Michelle Wang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Michelle Wang has a total of 62 co-authors. Here are the top co-authors by number of shared publications:
- Atul Janardhan Butte
Bakar Computational Health Sciences Institute, University of California, San Francisco, San Francisco, USA
3 shared publications - Artur Belov
Office of Biostatistics and Epidemiology, Center for Biologics Evaluation and Research, US Food and Drug Administration, Silver Spring, MD
2 shared publications - Aryan Ayati
Department of Medicine, University of California, San Francisco (UCSF), San Francisco, USA
2 shared publications Ayan Patel
Bakar Computational Health Sciences Institute, University of California San Francisco, San Francisco, CA
2 shared publications- Barbee Whitaker
Office of Biostatistics and Epidemiology, Center for Biologics Evaluation and Research, US Food and Drug Administration, Silver Spring, MD
2 shared publications Benjamin Rubin
Bakar Computational Health Sciences Institute, University of California San Francisco, San Francisco, CA
2 shared publications
Latest Publications
- Open Access
- 30 October 2026
- Published by Springer Nature in Oncology and Therapy
- p. 1-21
Real-World Treatment Patterns and Outcomes of Zanubrutinib and Acalabrutinib in Patients with Chronic Lymphocytic Leukemia and Small Lymphocytic Lymphoma at University of California Academic Health Centers: A Retrospective Cohort Study
- 30 October 2026
- Published by Springer Nature in Oncology and Therapy
- p. 1-21
Open AccessReal-World Treatment Patterns and Outcomes of Zanubrutinib and Acalabrutinib in Patients with Chronic Lymphocytic Leukemia and Small Lymphocytic Lymphoma at University of California Academic Health Centers: A Retrospective Cohort Study
- Open Access
- 14 November 2025
- Published by American Society of Hematology in Blood Advances
- Vol. 9 (22) , 5837-5852
An AI model classifies risks of early relapse post–CAR T-cell therapy in a multicenter real-world population with DLBCL
Accumulating real-world (RW) evidence of axicabtagene ciloleucel (axi-cel) has demonstrated comparable performance to that of pivotal trials. However, ∼57% of patients eventually relapse, with most requiring additional therapies. Being able to identify patients with risk of early relapse enables clinicians to consider additional interventions to extend survival outcomes. This study aimed to first comprehensively evaluate the RW performance of axi-cel in the multicenter University of California Health Systems using automated computational approaches. Second, we developed a decision tree machine learning (ML) model to identify patients with risks of early relapse within 6 months. A total of 416 adult patients with diffuse large B-cell lymphoma (DLBCL) receiving axi-cel between 2017 and 2024 were included in the study. The median progression-free survival (PFS) and overall survival (OS) were 10.1 and 54.4 months; the 18-month PFS and OS rates were 41.6% and 65.5%, respectively. Severe CRS and ICANS were observed in 18.8% and 32.5% of patients. The ML model, relying on age and 6 routinely measured laboratory tests (lactate dehydrogenase, C-reactive protein, ferritin, hematocrit, platelet count, and prothrombin time), achieved a high area under the receiver operating characteristic curve score of 0.82. The decision curve analysis indicated positive net benefit of the model across a broad range (0-0.7) of decision thresholds, suggesting clinical utility in diverse scenarios. This study further confirmed the RW performance of axi-cel in diverse populations. Our ML model represented a novel approach to identify patients that may benefit from additional interventions to extend survival outcomes. Following prospective confirmation study, our model and decision tree approach may support clinical decision making in patients with high-risk DLBCL.
- 14 November 2025
- Published by American Society of Hematology in Blood Advances
- Vol. 9 (22) , 5837-5852
Open AccessAn AI model classifies risks of early relapse post–CAR T-cell therapy in a multicenter real-world population with DLBCL
Accumulating real-world (RW) evidence of axicabtagene ciloleucel (axi-cel) has demonstrated comparable performance to that of pivotal trials. However, ∼57% of patients eventually relapse, with most requiring additional therapies. Being able to identify patients with risk of early relapse enables clinicians to consider additional interventions to extend survival outcomes. This study aimed to first comprehensively evaluate the RW performance of axi-cel in the multicenter University of California Health Systems using automated computational approaches. Second, we developed a decision tree machine learning (ML) model to identify patients with risks of early relapse within 6 months. A total of 416 adult patients with diffuse large B-cell lymphoma (DLBCL) receiving axi-cel between 2017 and 2024 were included in the study. The median progression-free survival (PFS) and overall survival (OS) were 10.1 and 54.4 months; the 18-month PFS and OS rates were 41.6% and 65.5%, respectively. Severe CRS and ICANS were observed in 18.8% and 32.5% of patients. The ML model, relying on age and 6 routinely measured laboratory tests (lactate dehydrogenase, C-reactive protein, ferritin, hematocrit, platelet count, and prothrombin time), achieved a high area under the receiver operating characteristic curve score of 0.82. The decision curve analysis indicated positive net benefit of the model across a broad range (0-0.7) of decision thresholds, suggesting clinical utility in diverse scenarios. This study further confirmed the RW performance of axi-cel in diverse populations. Our ML model represented a novel approach to identify patients that may benefit from additional interventions to extend survival outcomes. Following prospective confirmation study, our model and decision tree approach may support clinical decision making in patients with high-risk DLBCL.
- 9 September 2025
- Published by American Diabetes Association in Diabetes Care
- Vol. 48 (11) , 1914-1919
Quantifying Barriers to Diabetic Eye Screening: A Two-Center Study at the University of California
OBJECTIVE: This study aimed to evaluate the diabetic eye disease screening continuum at two academic centers and identify its barriers. RESEARCH DESIGN AND METHODS: We analyzed health records from the University of California, San Francisco, and University of California, Irvine, to identify primary care patients needing diabetic eye screening. We tracked referrals, screenings, diagnoses, and treatments to evaluate predictors and the impact of an automated referral system. We analyzed physician notes using GPT-4o to determine reasons for missed screenings. RESULTS: Of 8,240 unscreened patients with type 2 diabetes mellitus (T2DM), 43% received a referral, and only 16% completed screening within 1 year. Demographic, provider, and socioeconomic factors predicted adherence, with referrals being the strongest predictor. An automated referral system could improve screening rates to 22–34%. Clinician notes cited comorbidities, scheduling challenges, logistical issues, coronavirus disease 2019, and personal circumstances as barriers. CONCLUSIONS: Many patients with T2DM remain unscreened after primary care visits. Although an automated referral system may partially improve adherence, additional tailored strategies are needed.
- 9 September 2025
- Published by American Diabetes Association in Diabetes Care
- Vol. 48 (11) , 1914-1919
Quantifying Barriers to Diabetic Eye Screening: A Two-Center Study at the University of California
OBJECTIVE: This study aimed to evaluate the diabetic eye disease screening continuum at two academic centers and identify its barriers. RESEARCH DESIGN AND METHODS: We analyzed health records from the University of California, San Francisco, and University of California, Irvine, to identify primary care patients needing diabetic eye screening. We tracked referrals, screenings, diagnoses, and treatments to evaluate predictors and the impact of an automated referral system. We analyzed physician notes using GPT-4o to determine reasons for missed screenings. RESULTS: Of 8,240 unscreened patients with type 2 diabetes mellitus (T2DM), 43% received a referral, and only 16% completed screening within 1 year. Demographic, provider, and socioeconomic factors predicted adherence, with referrals being the strongest predictor. An automated referral system could improve screening rates to 22–34%. Clinician notes cited comorbidities, scheduling challenges, logistical issues, coronavirus disease 2019, and personal circumstances as barriers. CONCLUSIONS: Many patients with T2DM remain unscreened after primary care visits. Although an automated referral system may partially improve adherence, additional tailored strategies are needed.
- 20 August 2025
- Published by American Diabetes Association in Clinical Diabetes
- Vol. 43 (4) , 545-553
Assessment of Large Language Models for Enhancing Diabetologist-Developed Personalized Treatment Plans in Complex Type 2 Diabetes
Treatment decisions for type 2 diabetes are complex and multifactorial. This study evaluated the utility of OpenAI’s generative pre-trained transformer-4 (GPT-4) large language model in enhancing treatment plans for people with diabetes. The authors assessed GPT-4’s ability to extract and critically evaluate diabetologist-developed treatment plans from patients’ clinical notes. GPT-4 accurately discerned personalized treatment goals but performed suboptimally when identifying plans that could be further improved and personalized. However, 40% of GPT-4’s reasoning and explanations were still considered helpful by experienced academic diabetologists, suggesting a potential broad benefit to less expert physicians, such as those in primary care settings.
- 20 August 2025
- Published by American Diabetes Association in Clinical Diabetes
- Vol. 43 (4) , 545-553
Assessment of Large Language Models for Enhancing Diabetologist-Developed Personalized Treatment Plans in Complex Type 2 Diabetes
Treatment decisions for type 2 diabetes are complex and multifactorial. This study evaluated the utility of OpenAI’s generative pre-trained transformer-4 (GPT-4) large language model in enhancing treatment plans for people with diabetes. The authors assessed GPT-4’s ability to extract and critically evaluate diabetologist-developed treatment plans from patients’ clinical notes. GPT-4 accurately discerned personalized treatment goals but performed suboptimally when identifying plans that could be further improved and personalized. However, 40% of GPT-4’s reasoning and explanations were still considered helpful by experienced academic diabetologists, suggesting a potential broad benefit to less expert physicians, such as those in primary care settings.
- 22 November 2024
- Published by SAGE Publications in Journal of Diabetes Science and Technology
- Vol. 19 (1) , 246-264
Artificial Intelligence to Diagnose Complications of Diabetes
Artificial intelligence (AI) is increasingly being used to diagnose complications of diabetes. Artificial intelligence is technology that enables computers and machines to simulate human intelligence and solve complicated problems. In this article, we address current and likely future applications for AI to be applied to diabetes and its complications, including pharmacoadherence to therapy, diagnosis of hypoglycemia, diabetic eye disease, diabetic kidney diseases, diabetic neuropathy, diabetic foot ulcers, and heart failure in diabetes. Artificial intelligence is advantageous because it can handle large and complex datasets from a variety of sources. With each additional type of data incorporated into a clinical picture of a patient, the calculation becomes increasingly complex and specific. Artificial intelligence is the foundation of emerging medical technologies; it will power the future of diagnosing diabetes complications.
- 22 November 2024
- Published by SAGE Publications in Journal of Diabetes Science and Technology
- Vol. 19 (1) , 246-264
Artificial Intelligence to Diagnose Complications of Diabetes
Artificial intelligence (AI) is increasingly being used to diagnose complications of diabetes. Artificial intelligence is technology that enables computers and machines to simulate human intelligence and solve complicated problems. In this article, we address current and likely future applications for AI to be applied to diabetes and its complications, including pharmacoadherence to therapy, diagnosis of hypoglycemia, diabetic eye disease, diabetic kidney diseases, diabetic neuropathy, diabetic foot ulcers, and heart failure in diabetes. Artificial intelligence is advantageous because it can handle large and complex datasets from a variety of sources. With each additional type of data incorporated into a clinical picture of a patient, the calculation becomes increasingly complex and specific. Artificial intelligence is the foundation of emerging medical technologies; it will power the future of diagnosing diabetes complications.
- Open Access
- 19 June 2023
- Published by Oxford University Press (OUP) in The Oncologist
- Vol. 28 (10) , e950-e959
Evaluating Survival After Hospitalization Due to Immune-Related Adverse Events From Checkpoint Inhibitors
As immune checkpoint inhibitors (CPI) are increasingly approved for cancer treatment, hospitalizations related to severe immune-related adverse events (irAE) will increase. Here, we identify patients hospitalized due to irAEs and describe survival outcomes across irAE, CPI, and cancer type. We identified patients hospitalized at our institution from January 2012 to December 2020 due to irAEs. Survival was analyzed using Kaplan-Meier survival curves with log-rank tests. Of 3137 patients treated with CPIs, 114 (3.6%) were hospitalized for irAEs, resulting in 124 hospitalizations. Gastrointestinal (GI)/hepatic, endocrine, and pulmonary irAEs were the most common causes of irAE-related hospitalization. After CPI initiation, the average time to hospitalization was 141 days. Median survival from hospital admission was 980 days. Patients hospitalized due to GI/hepatic and endocrine irAEs had longer median survival than patients with pulmonary irAEs (795 and 949 days vs. 83 days [P < .001]). Patients with melanoma and renal cell carcinoma had longer median survival than patients with lung cancer (2792 days and not reached vs. 159 days [P < .001]). There was longer median survival in the combination group compared to the PD-(L)1 group (1471 vs. 529 days [P = .04]). As CPI use increases, irAE-related hospitalizations will as well. These findings suggest that among patients hospitalized for irAEs, survival differs by irAE and cancer type, with worse survival for patients with irAE pneumonitis or lung cancer. This real-world data contributes to research pertaining to hospitalization due to severe irAEs, which may inform patient counseling and treatment decision-making.
- 19 June 2023
- Published by Oxford University Press (OUP) in The Oncologist
- Vol. 28 (10) , e950-e959
Open AccessEvaluating Survival After Hospitalization Due to Immune-Related Adverse Events From Checkpoint Inhibitors
As immune checkpoint inhibitors (CPI) are increasingly approved for cancer treatment, hospitalizations related to severe immune-related adverse events (irAE) will increase. Here, we identify patients hospitalized due to irAEs and describe survival outcomes across irAE, CPI, and cancer type. We identified patients hospitalized at our institution from January 2012 to December 2020 due to irAEs. Survival was analyzed using Kaplan-Meier survival curves with log-rank tests. Of 3137 patients treated with CPIs, 114 (3.6%) were hospitalized for irAEs, resulting in 124 hospitalizations. Gastrointestinal (GI)/hepatic, endocrine, and pulmonary irAEs were the most common causes of irAE-related hospitalization. After CPI initiation, the average time to hospitalization was 141 days. Median survival from hospital admission was 980 days. Patients hospitalized due to GI/hepatic and endocrine irAEs had longer median survival than patients with pulmonary irAEs (795 and 949 days vs. 83 days [P < .001]). Patients with melanoma and renal cell carcinoma had longer median survival than patients with lung cancer (2792 days and not reached vs. 159 days [P < .001]). There was longer median survival in the combination group compared to the PD-(L)1 group (1471 vs. 529 days [P = .04]). As CPI use increases, irAE-related hospitalizations will as well. These findings suggest that among patients hospitalized for irAEs, survival differs by irAE and cancer type, with worse survival for patients with irAE pneumonitis or lung cancer. This real-world data contributes to research pertaining to hospitalization due to severe irAEs, which may inform patient counseling and treatment decision-making.
- Open Access
- 29 May 2023
- Published by Wiley in Transfusion
- Vol. 63 (7) , 1298-1309
Novel computational methods on electronic health record yields new estimates of transfusion‐associated circulatory overload in populations enriched with high‐risk patients
Background: Transfusion‐associated circulatory overload (TACO) is a severe adverse reaction (AR) contributing to the leading cause of mortality associated with transfusions. As strategies to mitigate TACO have been increasingly adopted, an update of prevalence rates and risk factors associated with TACO using the growing sources of electronic health record (EHR) data can help understand transfusion safety.Study Design and Methods: This retrospective study aimed to provide a timely and reproducible assessment of prevalence rates and risk factors associated with TACO. Novel natural language processing methods, now made publicly available on GitHub, were developed to extract ARs from 3178 transfusion reaction reports. Other patient‐level data were extracted computationally from UCSF EHR between 2012 and 2022. The odds ratio estimates of risk factors were calculated using a multivariate logistic regression analysis with case‐to‐control matched on sex and age at a ratio of 1:5.Results: A total of 56,208 patients received transfusions (total 573,533 units) at UCSF during the study period and 102 patients developed TACO. The prevalence of TACO was estimated to be 0.2% per patient (102/total 56,208). Patients with a history of coagulopathy (OR, 1.36; 95% CI, 1.04–1.79) and transplant (OR, 1.99; 95% CI, 1.48–2.68) were associated with increased odds of TACO.Discussion: While TACO is a serious AR, events remained rare, even in populations enriched with high‐risk patients. Novel computational methods can be used to find and continually surveil for transfusion ARs. Results suggest that patients with history or presence of coagulopathy and organ transplant should be carefully monitored to mitigate potential risks of TACO.
- 29 May 2023
- Published by Wiley in Transfusion
- Vol. 63 (7) , 1298-1309
Open AccessNovel computational methods on electronic health record yields new estimates of transfusion‐associated circulatory overload in populations enriched with high‐risk patients
Background: Transfusion‐associated circulatory overload (TACO) is a severe adverse reaction (AR) contributing to the leading cause of mortality associated with transfusions. As strategies to mitigate TACO have been increasingly adopted, an update of prevalence rates and risk factors associated with TACO using the growing sources of electronic health record (EHR) data can help understand transfusion safety.Study Design and Methods: This retrospective study aimed to provide a timely and reproducible assessment of prevalence rates and risk factors associated with TACO. Novel natural language processing methods, now made publicly available on GitHub, were developed to extract ARs from 3178 transfusion reaction reports. Other patient‐level data were extracted computationally from UCSF EHR between 2012 and 2022. The odds ratio estimates of risk factors were calculated using a multivariate logistic regression analysis with case‐to‐control matched on sex and age at a ratio of 1:5.Results: A total of 56,208 patients received transfusions (total 573,533 units) at UCSF during the study period and 102 patients developed TACO. The prevalence of TACO was estimated to be 0.2% per patient (102/total 56,208). Patients with a history of coagulopathy (OR, 1.36; 95% CI, 1.04–1.79) and transplant (OR, 1.99; 95% CI, 1.48–2.68) were associated with increased odds of TACO.Discussion: While TACO is a serious AR, events remained rare, even in populations enriched with high‐risk patients. Novel computational methods can be used to find and continually surveil for transfusion ARs. Results suggest that patients with history or presence of coagulopathy and organ transplant should be carefully monitored to mitigate potential risks of TACO.
- Open Access
- 15 May 2023
- Published by Oxford University Press (OUP) in Journal of the American Medical Informatics Association
- Vol. 30 (7) , 1323-1332
Bottom-up and top-down paradigms of artificial intelligence research approaches to healthcare data science using growing real-world big data
Objectives: As the real-world electronic health record (EHR) data continue to grow exponentially, novel methodologies involving artificial intelligence (AI) are becoming increasingly applied to enable efficient data-driven learning and, ultimately, to advance healthcare. Our objective is to provide readers with an understanding of evolving computational methods and help in deciding on methods to pursue. Target Audience: The sheer diversity of existing methods presents a challenge for health scientists who are beginning to apply computational methods to their research. Therefore, this tutorial is aimed at scientists working with EHR data who are early entrants into the field of applying AI methodologies. Scope: This manuscript describes the diverse and growing AI research approaches in healthcare data science and categorizes them into 2 distinct paradigms, the bottom-up and top-down paradigms to provide health scientists venturing into artificial intelligent research with an understanding of the evolving computational methods and help in deciding on methods to pursue through the lens of real-world healthcare data.
- 15 May 2023
- Published by Oxford University Press (OUP) in Journal of the American Medical Informatics Association
- Vol. 30 (7) , 1323-1332
Open AccessBottom-up and top-down paradigms of artificial intelligence research approaches to healthcare data science using growing real-world big data
Objectives: As the real-world electronic health record (EHR) data continue to grow exponentially, novel methodologies involving artificial intelligence (AI) are becoming increasingly applied to enable efficient data-driven learning and, ultimately, to advance healthcare. Our objective is to provide readers with an understanding of evolving computational methods and help in deciding on methods to pursue. Target Audience: The sheer diversity of existing methods presents a challenge for health scientists who are beginning to apply computational methods to their research. Therefore, this tutorial is aimed at scientists working with EHR data who are early entrants into the field of applying AI methodologies. Scope: This manuscript describes the diverse and growing AI research approaches in healthcare data science and categorizes them into 2 distinct paradigms, the bottom-up and top-down paradigms to provide health scientists venturing into artificial intelligent research with an understanding of the evolving computational methods and help in deciding on methods to pursue through the lens of real-world healthcare data.
- Open Access
- 18 January 2022
- Published by Public Library of Science (PLoS) in PLOS Digital Health
- Vol. 1 (1) , e0000008
Open challenges in developing digital therapeutics in the United States
- 18 January 2022
- Published by Public Library of Science (PLoS) in PLOS Digital Health
- Vol. 1 (1) , e0000008
Open AccessOpen challenges in developing digital therapeutics in the United States