatGPT-4o in Tissue Identification

    by Parth Shah, David J. Boughanem, John M. Templeton, Marzenna Wiranowska, Karim Hanna

    Introduction One of the most promising avenues of artificial intelligence (AI) integration into medicine is its examination, evaluation, and characterization of pathological slides. The use of large language models (LLMs), the AI model subtype that is becoming increasingly popular, in pathological applications remains unexplored. This study investigates the histological image recognition capabilities of the multimodal models Gemini 1.5 Flash, ChatGPT-4o, and Claude 3.5 Sonnet and assesses their suitability for clinical or medical education use. Methods The models were evaluated using 300 digital histology images derived from the University of South Florida Morsani College of Medicine Virtual Microscopy database, with a prompt to ascertain each model's ability to identify tissue type and plane of sectioning used. The images included the three subtypes in both longitudinal and transverse planes of sectioning. Standard machine learning metrics such as precision, recall, accuracy, and F1 score were used to classify and evaluate each model’s abilities. Results In the prediction of tissue type, OpenAI’s ChatGPT had the highest metrics with an F1 score of 0.772, while Claude yielded an F1 score of 0.380, and Gemini produced a 0.460 F1 score. In the prediction of sectioning, ChatGPT produced an F1 score of 0.396, while Claude produced a value of 0.472, and Gemini yielded 0.344. Conclusion Overall, the results indicate that ChatGPT is most effective at identifying tissues. However, the inaccuracy demonstrated in evaluating sectioning compared to other models leaves room for improvement in its overall accuracy across varying tissue samples to reliably supplement medical education or clinical use.


  • Open Access

    Impact of Pharmacist-Led Continuous Glucose Monitoring on Clinical Outcomes in People With Type 2 Diabetes in Primary Care: Protocol for a Prospective Cohort Study

    by Kevin Cowart, Raechel T White, Kevin Olson, Nicholas W Carris, Karim Hanna, Janice Zgibor

    Background: Continuous glucose monitoring (CGM) is increasingly being recognized as the new standard of care for glycemic monitoring in people with type 2 diabetes (T2D). However, despite advances in therapeutics and technology, glycemic control remains suboptimal. Team-based approaches involving pharmacists, particularly in primary care, have shown to be effective in addressing these shortcomings yet have not been rigorously evaluated in the literature. Objective: Herein we present the protocol for a study that seeks to evaluate the change in hemoglobin A1c (HbA1c) in people with T2D using CGM under a pharmacist-led approach as compared with a pharmacist-led approach using no CGM (only self-monitoring blood glucose with a glucometer). We will also assess changes in CGM-derived glycemic outcomes, health behavior, and safety outcomes among the pharmacist-led CGM cohort. Methods: This is a 12-week prospective cohort study in an academic family medicine department. We will enroll adults with T2D and a HbA1c level of ≥8%. Participants in the intervention cohort will wear a CGM sensor (FreeStyle Libre 2) for 12 weeks and receive structured diabetes self-management education and support from a pharmacist. Each participant in the intervention group will have 5 visits with a pharmacist. The primary objective is the between-group difference in change in HbA1c levels from baseline to 12 weeks between the intervention and historical cohort. Secondary objectives include a change in CGM-derived metrics among the intervention group from baseline to 12 weeks, and a change in health behavior via the Summary of Diabetes Self-Care Activities measure from baseline to 12 weeks in the intervention cohort. A CGM survey will also be administered to participants in the intervention cohort to evaluate changes in diet, physical activity, general lifestyle, and medication adherence. Safety endpoints will also be evaluated. The primary and secondary outcomes will be analyzed within and between groups using descriptive statistics, with a multivariable regression analysis conducted as appropriate to adjust for potential known confounding effects. Results: This study was funded in July 2023. We began enrolling participants in December 2024. At the time of writing, 3 participants have been enrolled. It is anticipated that we will conclude this study in December 2025 and expect to disseminate results in March 2026. Conclusions: Results of this study will further elucidate the role of pharmacist-led CGM in primary care. Trial Registration: ClinicalTrials.gov NCT06572306; https://clinicaltrials.gov/study/NCT06572306 International Registered Report Identifier (IRRID): PRR1-10.2196/67014


  • Top Cited Papers
    Open Access

    Generative Artificial Intelligence in Medical Education—Policies and Training at US Osteopathic Medical Schools: Descriptive Cross-Sectional Survey

    by Tsunagu Ichikawa, Elizabeth Olsen, Arathi Vinod, Noah Glenn, Karim Hanna, Gregg C Lund, Stacey Pierce-Talsma

    Background: Interest has recently increased in generative artificial intelligence (GenAI), a subset of artificial intelligence that can create new content. Although the publicly available GenAI tools are not specifically trained in the medical domain, they have demonstrated proficiency in a wide range of medical assessments. The future integration of GenAI in medicine remains unknown. However, the rapid availability of GenAI with a chat interface and the potential risks and benefits are the focus of great interest. As with any significant medical advancement or change, medical schools must adapt their curricula to equip students with the skills necessary to become successful physicians. Furthermore, medical schools must ensure that faculty members have the skills to harness these new opportunities to increase their effectiveness as educators. How medical schools currently fulfill their responsibilities is unclear. Colleges of Osteopathic Medicine (COMs) in the United States currently train a significant proportion of the total number of medical students. These COMs are in academic settings ranging from large public research universities to small private institutions. Therefore, studying COMs will offer a representative sample of the current GenAI integration in medical education. Objective: This study aims to describe the policies and training regarding the specific aspect of GenAI in US COMs, targeting students, faculty, and administrators. Methods: Web-based surveys were sent to deans and Student Government Association (SGA) presidents of the main campuses of fully accredited US COMs. The dean survey included questions regarding current and planned policies and training related to GenAI for students, faculty, and administrators. The SGA president survey included only those questions related to current student policies and training. Results: Responses were received from 81% (26/32) of COMs surveyed. This included 47% (15/32) of the deans and 50% (16/32) of the SGA presidents (with 5 COMs represented by both the deans and the SGA presidents). Most COMs did not have a policy on the student use of GenAI, as reported by the dean (14/15, 93%) and the SGA president (14/16, 88%). Of the COMs with no policy, 79% (11/14) had no formal plans for policy development. Only 1 COM had training for students, which focused entirely on the ethics of using GenAI. Most COMs had no formal plans to provide mandatory (11/14, 79%) or elective (11/15, 73%) training. No COM had GenAI policies for faculty or administrators. Eighty percent had no formal plans for policy development. Furthermore, 33.3% (5/15) of COMs had faculty or administrator GenAI training. Except for examination question development, there was no training to increase faculty or administrator capabilities and efficiency or to decrease their workload. Conclusions: The survey revealed that most COMs lack GenAI policies and training for students, faculty, and administrators. The few institutions with policies or training were extremely limited in scope. Most institutions without current training or policies had no formal plans for development. The lack of current policies and training initiatives suggests inadequate preparedness for integrating GenAI into the medical school environment, therefore, relegating the responsibility for ethical guidance and training to the individual COM member.


  • Open Access

    Impact of Pharmacist-Led Continuous Glucose Monitoring on Clinical Outcomes in People With Type 2 Diabetes in Primary Care: Protocol for a Prospective Cohort Study (Preprint)

    by Kevin Cowart, Raechel T White, Kevin Olson, Nicholas W Carris, Karim Hanna, Janice Zgibor

    BACKGROUND Continuous glucose monitoring (CGM) is increasingly being recognized as the new standard of care for glycemic monitoring in people with type 2 diabetes (T2D). However, despite advances in therapeutics and technology, glycemic control remains suboptimal. Team-based approaches involving pharmacists, particularly in primary care, have shown to be effective in addressing these shortcomings yet have not been rigorously evaluated in the literature. OBJECTIVE Herein we present the protocol for a study that seeks to evaluate the change in hemoglobin A1c (HbA1c) in people with T2D using CGM under a pharmacist-led approach as compared with a pharmacist-led approach using no CGM (only self-monitoring blood glucose with a glucometer). We will also assess changes in CGM-derived glycemic outcomes, health behavior, and safety outcomes among the pharmacist-led CGM cohort. METHODS This is a 12-week prospective cohort study in an academic family medicine department. We will enroll adults with T2D and a HbA1c level of ≥8%. Participants in the intervention cohort will wear a CGM sensor (FreeStyle Libre 2) for 12 weeks and receive structured diabetes self-management education and support from a pharmacist. Each participant in the intervention group will have 5 visits with a pharmacist. The primary objective is the between-group difference in change in HbA1c levels from baseline to 12 weeks between the intervention and historical cohort. Secondary objectives include a change in CGM-derived metrics among the intervention group from baseline to 12 weeks, and a change in health behavior via the Summary of Diabetes Self-Care Activities measure from baseline to 12 weeks in the intervention cohort. A CGM survey will also be administered to participants in the intervention cohort to evaluate changes in diet, physical activity, general lifestyle, and medication adherence. Safety endpoints will also be evaluated. The primary and secondary outcomes will be analyzed within and between groups using descriptive statistics, with a multivariable regression analysis conducted as appropriate to adjust for potential known confounding effects. RESULTS This study was funded in July 2023. We began enrolling participants in December 2024. At the time of writing, 3 participants have been enrolled. It is anticipated that we will conclude this study in December 2025 and expect to disseminate results in March 2026. CONCLUSIONS Results of this study will further elucidate the role of pharmacist-led CGM in primary care. CLINICALTRIAL ClinicalTrials.gov NCT06572306; https://clinicaltrials.gov/study/NCT06572306 INTERNATIONAL REGISTERED REPORT PRR1-10.2196/67014

  • Open Access

    Performance of Language Models on the Family Medicine In-Training Exam

    by Rana E. Hanna, Logan R. Smith, Rahul Mhaskar, Karim Hanna

    Background and Objectives: Artificial intelligence (AI), such as ChatGPT and Bard, has gained popularity as a tool in medical education. The use of AI in family medicine has not yet been assessed. The objective of this study is to compare the performance of three large language models (LLMs; ChatGPT 3.5, ChatGPT 4.0, and Google Bard) on the family medicine in-training exam (ITE). Methods: The 193 multiple-choice questions of the 2022 ITE, written by the American Board of Family Medicine, were inputted in ChatGPT 3.5, ChatGPT 4.0, and Bard. The LLMs’ performance was then scored and scaled. Results: ChatGPT 4.0 scored 167/193 (86.5%) with a scaled score of 730 out of 800. According to the Bayesian score predictor, ChatGPT 4.0 has a 100% chance of passing the family medicine board exam. ChatGPT 3.5 scored 66.3%, translating to a scaled score of 400 and an 88% chance of passing the family medicine board exam. Bard scored 64.2%, with a scaled score of 380 and an 85% chance of passing the boards. Compared to the national average of postgraduate year 3 residents, only ChatGPT 4.0 surpassed the residents’ mean of 68.4%. Conclusions: ChatGPT 4.0 was the only LLM that outperformed the family medicine postgraduate year 3 residents’ national averages on the 2022 ITE, providing robust explanations and demonstrating its potential use in delivering background information on common medical concepts that appear on board exams.


  • Open Access

    Artificial Intelligence-Prompted Explanations of Common Primary Care Diagnoses

    by Mafaz Kattih, Max Bressler, Logan R. Smith, Anthony Schinelli, Rahul Mhaskar, Karim Hanna

    Background: Artificial intelligence (AI)-generated explanations about medical topics may be clearer and more accessible than traditional evidence-based sources, enhancing patient understanding and autonomy. We evaluated different AI explanations for patients about common diagnoses to aid in patient care. Methods: We prompted ChatGPT 3.5 , Google Bard, HuggingChat, and Claude 2 separately to generate a short patient education paragraph about seven common diagnoses. We used the Flesch Reading Ease (FRE) and Flesch-Kincaid Grade Level (FKGL) to evaluate the readability and grade level of the responses. We used the Agency for Healthcare Research and Quality’s Patient Education Materials Assessment Tool (PEMAT) grading rubric to evaluate the understandability and actionability of responses. Results: Claude 2 demonstrated scores of FRE (67.0), FKGL (7.4), and PEMAT, 69% for understandability, and 34% for actionability. ChatGPT scores were FRE (58.5), FKGL (9.3), PEMAT (69% and 31%, respectively). Google Bard scores were FRE (50.1), FKGL (9.9), PEMAT (52% and 23%). HuggingChat scores were FRE (48.7) and FKGL (11.6), PEMAT (57% and 29%). Conclusion: Claude 2 and ChatGPT demonstrated superior readability and understandability, but practical application and patient outcomes need further exploration. This study is limited by the rapid development of these tools with newer improved models replacing the older ones. Additionally, the accuracy and clarity of AI responses is based on that of the user-generated response. The PEMAT grading rubric is also mainly used for patient information leaflets that include visual aids and may contain subjective evaluations.