Jiancheng Ye
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
Jiancheng Ye has a total of 153 co-authors. Here are the top co-authors by number of shared publications:
Jiarui Hai
Johns Hopkins University, Baltimore, USA
8 shared publicationsHaoxin Chen
The University of Hong Kong, Hong Kong, China
7 shared publicationsMalak Abu Hashish
Touro University College
6 shared publicationsSirui Ding
Bakar Computational Health Sciences Institute, University of California San Francisco, San Francisco, USA
6 shared publicationsSophie Bronstein
Weill Cornell Medicine, Cornell University, New York, USA
5 shared publicationsZidan Wang
Department of Statistics, Northwestern University
5 shared publications
Latest Publications
- Open Access
Biomarker-Informed Phenotyping of the Heart Failure Continuum: A Population-Based Evaluation of the Second Universal Definition
Objective: To develop and evaluate a biomarker-informed, computable phenotype of the heart failure (HF) continuum, aligned with 2026 AHA/ACC/ESC/WHF Expert Consensus, using NHANES and linked mortality data, and to determine whether this phenotype identifies subgroups with distinct mortality risks. Methods: We analyzed 11,477 U.S. adults aged ≥20 years from the 1999–2004 NHANES cycles who participated in the surplus-sera biomarker substudy and were linked to National Death Index follow-up. Participants were classified into four phenotypes according to the 2026 Universal Definition of Heart Failure: no apparent risk, Stage A (at risk), Stage B (elevated NT-proBNP without self-reported HF), and Stage C (self-reported HF). Survey-weighted Kaplan–Meier curves and nested, multiply imputed Cox proportional hazards models, with sequential adjustment for demographic, socioeconomic, and cardiometabolic covariates, were used to estimate associations with mortality. Pairwise phenotype comparisons, standardized adjusted survival estimates, and restricted mean survival times were also calculated. Incremental prognostic value was evaluated by comparing discrimination and reclassification metrics between the full phenotype and a conventional three-level model that excluded NT-proBNP. Results: Weighted prevalences were 16.9%, 70.0%, 10.7%, and 2.4% for no apparent risk, Stage A, Stage B, and Stage C, respectively. In fully adjusted models, the hazard ratios (HRs) relative to no apparent risk were 1.58 (95% CI, 1.08–2.30), 2.88 (95% CI, 2.00–4.16), and 2.56 (95% CI, 1.61–4.09), respectively. Stage B was associated with significantly higher mortality risk than Stage A (HR, 1.83; 95% CI, 1.62–2.07), whereas the difference between Stage C and Stage B was not statistically significant (HR, 0.89; 95% CI, 0.71–1.11). The association for Stage B attenuated less across sequentially adjusted models than that for Stage C. Adjusted 10-year mortality risks were 15.2% for Stage B and 14.2% for Stage C. Incorporating Stage B into the conventional model yielded only marginal improvements in discrimination (ΔC-statistic, 0.002; 95% CI, 0.001– 0.002). Conclusion: A biomarker-informed phenotype of the HF continuum can be operationalized using national survey data and identifies a pre-HF subgroup with an adjusted mortality risk not significantly different from that of individuals with overt HF, despite only marginal gains in prognostic discrimination.
Open AccessBiomarker-Informed Phenotyping of the Heart Failure Continuum: A Population-Based Evaluation of the Second Universal Definition
Objective: To develop and evaluate a biomarker-informed, computable phenotype of the heart failure (HF) continuum, aligned with 2026 AHA/ACC/ESC/WHF Expert Consensus, using NHANES and linked mortality data, and to determine whether this phenotype identifies subgroups with distinct mortality risks. Methods: We analyzed 11,477 U.S. adults aged ≥20 years from the 1999–2004 NHANES cycles who participated in the surplus-sera biomarker substudy and were linked to National Death Index follow-up. Participants were classified into four phenotypes according to the 2026 Universal Definition of Heart Failure: no apparent risk, Stage A (at risk), Stage B (elevated NT-proBNP without self-reported HF), and Stage C (self-reported HF). Survey-weighted Kaplan–Meier curves and nested, multiply imputed Cox proportional hazards models, with sequential adjustment for demographic, socioeconomic, and cardiometabolic covariates, were used to estimate associations with mortality. Pairwise phenotype comparisons, standardized adjusted survival estimates, and restricted mean survival times were also calculated. Incremental prognostic value was evaluated by comparing discrimination and reclassification metrics between the full phenotype and a conventional three-level model that excluded NT-proBNP. Results: Weighted prevalences were 16.9%, 70.0%, 10.7%, and 2.4% for no apparent risk, Stage A, Stage B, and Stage C, respectively. In fully adjusted models, the hazard ratios (HRs) relative to no apparent risk were 1.58 (95% CI, 1.08–2.30), 2.88 (95% CI, 2.00–4.16), and 2.56 (95% CI, 1.61–4.09), respectively. Stage B was associated with significantly higher mortality risk than Stage A (HR, 1.83; 95% CI, 1.62–2.07), whereas the difference between Stage C and Stage B was not statistically significant (HR, 0.89; 95% CI, 0.71–1.11). The association for Stage B attenuated less across sequentially adjusted models than that for Stage C. Adjusted 10-year mortality risks were 15.2% for Stage B and 14.2% for Stage C. Incorporating Stage B into the conventional model yielded only marginal improvements in discrimination (ΔC-statistic, 0.002; 95% CI, 0.001– 0.002). Conclusion: A biomarker-informed phenotype of the HF continuum can be operationalized using national survey data and identifies a pre-HF subgroup with an adjusted mortality risk not significantly different from that of individuals with overt HF, despite only marginal gains in prognostic discrimination.
- Open Access
Predicting the Absence and Gaps in Recorded Blood Pressure Follow-up Among Patients With Hypertension
Background: Electronic health record (EHR) data can help identify patients who would benefit from blood pressure reassessment after an elevated reading. Objective: To develop and internally validate a model to predict the absence of a recorded blood pressure measurement after an elevated reading among adults with hypertension; evaluate the potential yield of targeted outreach and model performance across subgroups; and determine whether recorded contact patterns provide additional information for predicting subsequent cardiovascular hospitalization. Methods: This retrospective cohort study used MIMIC-IV data. Adults with hypertension, a qualifying systolic blood pressure reading >140 mmHg, defined as the first such reading on or after index, and 90 days of subsequent database coverage and survival were included. The primary outcome was the absence of a recorded blood pressure measurement within 90 days of the index reading. Gradient-boosted trees incorporating 62 demographic, clinical, and contact-history features were compared with logistic regression models using a random 6:2:2 training, validation, and test split, with isotonic recalibration. Nested cross-validation was used to assess the tree model. Model discrimination, calibration, simulated outreach yield, and subgroup performance were evaluated. Results: Among 30,942 patients, 10,899 (35.2%) had no recorded blood pressure measurement within 90 days. In the test set (n = 6,189), the gradient-boosted tree model achieved an area under the receiver operating characteristic curve (AUROC) of 0.740 (95% CI, 0.727-0.753), compared with 0.698 for penalized logistic regression and 0.687 for two-variable logistic regression. The calibration slope was 0.91, and the area under the precision-recall curve was 0.568 (95% CI, 0.546- 0.593). Selecting the 20% of patients with the highest predicted risk identified 34.2% (95% CI, 32.7-35.7) of patients without a recorded measurement, with a positive predictive value of 60.4% (95% CI, 57.4-63.2). In out-of-fold subgroup analyses using a common selection threshold corresponding to an overall selection rate of 20%, sensitivity was 21.5% among Black patients, 23.3% among Hispanic patients, and 37.2% among White patients. Among 15,200 participants in the landmark analysis, 664 cardiovascular hospitalizations occurred. Increasing and sustained contact patterns were associated with modestly higher hazards of cardiovascular hospitalization, with the association primarily attributable to heart failure hospitalization. For sustained versus minimal contact, the hazard ratio for the composite outcome was 1.40 (95% CI, 1.06-1.85). However, adding contact gaps and trajectories improved the C-statistic by only 0.0023 (95% CI, -0.0006 to 0.0056). Conclusions: The model showed potential for prioritizing blood pressure reassessment outreach, although sensitivity varied across racial and ethnic groups. Recorded contact patterns provided little incremental discrimination for cardiovascular hospitalization.
Open AccessPredicting the Absence and Gaps in Recorded Blood Pressure Follow-up Among Patients With Hypertension
Background: Electronic health record (EHR) data can help identify patients who would benefit from blood pressure reassessment after an elevated reading. Objective: To develop and internally validate a model to predict the absence of a recorded blood pressure measurement after an elevated reading among adults with hypertension; evaluate the potential yield of targeted outreach and model performance across subgroups; and determine whether recorded contact patterns provide additional information for predicting subsequent cardiovascular hospitalization. Methods: This retrospective cohort study used MIMIC-IV data. Adults with hypertension, a qualifying systolic blood pressure reading >140 mmHg, defined as the first such reading on or after index, and 90 days of subsequent database coverage and survival were included. The primary outcome was the absence of a recorded blood pressure measurement within 90 days of the index reading. Gradient-boosted trees incorporating 62 demographic, clinical, and contact-history features were compared with logistic regression models using a random 6:2:2 training, validation, and test split, with isotonic recalibration. Nested cross-validation was used to assess the tree model. Model discrimination, calibration, simulated outreach yield, and subgroup performance were evaluated. Results: Among 30,942 patients, 10,899 (35.2%) had no recorded blood pressure measurement within 90 days. In the test set (n = 6,189), the gradient-boosted tree model achieved an area under the receiver operating characteristic curve (AUROC) of 0.740 (95% CI, 0.727-0.753), compared with 0.698 for penalized logistic regression and 0.687 for two-variable logistic regression. The calibration slope was 0.91, and the area under the precision-recall curve was 0.568 (95% CI, 0.546- 0.593). Selecting the 20% of patients with the highest predicted risk identified 34.2% (95% CI, 32.7-35.7) of patients without a recorded measurement, with a positive predictive value of 60.4% (95% CI, 57.4-63.2). In out-of-fold subgroup analyses using a common selection threshold corresponding to an overall selection rate of 20%, sensitivity was 21.5% among Black patients, 23.3% among Hispanic patients, and 37.2% among White patients. Among 15,200 participants in the landmark analysis, 664 cardiovascular hospitalizations occurred. Increasing and sustained contact patterns were associated with modestly higher hazards of cardiovascular hospitalization, with the association primarily attributable to heart failure hospitalization. For sustained versus minimal contact, the hazard ratio for the composite outcome was 1.40 (95% CI, 1.06-1.85). However, adding contact gaps and trajectories improved the C-statistic by only 0.0023 (95% CI, -0.0006 to 0.0056). Conclusions: The model showed potential for prioritizing blood pressure reassessment outreach, although sensitivity varied across racial and ethnic groups. Recorded contact patterns provided little incremental discrimination for cardiovascular hospitalization.
- Open Access
- 17 October 2026
- Published by Elsevier in SSRN Electronic Journal
Evaluation of Large Language Models to Characterize Reasons for Treatment Discontinuation from Unstructured Clinical Notes in Patients with Gastrointestinal Cancers: A Feasibility Study
BACKGROUND: Treatment discontinuation, which may result from regimen-related adverse events, disease progression, or death, is an important indicator of safety and effectiveness in real-world gastrointestinal (GI) cancer care delivery. Manual review of clinical notes is currently required to accurately identify reasons for discontinuation, which is time-consuming, burdensome, and expensive. Advances in generative artificial intelligence (AI) may help to overcome these barriers while preserving high clinical accuracy. This pilot study assesses the feasibility of using large language models (LLMs) to characterize reasons for discontinuing systemic therapy from clinical notes in GI oncology. METHODS: We included 297 patients with GI cancers who discontinued systemic therapy within six months of initiation in the Mass General Brigham health system between 01/2019 and 01/2024. Clinical notes documented before and after three months of discontinuation were retrieved. The cohort was divided into development (n=210, 37,804 notes) and testing (n=87, 15,621 notes) sets. Manual chart review based on clinical experience established a ground truth for the following discontinuation reasons: death, disease progression, and treatment intolerance. Zero-shot, few-shot, and error-analysis-based prompting strategies were applied via Microsoft Azure OpenAI Service using ChatGPT-4.1 and Claude-4.6 to analyze clinical notes within three months of discontinuation. We developed models to output discontinuation reason and the supporting evidence used to make that determination. Model performance was evaluated using accuracy, Cohen’s kappa, and sensitivity for each discontinuation reason. RESULTS: Among the 297 patients, progression was the most common reason for discontinuation (47.5%), followed by death (38.0%) and intolerance (14.5%). Approximately 59% of patients had stage IV cancers, and the majority were receiving treatment with palliative intent (85.5%). In the testing set, GPT-4.1 achieved 92.0% overall accuracy and 0.87 Cohen’s kappa, whereas Claude-4.6 achieved 89.7% accuracy and 0.83 Cohen’s kappa. For GPT-4.1, sensitivity was 97.0% for death, 91.7% for treatment intolerance, and 88.1% for progression. For Claude-4.6, sensitivity was 90.9% for death, 91.7% for treatment intolerance, and 88.1% for progression. Discussion: Strong alignment between manual chart review and LLMs supports the potential for generative AI to transform labor-intensive chart abstraction into scalable, reproducible quality measurement, enabling rapid surveillance of treatment safety, effectiveness, and care delivery.
- 17 October 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessEvaluation of Large Language Models to Characterize Reasons for Treatment Discontinuation from Unstructured Clinical Notes in Patients with Gastrointestinal Cancers: A Feasibility Study
BACKGROUND: Treatment discontinuation, which may result from regimen-related adverse events, disease progression, or death, is an important indicator of safety and effectiveness in real-world gastrointestinal (GI) cancer care delivery. Manual review of clinical notes is currently required to accurately identify reasons for discontinuation, which is time-consuming, burdensome, and expensive. Advances in generative artificial intelligence (AI) may help to overcome these barriers while preserving high clinical accuracy. This pilot study assesses the feasibility of using large language models (LLMs) to characterize reasons for discontinuing systemic therapy from clinical notes in GI oncology. METHODS: We included 297 patients with GI cancers who discontinued systemic therapy within six months of initiation in the Mass General Brigham health system between 01/2019 and 01/2024. Clinical notes documented before and after three months of discontinuation were retrieved. The cohort was divided into development (n=210, 37,804 notes) and testing (n=87, 15,621 notes) sets. Manual chart review based on clinical experience established a ground truth for the following discontinuation reasons: death, disease progression, and treatment intolerance. Zero-shot, few-shot, and error-analysis-based prompting strategies were applied via Microsoft Azure OpenAI Service using ChatGPT-4.1 and Claude-4.6 to analyze clinical notes within three months of discontinuation. We developed models to output discontinuation reason and the supporting evidence used to make that determination. Model performance was evaluated using accuracy, Cohen’s kappa, and sensitivity for each discontinuation reason. RESULTS: Among the 297 patients, progression was the most common reason for discontinuation (47.5%), followed by death (38.0%) and intolerance (14.5%). Approximately 59% of patients had stage IV cancers, and the majority were receiving treatment with palliative intent (85.5%). In the testing set, GPT-4.1 achieved 92.0% overall accuracy and 0.87 Cohen’s kappa, whereas Claude-4.6 achieved 89.7% accuracy and 0.83 Cohen’s kappa. For GPT-4.1, sensitivity was 97.0% for death, 91.7% for treatment intolerance, and 88.1% for progression. For Claude-4.6, sensitivity was 90.9% for death, 91.7% for treatment intolerance, and 88.1% for progression. Discussion: Strong alignment between manual chart review and LLMs supports the potential for generative AI to transform labor-intensive chart abstraction into scalable, reproducible quality measurement, enabling rapid surveillance of treatment safety, effectiveness, and care delivery.
- Open Access
- 25 September 2026
- Published by Springer Nature in BMC Medicine
Learning health systems for cardiovascular health and health care
Cardiovascular diseases (CVDs) remain the leading cause of morbidity and mortality worldwide, yet the translation of established prevention and treatment strategies into routine clinical practice remains inconsistent. This review examines the application of Learning Health System (LHS) principles to cardiovascular care, synthesizing contemporary evidence on artificial intelligence (AI) and machine learning (ML), clinical decision support (CDS), wearable and remote patient monitoring, interoperable data infrastructures and common data models, patient-generated health data (PGHD), patient-reported outcomes (PROs), implementation science, community engagement, and the governance of AI-enabled technologies. Cardiovascular health and care represent promising but complex contexts for LHS implementation, characterized by time-sensitive acute events, longitudinal disease management, guideline-directed therapies, increasingly extensive clinical and physiological data, costly interventions, well-established registries, and persistent disparities in access, quality, and outcomes. Opportunities exist to strengthen evidence generation, clinical decision-making, and continuous quality improvement while addressing major implementation challenges, including fragmented multipayer systems, misaligned incentives, substantial upfront infrastructure requirements, heterogeneous data quality, algorithmic bias and model drift, clinician workload, regulatory and liability uncertainties, and inequitable access to digital technologies. Case studies from the Veterans Health Administration, PCORnet and PCORI initiatives, Kaiser Permanente, and the American Heart Association’s Get With The Guidelines program demonstrate the feasibility of key LHS functions while also underscoring contextual factors that may limit their generalizability and scalability. Cardiovascular LHSs have the potential to accelerate the translation of evidence into practice, but their effectiveness depends on the integration of robust technical infrastructure with effective governance, sustainable financing, systematic measurement of health equity, continuous monitoring of algorithmic performance, pragmatic implementation strategies, and rigorous evaluation of patient-centered outcomes across diverse and fragmented care environments.
- 25 September 2026
- Published by Springer Nature in BMC Medicine
Open AccessLearning health systems for cardiovascular health and health care
Cardiovascular diseases (CVDs) remain the leading cause of morbidity and mortality worldwide, yet the translation of established prevention and treatment strategies into routine clinical practice remains inconsistent. This review examines the application of Learning Health System (LHS) principles to cardiovascular care, synthesizing contemporary evidence on artificial intelligence (AI) and machine learning (ML), clinical decision support (CDS), wearable and remote patient monitoring, interoperable data infrastructures and common data models, patient-generated health data (PGHD), patient-reported outcomes (PROs), implementation science, community engagement, and the governance of AI-enabled technologies. Cardiovascular health and care represent promising but complex contexts for LHS implementation, characterized by time-sensitive acute events, longitudinal disease management, guideline-directed therapies, increasingly extensive clinical and physiological data, costly interventions, well-established registries, and persistent disparities in access, quality, and outcomes. Opportunities exist to strengthen evidence generation, clinical decision-making, and continuous quality improvement while addressing major implementation challenges, including fragmented multipayer systems, misaligned incentives, substantial upfront infrastructure requirements, heterogeneous data quality, algorithmic bias and model drift, clinician workload, regulatory and liability uncertainties, and inequitable access to digital technologies. Case studies from the Veterans Health Administration, PCORnet and PCORI initiatives, Kaiser Permanente, and the American Heart Association’s Get With The Guidelines program demonstrate the feasibility of key LHS functions while also underscoring contextual factors that may limit their generalizability and scalability. Cardiovascular LHSs have the potential to accelerate the translation of evidence into practice, but their effectiveness depends on the integration of robust technical infrastructure with effective governance, sustainable financing, systematic measurement of health equity, continuous monitoring of algorithmic performance, pragmatic implementation strategies, and rigorous evaluation of patient-centered outcomes across diverse and fragmented care environments.
- Open Access
- 19 August 2026
- Published by Frontiers Media SA in Frontiers in Climate
- Vol. 8, 1703142
Climate change and hypertension: global trends, temporal patterns, and regional insights
Objective: Climate change has emerged as a potential cardiovascular health risk factor. This study describes temporal and geographic patterns in temperature anomalies alongside global trends in hypertension prevalence, treatment, and control. Methods: We conducted a descriptive observational synthesis using temperature anomaly data from the National Centers for Environmental Information spanning 1850–2023 and global hypertension data from the NCD Risk Factor Collaboration. Temperature anomalies were analyzed at continental and global scales. Hypertension outcomes included prevalence, mean systolic and diastolic blood pressure (BP), diagnosis rates, treatment rates, and control rates. Data were stratified by geographic region and sex to identify differential patterns. Results: Global temperature anomalies increased dramatically from predominantly negative values before 1980 to consistently positive values thereafter, reaching 1.43 °C by 2023. Regional warming was greatest in Asia (peak 1.73 °C in 2015), Europe (1.67 °C in 2015; peak 1.89 °C in 2007), and North America (1.34 °C in 2015; peak 1.67 °C, recorded in 1999, 2010, and 2012). During the same period in which temperature anomalies intensified, global hypertension management improved substantially. Women consistently had more favorable hypertension indicators across all measures, with raised BP prevalence declining from 26.1% to 20.1% (1975–2015), compared to men's decline from 29.5% to 24.1%. High-income regions, particularly Europe and North America, achieved significant improvements in hypertension outcomes. Conversely, many African countries showed concerning trends with rising hypertension prevalence coinciding with elevated temperatures and limited healthcare infrastructure. Discussion and conclusion: The observed population-level patterns linking climate change and hypertension are complex and appear strongly shaped by socioeconomic factors, healthcare infrastructure, and demographic characteristics. Despite intensifying temperature anomalies globally, observed population-level hypertension trends varied substantially across regions, with apparent improvements concentrated in higher-income settings with stronger healthcare infrastructure and persistent or rising prevalence in many lower-income settings. Whether, and to what extent, these contrasts reflect climate adaptation through healthcare and economic resources, rather than other concurrent drivers, cannot be determined from the descriptive design of the present study and warrants dedicated inferential investigation. Resource-constrained areas may face compounding risks from climate change and limited healthcare infrastructure. Sex-specific differences suggest that targeted interventions may be necessary to address differential vulnerabilities. Future climate-health adaptation strategies should prioritize healthcare infrastructure strengthening, particularly in vulnerable regions experiencing both significant warming and limited adaptive capacity.
- 19 August 2026
- Published by Frontiers Media SA in Frontiers in Climate
- Vol. 8, 1703142
Open AccessClimate change and hypertension: global trends, temporal patterns, and regional insights
Objective: Climate change has emerged as a potential cardiovascular health risk factor. This study describes temporal and geographic patterns in temperature anomalies alongside global trends in hypertension prevalence, treatment, and control. Methods: We conducted a descriptive observational synthesis using temperature anomaly data from the National Centers for Environmental Information spanning 1850–2023 and global hypertension data from the NCD Risk Factor Collaboration. Temperature anomalies were analyzed at continental and global scales. Hypertension outcomes included prevalence, mean systolic and diastolic blood pressure (BP), diagnosis rates, treatment rates, and control rates. Data were stratified by geographic region and sex to identify differential patterns. Results: Global temperature anomalies increased dramatically from predominantly negative values before 1980 to consistently positive values thereafter, reaching 1.43 °C by 2023. Regional warming was greatest in Asia (peak 1.73 °C in 2015), Europe (1.67 °C in 2015; peak 1.89 °C in 2007), and North America (1.34 °C in 2015; peak 1.67 °C, recorded in 1999, 2010, and 2012). During the same period in which temperature anomalies intensified, global hypertension management improved substantially. Women consistently had more favorable hypertension indicators across all measures, with raised BP prevalence declining from 26.1% to 20.1% (1975–2015), compared to men's decline from 29.5% to 24.1%. High-income regions, particularly Europe and North America, achieved significant improvements in hypertension outcomes. Conversely, many African countries showed concerning trends with rising hypertension prevalence coinciding with elevated temperatures and limited healthcare infrastructure. Discussion and conclusion: The observed population-level patterns linking climate change and hypertension are complex and appear strongly shaped by socioeconomic factors, healthcare infrastructure, and demographic characteristics. Despite intensifying temperature anomalies globally, observed population-level hypertension trends varied substantially across regions, with apparent improvements concentrated in higher-income settings with stronger healthcare infrastructure and persistent or rising prevalence in many lower-income settings. Whether, and to what extent, these contrasts reflect climate adaptation through healthcare and economic resources, rather than other concurrent drivers, cannot be determined from the descriptive design of the present study and warrants dedicated inferential investigation. Resource-constrained areas may face compounding risks from climate change and limited healthcare infrastructure. Sex-specific differences suggest that targeted interventions may be necessary to address differential vulnerabilities. Future climate-health adaptation strategies should prioritize healthcare infrastructure strengthening, particularly in vulnerable regions experiencing both significant warming and limited adaptive capacity.
- Open Access
- 29 July 2026
- Published by Springer Nature in BMC Medical Imaging
- Vol. 26 (1)
Medical Segment Anything Model (MedSAM): methodological innovations, evaluation practices, and clinical applications
Background: The Medical Segment Anything Model (MedSAM) has emerged as a prominent foundation model for medical image segmentation, prompting rapid growth in related research. Despite this proliferation, the literature remains fragmented across multimodal paradigms and lacks rigorous synthesis of methodological innovations, clinical safety evidence, algorithmic bias, and real-world implementation barriers—gaps that limit the field’s ability to translate algorithmic progress into clinical impact. Objective: To systematically identify, classify, and synthesize published research on MedSAM and its derivatives, providing a comprehensive overview of methodological developments, evaluation findings, and clinical translation progress. Methods: This scoping review followed the PRISMA-ScR guideline. We conducted a systematic search across six major databases (PubMed, Embase, Web of Science, Scopus, IEEE Xplore, ACM Digital Library) through June 2025. Results: From 1,482 initial records, 45 studies met inclusion criteria. Innovation-oriented studies ( n = 32) were classified along four methodological axes spanning multimodal paradigms: Architecture Enhancement (encoder/decoder refinements for efficiency and precision), Prompt Strategy (textual, spatial, and multimodal prompting mechanisms), Efficiency and Deployability (quantization, distillation, and system integration), and Pipeline and Application (integration into clinical workflows). Evaluation-oriented studies ( n = 13) revealed that MedSAM outperforms general-purpose SAM by 5–20% on structured modalities (CT/MRI) but lags behind task-specific networks for delicate structures. Notably, fairness analyses from two studies (combined n = 1,757 subjects) report Dice disparities of up to 10% across sex, age, and BMI strata, raising concerns about potential demographic bias that warrants further investigation in larger, multi-center cohorts before firm clinical-safety conclusions can be drawn. Real-world deployment is further constrained by prompt sensitivity (5–15% Dice variation), cross-modality domain shift, regulatory compliance gaps (FDA/CE marking, HIPAA, GDPR), and absent multi-center prospective validation, all of which represent principal implementation hurdles for clinical adoption. Conclusions: MedSAM represents a significant advancement in medical image segmentation, demonstrating a clear trajectory from large-scale architectures toward lightweight, prompt-aware, and deployable models. This synthesis provides a structured critical appraisal indicating that the evidence base remains at the proof-of-concept stage, with most studies retrospective and single-center, and emerging signals of algorithmic bias, unresolved cross-modality failure modes, and incomplete regulatory pathways represent safety and implementation considerations that should be addressed before multimodal MedSAM-based systems are more broadly deployed in clinical imaging workflows.
- 29 July 2026
- Published by Springer Nature in BMC Medical Imaging
- Vol. 26 (1)
Open AccessMedical Segment Anything Model (MedSAM): methodological innovations, evaluation practices, and clinical applications
Background: The Medical Segment Anything Model (MedSAM) has emerged as a prominent foundation model for medical image segmentation, prompting rapid growth in related research. Despite this proliferation, the literature remains fragmented across multimodal paradigms and lacks rigorous synthesis of methodological innovations, clinical safety evidence, algorithmic bias, and real-world implementation barriers—gaps that limit the field’s ability to translate algorithmic progress into clinical impact. Objective: To systematically identify, classify, and synthesize published research on MedSAM and its derivatives, providing a comprehensive overview of methodological developments, evaluation findings, and clinical translation progress. Methods: This scoping review followed the PRISMA-ScR guideline. We conducted a systematic search across six major databases (PubMed, Embase, Web of Science, Scopus, IEEE Xplore, ACM Digital Library) through June 2025. Results: From 1,482 initial records, 45 studies met inclusion criteria. Innovation-oriented studies ( n = 32) were classified along four methodological axes spanning multimodal paradigms: Architecture Enhancement (encoder/decoder refinements for efficiency and precision), Prompt Strategy (textual, spatial, and multimodal prompting mechanisms), Efficiency and Deployability (quantization, distillation, and system integration), and Pipeline and Application (integration into clinical workflows). Evaluation-oriented studies ( n = 13) revealed that MedSAM outperforms general-purpose SAM by 5–20% on structured modalities (CT/MRI) but lags behind task-specific networks for delicate structures. Notably, fairness analyses from two studies (combined n = 1,757 subjects) report Dice disparities of up to 10% across sex, age, and BMI strata, raising concerns about potential demographic bias that warrants further investigation in larger, multi-center cohorts before firm clinical-safety conclusions can be drawn. Real-world deployment is further constrained by prompt sensitivity (5–15% Dice variation), cross-modality domain shift, regulatory compliance gaps (FDA/CE marking, HIPAA, GDPR), and absent multi-center prospective validation, all of which represent principal implementation hurdles for clinical adoption. Conclusions: MedSAM represents a significant advancement in medical image segmentation, demonstrating a clear trajectory from large-scale architectures toward lightweight, prompt-aware, and deployable models. This synthesis provides a structured critical appraisal indicating that the evidence base remains at the proof-of-concept stage, with most studies retrospective and single-center, and emerging signals of algorithmic bias, unresolved cross-modality failure modes, and incomplete regulatory pathways represent safety and implementation considerations that should be addressed before multimodal MedSAM-based systems are more broadly deployed in clinical imaging workflows.
- Open Access
Latent Class Trajectory Phenotypes of Longitudinal Visit and Follow-up Patterns Among Patients with Hypertension
Objective: To identify latent phenotypes of observed electronic health record (EHR) contact among patients with hypertension, evaluate their reproducibility across temporal resolutions and sensitivity to administrative censoring, and examine associations with treatment documentation and blood pressure (BP) control after adjustment for comorbidity burden and measurement opportunity. Methods: We conducted a retrospective cohort study of patients with hypertension and at least two recorded visits (N = 26,710). Latent class trajectory modelling of repeated binary visit indicators within a Bernoulli finite mixture framework was performed using 1-month and 3-month intervals over 36 months. Models were selected using the Bayesian Information Criterion and GRoLTS-recommended metrics. Multivariable linear regression examined associations between trajectory phenotypes and treatment documentation and BP control after adjustment for age, sex, race, baseline BP, and comorbidity burden. BP outcomes were also compared at 12, 24, and 36 months. Results: Four observed EHR contact trajectory groups were identified at both temporal resolutions (1-month: relative entropy 0.821, minimum APP 0.770, minimum OCC 6.49; 3-month: relative entropy 0.771, minimum APP 0.778, minimum OCC 5.39), with excellent bootstrap reproducibility (mean ARI 0.966 - 0.969). The four-group structure was replicated in the ≥ 24-month subgroup, although this represented only 20.8% of the cohort. Overall, 79.2% of patients had <24 months of follow-up, with censoring concentrated in the lowest-contact groups, indicating that these trajectories reflect observed EHR contact under variable administrative observation rather than patient disengagement. After adjustment, higher-contact groups had consistently higher treatment documentation rates than the low observed follow-up group, whereas differences in BP control were small and inconsistent. The highest-contact group did not achieve the lowest BP at any fixed time point despite the largest first-to-last BP reduction, demonstrating bias from differential observation window length. Conclusion: Four internally reproducible phenotypes of observed EHR contact were identified, but their trajectories were substantially influenced by administrative censoring and lack external validation. Contact phenotype was associated with treatment documentation but only weakly with BP control after adjustment.
Open AccessLatent Class Trajectory Phenotypes of Longitudinal Visit and Follow-up Patterns Among Patients with Hypertension
Objective: To identify latent phenotypes of observed electronic health record (EHR) contact among patients with hypertension, evaluate their reproducibility across temporal resolutions and sensitivity to administrative censoring, and examine associations with treatment documentation and blood pressure (BP) control after adjustment for comorbidity burden and measurement opportunity. Methods: We conducted a retrospective cohort study of patients with hypertension and at least two recorded visits (N = 26,710). Latent class trajectory modelling of repeated binary visit indicators within a Bernoulli finite mixture framework was performed using 1-month and 3-month intervals over 36 months. Models were selected using the Bayesian Information Criterion and GRoLTS-recommended metrics. Multivariable linear regression examined associations between trajectory phenotypes and treatment documentation and BP control after adjustment for age, sex, race, baseline BP, and comorbidity burden. BP outcomes were also compared at 12, 24, and 36 months. Results: Four observed EHR contact trajectory groups were identified at both temporal resolutions (1-month: relative entropy 0.821, minimum APP 0.770, minimum OCC 6.49; 3-month: relative entropy 0.771, minimum APP 0.778, minimum OCC 5.39), with excellent bootstrap reproducibility (mean ARI 0.966 - 0.969). The four-group structure was replicated in the ≥ 24-month subgroup, although this represented only 20.8% of the cohort. Overall, 79.2% of patients had <24 months of follow-up, with censoring concentrated in the lowest-contact groups, indicating that these trajectories reflect observed EHR contact under variable administrative observation rather than patient disengagement. After adjustment, higher-contact groups had consistently higher treatment documentation rates than the low observed follow-up group, whereas differences in BP control were small and inconsistent. The highest-contact group did not achieve the lowest BP at any fixed time point despite the largest first-to-last BP reduction, demonstrating bias from differential observation window length. Conclusion: Four internally reproducible phenotypes of observed EHR contact were identified, but their trajectories were substantially influenced by administrative censoring and lack external validation. Contact phenotype was associated with treatment documentation but only weakly with BP control after adjustment.
- Open Access
Temporal Patterns and Association Between FIFA World Cup Tournament Periods and Emergency Department–Treated Injuries
Background: Major soccer tournaments may temporarily change recreational soccer activity, community gatherings, and injury-prevention needs, but evidence for population-level emergency department (ED) injury patterns during these events is limited. Understanding whether ED-treated soccer injury burden changes during Men’s FIFA World Cup periods may help inform surveillance readiness and prevention planning for future tournaments. Objective: To evaluate whether Men’s FIFA World Cup tournament periods temporally coincided with changes in ED-treated soccer-coded injury burden in the United States and to assess the implications for public health surveillance and injury-prevention preparedness. Methods: We conducted a retrospective, repeated cross-sectional calendar-period analysis of publicly available national ED injury surveillance records from 1999 through 2025. Soccer-coded injuries were identified using product code 1267 in any available product field. The primary exposure was the set of official Men’s FIFA World Cup tournament dates from 2002, 2006, 2010, 2014, 2018, and 2022. Tournament dates were compared with matched same-calendar dates in adjacent years, excluding dates that overlapped other FIFA World Cup tournament windows. The primary estimands were the mean daily difference and ratio in weighted national ED-treated soccer-coded injury estimates between tournament and matched-control periods. Results: The analytic cohort included 170,679 soccer-coded ED cases, corresponding to an estimated 5,366,681 ED-treated soccer-coded injuries nationally. Mean daily weighted estimates were 453.1 during Men’s World Cup tournament dates and 384.1 during matched control dates. The absolute mean daily difference was 68.9 injuries per day (95% CI, -0.5 to 138.3), and the mean daily ratio was 1.18 (95% CI, 1.00 to 1.39). Tournament-specific estimates were heterogeneous, with a near-null estimate for the 2022 winter tournament and higher estimates for prior summer tournaments. Conclusions: Mean daily injury estimates were modestly higher during tournament periods, although the confidence interval for the absolute difference included zero and estimates varied across tournaments. The findings suggest that major soccer events offer valuable opportunities for real-time ED injury surveillance, targeted recreational soccer injury-prevention messaging, concussion awareness, and coordinated public health preparedness for community- and fan-related injury patterns during future tournaments.
Open AccessTemporal Patterns and Association Between FIFA World Cup Tournament Periods and Emergency Department–Treated Injuries
Background: Major soccer tournaments may temporarily change recreational soccer activity, community gatherings, and injury-prevention needs, but evidence for population-level emergency department (ED) injury patterns during these events is limited. Understanding whether ED-treated soccer injury burden changes during Men’s FIFA World Cup periods may help inform surveillance readiness and prevention planning for future tournaments. Objective: To evaluate whether Men’s FIFA World Cup tournament periods temporally coincided with changes in ED-treated soccer-coded injury burden in the United States and to assess the implications for public health surveillance and injury-prevention preparedness. Methods: We conducted a retrospective, repeated cross-sectional calendar-period analysis of publicly available national ED injury surveillance records from 1999 through 2025. Soccer-coded injuries were identified using product code 1267 in any available product field. The primary exposure was the set of official Men’s FIFA World Cup tournament dates from 2002, 2006, 2010, 2014, 2018, and 2022. Tournament dates were compared with matched same-calendar dates in adjacent years, excluding dates that overlapped other FIFA World Cup tournament windows. The primary estimands were the mean daily difference and ratio in weighted national ED-treated soccer-coded injury estimates between tournament and matched-control periods. Results: The analytic cohort included 170,679 soccer-coded ED cases, corresponding to an estimated 5,366,681 ED-treated soccer-coded injuries nationally. Mean daily weighted estimates were 453.1 during Men’s World Cup tournament dates and 384.1 during matched control dates. The absolute mean daily difference was 68.9 injuries per day (95% CI, -0.5 to 138.3), and the mean daily ratio was 1.18 (95% CI, 1.00 to 1.39). Tournament-specific estimates were heterogeneous, with a near-null estimate for the 2022 winter tournament and higher estimates for prior summer tournaments. Conclusions: Mean daily injury estimates were modestly higher during tournament periods, although the confidence interval for the absolute difference included zero and estimates varied across tournaments. The findings suggest that major soccer events offer valuable opportunities for real-time ED injury surveillance, targeted recreational soccer injury-prevention messaging, concussion awareness, and coordinated public health preparedness for community- and fan-related injury patterns during future tournaments.
- 13 June 2026
- Published by Oxford University Press (OUP) in Journal of the American Medical Informatics Association
- Vol. 33 (9) , 1822-1833
Large language model biases in health care: a scoping review and call for an integrated assessment framework
Objectives: To conduct a scoping review of bias assessment in studies applying large language models (LLMs) to health data and to synthesize their prevailing conceptualization of bias. Materials and Methods: Following PRISMA guidelines, we queried PubMed and Scopus. Two annotators screened titles, abstracts, and full texts for eligibility, calibrating their assessments throughout the process. For included studies, we extracted and summarized data on LLMs (name and version, development domain, open- or closed-sourced status, and commercial or academic origin), natural language processing tasks (task formulation, gold-standard dataset, evaluation metrics, prompting or fine-tuning strategies), and biases (type, assessment, and bias summary). Results: Of the 1585 records retrieved, 76 papers met the eligibility criteria for full review. Among these, 59 reported identifying bias. Three major conceptualizations of bias emerged: behavioral output bias (nonstereotyping and stereotyping), predictive outcome bias, and representational bias. Studies generally adopted an observational approach (measuring bias using the existing dataset) or an experimental approach (altering prompts, eg, with different demographic information, and comparing outputs). Discussion and Conclusion: Behavioral output bias and predictive outcome bias, both of which emphasize parity, dominate existing studies. Whether evaluated against external accuracy or internal equality benchmarks, these approaches often assume that equal performance across groups is inherently desirable. Treating all disparities as bias risks conflating poor model behavior with real-world disparities, and researchers should remain aware of potential tradeoffs between parity and accuracy objectives. We introduce an integrated framework that combines parity and accuracy benchmarks and encourages transparent, context-aware interpretation of group differences.
- 13 June 2026
- Published by Oxford University Press (OUP) in Journal of the American Medical Informatics Association
- Vol. 33 (9) , 1822-1833
Large language model biases in health care: a scoping review and call for an integrated assessment framework
Objectives: To conduct a scoping review of bias assessment in studies applying large language models (LLMs) to health data and to synthesize their prevailing conceptualization of bias. Materials and Methods: Following PRISMA guidelines, we queried PubMed and Scopus. Two annotators screened titles, abstracts, and full texts for eligibility, calibrating their assessments throughout the process. For included studies, we extracted and summarized data on LLMs (name and version, development domain, open- or closed-sourced status, and commercial or academic origin), natural language processing tasks (task formulation, gold-standard dataset, evaluation metrics, prompting or fine-tuning strategies), and biases (type, assessment, and bias summary). Results: Of the 1585 records retrieved, 76 papers met the eligibility criteria for full review. Among these, 59 reported identifying bias. Three major conceptualizations of bias emerged: behavioral output bias (nonstereotyping and stereotyping), predictive outcome bias, and representational bias. Studies generally adopted an observational approach (measuring bias using the existing dataset) or an experimental approach (altering prompts, eg, with different demographic information, and comparing outputs). Discussion and Conclusion: Behavioral output bias and predictive outcome bias, both of which emphasize parity, dominate existing studies. Whether evaluated against external accuracy or internal equality benchmarks, these approaches often assume that equal performance across groups is inherently desirable. Treating all disparities as bias risks conflating poor model behavior with real-world disparities, and researchers should remain aware of potential tradeoffs between parity and accuracy objectives. We introduce an integrated framework that combines parity and accuracy benchmarks and encourages transparent, context-aware interpretation of group differences.
- Open Access
Retention and loss to follow-up among patients with hypertension in primary care: a multi-practice cohort study
Effective hypertension management depends on sustained engagement with primary care, and there is a need to understand the magnitude and determinants of follow-up loss in real-world primary care. We analyzed electronic health record (EHR) data from 26,541 patients with hypertension across primary care practices participating in the EvidenceNOW quality-improvement initiative. We characterized retention in care, longitudinal blood pressure (BP) control, and predictors of loss to follow-up using descriptive statistics, cumulative retention curves, and multivariable Cox proportional-hazards regression. At baseline, mean systolic and diastolic BP were 140.0 ± 20.6 and 84.7 ± 13.0 mmHg, respectively; only 10.7% (95% CI 10.4–11.1) of patients had controlled BP and 18.1% never returned for any follow-up visit. Among the 21,729 patients who had ≥1 follow-up encounter, retention declined steeply over time—from 59.9% at 6 months to 16.3% at 36 months. Patients identifying as Black/African American (adjusted hazard ratio [aHR] 1.44; 95% CI 1.33–1.56), Hispanic/Latino (aHR 1.43; 1.35–1.52), or Other race/ethnicity (aHR 1.50; 1.41–1.59) had significantly higher hazards of being lost to follow-up than White patients, whereas older age, female sex, comorbid diabetes, heart failure, chronic kidney disease, stroke, and baseline BP control were each independently protective. Among patients retained for at least 12 months, BP control rose to 63.7% and remained near 64–66% through 36 months. These findings reveal a substantial and inequitable longitudinal care-engagement gap that is likely a principal driver of suboptimal hypertension control in the United States and identify actionable demographic and clinical targets for primary-care retention interventions.
Open AccessRetention and loss to follow-up among patients with hypertension in primary care: a multi-practice cohort study
Effective hypertension management depends on sustained engagement with primary care, and there is a need to understand the magnitude and determinants of follow-up loss in real-world primary care. We analyzed electronic health record (EHR) data from 26,541 patients with hypertension across primary care practices participating in the EvidenceNOW quality-improvement initiative. We characterized retention in care, longitudinal blood pressure (BP) control, and predictors of loss to follow-up using descriptive statistics, cumulative retention curves, and multivariable Cox proportional-hazards regression. At baseline, mean systolic and diastolic BP were 140.0 ± 20.6 and 84.7 ± 13.0 mmHg, respectively; only 10.7% (95% CI 10.4–11.1) of patients had controlled BP and 18.1% never returned for any follow-up visit. Among the 21,729 patients who had ≥1 follow-up encounter, retention declined steeply over time—from 59.9% at 6 months to 16.3% at 36 months. Patients identifying as Black/African American (adjusted hazard ratio [aHR] 1.44; 95% CI 1.33–1.56), Hispanic/Latino (aHR 1.43; 1.35–1.52), or Other race/ethnicity (aHR 1.50; 1.41–1.59) had significantly higher hazards of being lost to follow-up than White patients, whereas older age, female sex, comorbid diabetes, heart failure, chronic kidney disease, stroke, and baseline BP control were each independently protective. Among patients retained for at least 12 months, BP control rose to 63.7% and remained near 64–66% through 36 months. These findings reveal a substantial and inequitable longitudinal care-engagement gap that is likely a principal driver of suboptimal hypertension control in the United States and identify actionable demographic and clinical targets for primary-care retention interventions.