Guanhua Chen
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
Guanhua Chen has a total of 410 co-authors. Here are the top co-authors by number of shared publications:
- Jared Davis Huling
Division of Biostatistics and Health Data Science University of Minnesota
7 shared publications - Jonathan L. Temte
Department of Family Medicine and Community Health University of Wisconsin Madison Wisconsin USA
7 shared publications - Maureen Goss
Department of Family Medicine and Community Health University of Wisconsin Madison Wisconsin USA
7 shared publications - Talat Alp Ikizler
Division of Nephrology and Hypertension, Vanderbilt University Medical Center, Nashville, Tennessee, USA
7 shared publications Aihua Bian
Department of Biostatistics, Vanderbilt University Medical Center, Nashville, Tennessee, USA
6 shared publications- Cristalyne Bell
Affiliation Department of Family Medicine and Community Health, School of Medicine and Public Health, University of Wisconsin, Madison, Wisconsin, United States of America
6 shared publications
Latest Publications
- 24 February 2026
- Published by Elsevier in Journal of Biomedical Informatics
- Vol. 177, 105001
Explainable multimodal deep learning models for variable-length sequences in critically ill patients
- 24 February 2026
- Published by Elsevier in Journal of Biomedical Informatics
- Vol. 177, 105001
Explainable multimodal deep learning models for variable-length sequences in critically ill patients
- Open Access
Differentially Private Two-Stage Empirical Risk Minimization and Applications to Individualized Treatment Rule
Differential Privacy (DP) provides a rigorous framework for deriving privacy-preserving estimators by injecting calibrated noise to mask individual contributions while preserving population-level insights. Its central challenge lies in the privacy-utility trade-off: calibrating noise levels to ensure robust protection without compromising statistical performance. Standard DP methods struggle with a particular class of two-stage problems prevalent in individualized treatment rules (ITRs) and causal inference. In these settings, data-dependent weights are first computed to satisfy distributional constraints, such as covariate balance, before the final parameter of interest is estimated. Current DP approaches often privatize stages independently, which either degrades weight efficacy-leading to biased and inconsistent estimates-or introduces excessive noise to account for worst-case scenarios. To address these challenges, we propose the Differentially Private Two-Stage Empirical Risk Minimization (DP-2ERM), a framework that injects a carefully calibrated noise only into the second stage while maintaining privacy for the entire pipeline and preserving the integrity of the first stage weights. Our theoretical contributions include deterministic bounds on weight perturbations across various widely used weighting methods, and probabilistic bounds on sensitivity for the final estimator. Simulations and real-world applications in ITR demonstrate that DP-2ERM significantly enhances utility over existing methods while providing rigorous privacy guarantees.
Open AccessDifferentially Private Two-Stage Empirical Risk Minimization and Applications to Individualized Treatment Rule
Differential Privacy (DP) provides a rigorous framework for deriving privacy-preserving estimators by injecting calibrated noise to mask individual contributions while preserving population-level insights. Its central challenge lies in the privacy-utility trade-off: calibrating noise levels to ensure robust protection without compromising statistical performance. Standard DP methods struggle with a particular class of two-stage problems prevalent in individualized treatment rules (ITRs) and causal inference. In these settings, data-dependent weights are first computed to satisfy distributional constraints, such as covariate balance, before the final parameter of interest is estimated. Current DP approaches often privatize stages independently, which either degrades weight efficacy-leading to biased and inconsistent estimates-or introduces excessive noise to account for worst-case scenarios. To address these challenges, we propose the Differentially Private Two-Stage Empirical Risk Minimization (DP-2ERM), a framework that injects a carefully calibrated noise only into the second stage while maintaining privacy for the entire pipeline and preserving the integrity of the first stage weights. Our theoretical contributions include deterministic bounds on weight perturbations across various widely used weighting methods, and probabilistic bounds on sensitivity for the final estimator. Simulations and real-world applications in ITR demonstrate that DP-2ERM significantly enhances utility over existing methods while providing rigorous privacy guarantees.
- Open Access
- 9 December 2025
- Published by Springer Nature in npj Digital Medicine
- Vol. 9 (1)
MoMA: a mixture-of-multimodal-agents architecture for enhancing clinical prediction modelling
Multimodal electronic health record (EHR) data provide richer, complementary insights into patient health compared to single-modality data. However, effectively integrating diverse data modalities for clinical prediction modeling remains challenging due to the substantial data requirements. We introduce a novel architecture, Mixture-of-Multimodal-Agents (MoMA), designed to leverage multiple large language model (LLM) agents for clinical prediction tasks using multimodal EHR data. MoMA employs specialized LLM agents (“specialist agents”) to convert non-textual modalities, such as medical images and laboratory results, into structured textual summaries. These summaries, together with clinical notes, are combined by another LLM (“aggregator agent”) to generate a unified multimodal summary, which is then used by a third LLM (“predictor agent”) to produce clinical predictions. Evaluating MoMA with different modality combinations and prediction settings, MoMA outperforms existing methods on three prediction tasks using private datasets, highlighting its enhanced accuracy and flexibility across various tasks.
- 9 December 2025
- Published by Springer Nature in npj Digital Medicine
- Vol. 9 (1)
Open AccessMoMA: a mixture-of-multimodal-agents architecture for enhancing clinical prediction modelling
Multimodal electronic health record (EHR) data provide richer, complementary insights into patient health compared to single-modality data. However, effectively integrating diverse data modalities for clinical prediction modeling remains challenging due to the substantial data requirements. We introduce a novel architecture, Mixture-of-Multimodal-Agents (MoMA), designed to leverage multiple large language model (LLM) agents for clinical prediction tasks using multimodal EHR data. MoMA employs specialized LLM agents (“specialist agents”) to convert non-textual modalities, such as medical images and laboratory results, into structured textual summaries. These summaries, together with clinical notes, are combined by another LLM (“aggregator agent”) to generate a unified multimodal summary, which is then used by a third LLM (“predictor agent”) to produce clinical predictions. Evaluating MoMA with different modality combinations and prediction settings, MoMA outperforms existing methods on three prediction tasks using private datasets, highlighting its enhanced accuracy and flexibility across various tasks.
- Open Access
MoMA: A Mixture-of-Multimodal-Agents Architecture for Enhancing Clinical Prediction Modelling
Multimodal electronic health record (EHR) data provide richer, complementary insights into patient health compared to single-modality data. However, effectively integrating diverse data modalities for clinical prediction modeling remains challenging due to the substantial data requirements. We introduce a novel architecture, Mixture-of-Multimodal-Agents (MoMA), designed to leverage multiple large language model (LLM) agents for clinical prediction tasks using multimodal EHR data. MoMA employs specialized LLM agents ("specialist agents") to convert non-textual modalities, such as medical images and laboratory results, into structured textual summaries. These summaries, together with clinical notes, are combined by another LLM ("aggregator agent") to generate a unified multimodal summary, which is then used by a third LLM ("predictor agent") to produce clinical predictions. Evaluating MoMA on three prediction tasks using real-world datasets with different modality combinations and prediction settings, MoMA outperforms current state-of-the-art methods, highlighting its enhanced accuracy and flexibility across various tasks.
Open AccessMoMA: A Mixture-of-Multimodal-Agents Architecture for Enhancing Clinical Prediction Modelling
Multimodal electronic health record (EHR) data provide richer, complementary insights into patient health compared to single-modality data. However, effectively integrating diverse data modalities for clinical prediction modeling remains challenging due to the substantial data requirements. We introduce a novel architecture, Mixture-of-Multimodal-Agents (MoMA), designed to leverage multiple large language model (LLM) agents for clinical prediction tasks using multimodal EHR data. MoMA employs specialized LLM agents ("specialist agents") to convert non-textual modalities, such as medical images and laboratory results, into structured textual summaries. These summaries, together with clinical notes, are combined by another LLM ("aggregator agent") to generate a unified multimodal summary, which is then used by a third LLM ("predictor agent") to produce clinical predictions. Evaluating MoMA on three prediction tasks using real-world datasets with different modality combinations and prediction settings, MoMA outperforms current state-of-the-art methods, highlighting its enhanced accuracy and flexibility across various tasks.
- Open Access
- 10 April 2025
- Published by JMIR Publications Inc. in JMIR AI
- Vol. 4 (1) , e67144
Comparison of Deep Learning Approaches Using Chest Radiographs for Predicting Clinical Deterioration: Retrospective Observational Study
Background: The early detection of clinical deterioration and timely intervention for hospitalized patients can improve patient outcomes. The currently existing early warning systems rely on variables from structured data, such as vital signs and laboratory values, and do not incorporate other potentially predictive data modalities. Because respiratory failure is a common cause of deterioration, chest radiographs are often acquired in patients with clinical deterioration, which may be informative for predicting their risk of intensive care unit (ICU) transfer. Objective: This study aimed to compare and validate different computer vision models and data augmentation approaches with chest radiographs for predicting clinical deterioration. Methods: This retrospective observational study included adult patients hospitalized at the University of Wisconsin Health System between 2009 and 2020 with an elevated electronic cardiac arrest risk triage (eCART) score, a validated clinical deterioration early warning score, on the medical-surgical wards. Patients with a chest radiograph obtained within 48 hours prior to the elevated score were included in this study. Five computer vision model architectures (VGG16, DenseNet121, Vision Transformer, ResNet50, and Inception V3) and four data augmentation methods (histogram normalization, random flip, random Gaussian noise, and random rotate) were compared using the area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPRC) for predicting clinical deterioration (ie, ICU transfer or ward death in the following 24 hours). Results: The study included 21,817 patient admissions, of which 1655 (7.6%) experienced clinical deterioration. The DenseNet121 model pretrained on chest radiograph datasets with histogram normalization and random Gaussian noise augmentation had the highest discrimination (AUROC 0.734 and AUPRC 0.414), while the vision transformer having 24 transformer blocks with random rotate augmentation had the lowest discrimination (AUROC 0.598). Conclusions: The study shows the potential of chest radiographs in deep learning models for predicting clinical deterioration. The DenseNet121 architecture pretrained with chest radiographs performed better than other architectures in most experiments, and the addition of histogram normalization with random Gaussian noise data augmentation may enhance the performance of DenseNet121 and pretrained VGG16 architectures.
- 10 April 2025
- Published by JMIR Publications Inc. in JMIR AI
- Vol. 4 (1) , e67144
Open AccessComparison of Deep Learning Approaches Using Chest Radiographs for Predicting Clinical Deterioration: Retrospective Observational Study
Background: The early detection of clinical deterioration and timely intervention for hospitalized patients can improve patient outcomes. The currently existing early warning systems rely on variables from structured data, such as vital signs and laboratory values, and do not incorporate other potentially predictive data modalities. Because respiratory failure is a common cause of deterioration, chest radiographs are often acquired in patients with clinical deterioration, which may be informative for predicting their risk of intensive care unit (ICU) transfer. Objective: This study aimed to compare and validate different computer vision models and data augmentation approaches with chest radiographs for predicting clinical deterioration. Methods: This retrospective observational study included adult patients hospitalized at the University of Wisconsin Health System between 2009 and 2020 with an elevated electronic cardiac arrest risk triage (eCART) score, a validated clinical deterioration early warning score, on the medical-surgical wards. Patients with a chest radiograph obtained within 48 hours prior to the elevated score were included in this study. Five computer vision model architectures (VGG16, DenseNet121, Vision Transformer, ResNet50, and Inception V3) and four data augmentation methods (histogram normalization, random flip, random Gaussian noise, and random rotate) were compared using the area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPRC) for predicting clinical deterioration (ie, ICU transfer or ward death in the following 24 hours). Results: The study included 21,817 patient admissions, of which 1655 (7.6%) experienced clinical deterioration. The DenseNet121 model pretrained on chest radiograph datasets with histogram normalization and random Gaussian noise augmentation had the highest discrimination (AUROC 0.734 and AUPRC 0.414), while the vision transformer having 24 transformer blocks with random rotate augmentation had the lowest discrimination (AUROC 0.598). Conclusions: The study shows the potential of chest radiographs in deep learning models for predicting clinical deterioration. The DenseNet121 architecture pretrained with chest radiographs performed better than other architectures in most experiments, and the addition of histogram normalization with random Gaussian noise data augmentation may enhance the performance of DenseNet121 and pretrained VGG16 architectures.
- Open Access
- 13 March 2025
- Published by Wiley
Effects of K-12 school district non-pharmaceutical interventions on community-level prevalence of acute respiratory infection during the COVID-19 pandemic
Background: Responding to the COVID-19 pandemic, kindergarten through 12 th grade schools implemented non-pharmaceutical interventions (NPIs). The effects of school-based NPIs on broader community levels of acute respiratory infection (ARI) have not been defined. We utilized an existing longitudinal cohort of households reporting weekly ARI cases to evaluate effects of evolving school districtwide NPIs on ARI activity at eight transition points from December 2019 through October 2022. Methods: Household ARI data was reported through GROVES (the GReat ORCHARDS Vaccine Effectiveness Study—a prospective cohort study based in the OSD). Participating GROVES families completed weekly online surveys with respiratory illness updates. Mixed effects logistic regression was used to examine the association between eight school-related transition events during the COVID-19 pandemic and changes in trajectory of ARI risk for GROVES family members, while accounting for family clusters. Transition events were assessed using a ±4-week window of community data. Results: Opening schools with maximal NPIs (mandated masking and physical distancing, with hybrid education) was not associated with increased community ARI activity. The four transition events associated with significant ARI risk trajectory increases included summer breaks (June 2020, p=0.001; June 2021, p=0.002), and the start of school with mandatory masking only (September 2021, pConclusions: School-based NPI implementation was associated with reduced risks for community ARI activity. Enhanced surveillance platforms such as the weekly online surveys used in this study are valuable tools for better understanding and monitoring SARS-CoV-2 and respiratory virus transmission in schools and surrounding communities.
- 13 March 2025
- Published by Wiley
Open AccessEffects of K-12 school district non-pharmaceutical interventions on community-level prevalence of acute respiratory infection during the COVID-19 pandemic
Background: Responding to the COVID-19 pandemic, kindergarten through 12 th grade schools implemented non-pharmaceutical interventions (NPIs). The effects of school-based NPIs on broader community levels of acute respiratory infection (ARI) have not been defined. We utilized an existing longitudinal cohort of households reporting weekly ARI cases to evaluate effects of evolving school districtwide NPIs on ARI activity at eight transition points from December 2019 through October 2022. Methods: Household ARI data was reported through GROVES (the GReat ORCHARDS Vaccine Effectiveness Study—a prospective cohort study based in the OSD). Participating GROVES families completed weekly online surveys with respiratory illness updates. Mixed effects logistic regression was used to examine the association between eight school-related transition events during the COVID-19 pandemic and changes in trajectory of ARI risk for GROVES family members, while accounting for family clusters. Transition events were assessed using a ±4-week window of community data. Results: Opening schools with maximal NPIs (mandated masking and physical distancing, with hybrid education) was not associated with increased community ARI activity. The four transition events associated with significant ARI risk trajectory increases included summer breaks (June 2020, p=0.001; June 2021, p=0.002), and the start of school with mandatory masking only (September 2021, pConclusions: School-based NPI implementation was associated with reduced risks for community ARI activity. Enhanced surveillance platforms such as the weekly online surveys used in this study are valuable tools for better understanding and monitoring SARS-CoV-2 and respiratory virus transmission in schools and surrounding communities.
- Open Access
- 28 January 2025
- Published by Oxford University Press (OUP) in Bioinformatics
- Vol. 41 (2)
Multidimensional scaling improves distance-based clustering for microbiome data
Motivation: Clustering patients into subgroups based on their microbial compositions can greatly enhance our understanding of the role of microbes in human health and disease etiology. Distance-based clustering methods, such as partitioning around medoids (PAM), are popular due to their computational efficiency and absence of distributional assumptions. However, the performance of these methods can be suboptimal when true cluster memberships are driven by differences in the abundance of only a few microbes, a situation known as the sparse signal scenario. Results: We demonstrate that classical multidimensional scaling (MDS), a widely used dimensionality reduction technique, effectively denoises microbiome data and enhances the clustering performance of distance-based methods. We propose a two-step procedure that first applies MDS to project high-dimensional microbiome data into a low-dimensional space, followed by distance-based clustering using the low-dimensional data. Our extensive simulations demonstrate that our procedure offers superior performance compared to directly conducting distance-based clustering under the sparse signal scenario. The advantage of our procedure is further showcased in several real data applications. Availability and implementation: The R package MDSMClust is available at https://github.com/wxy929/MDS-project.
- 28 January 2025
- Published by Oxford University Press (OUP) in Bioinformatics
- Vol. 41 (2)
Open AccessMultidimensional scaling improves distance-based clustering for microbiome data
Motivation: Clustering patients into subgroups based on their microbial compositions can greatly enhance our understanding of the role of microbes in human health and disease etiology. Distance-based clustering methods, such as partitioning around medoids (PAM), are popular due to their computational efficiency and absence of distributional assumptions. However, the performance of these methods can be suboptimal when true cluster memberships are driven by differences in the abundance of only a few microbes, a situation known as the sparse signal scenario. Results: We demonstrate that classical multidimensional scaling (MDS), a widely used dimensionality reduction technique, effectively denoises microbiome data and enhances the clustering performance of distance-based methods. We propose a two-step procedure that first applies MDS to project high-dimensional microbiome data into a low-dimensional space, followed by distance-based clustering using the low-dimensional data. Our extensive simulations demonstrate that our procedure offers superior performance compared to directly conducting distance-based clustering under the sparse signal scenario. The advantage of our procedure is further showcased in several real data applications. Availability and implementation: The R package MDSMClust is available at https://github.com/wxy929/MDS-project.
- Open Access
- 9 January 2025
- Published by Ubiquity Press, Ltd. in Global Heart
- Vol. 20 (1) , 39
Community Health Workers Equipped with an mHealth Application Can Accurately Diagnose Hypertension in Rural Guatemala
Background: Hypertension is a leading global cause of morbidity and mortality and is increasing in low- and middle-income countries, where unawareness of hypertension is a primary obstacle to management. Community health workers (CHWs) in combination with mobile health (mHealth) tools are increasingly used in LMIC health systems to strengthen primary care infrastructure. In this study, we applied this care model to hypertension in rural Guatemala by comparing the accuracy of CHWs equipped with an mHealth clinical decision support application in diagnosing hypertension to concurrent physician evaluation. Methods: We performed a prospective diagnostic accuracy study in which adults from rural Guatemalan communities were assessed independently by a CHW aided by a mHealth application and a physician. Assessment included medical history; measurement of blood pressure, height and weight; and determination of hypertension status. CHW-physician agreement on hypertension status and past medical history elements was assessed by Kappa analysis and proportional agreement, with a priori thresholds of Kappa = 0.61 and agreement of 90%. Agreement on patient measurements was evaluated using Bland-Altman and regression analyses. Results: Of 359 participants enrolled, 47 (13%) were confirmed to have hypertension and another 11 (3%) had possible hypertension. CHW-physician agreement was high for hypertension diagnosis, with Kappa = 0.8 (95% CI = 0.72, 0.88) and overall agreement 92.8% (95% CI = 90.1%, 95.4%). Bland-Altman analysis showed small biases toward lower systolic blood pressure, higher height, and lower BMI measurements by CHWs. Most patient history characteristics showed moderate to almost perfect (Kappa: 0.41–1) agreement between physicians and CHWs. Conclusions: In this study based in rural Guatemala, CHWs using a mHealth clinical decision support application were found to screen adult patients for hypertension with similar accuracy to a physician. This approach could be adapted to other low-resource settings to reduce the burden of undiagnosed and untreated hypertension.
- 9 January 2025
- Published by Ubiquity Press, Ltd. in Global Heart
- Vol. 20 (1) , 39
Open AccessCommunity Health Workers Equipped with an mHealth Application Can Accurately Diagnose Hypertension in Rural Guatemala
Background: Hypertension is a leading global cause of morbidity and mortality and is increasing in low- and middle-income countries, where unawareness of hypertension is a primary obstacle to management. Community health workers (CHWs) in combination with mobile health (mHealth) tools are increasingly used in LMIC health systems to strengthen primary care infrastructure. In this study, we applied this care model to hypertension in rural Guatemala by comparing the accuracy of CHWs equipped with an mHealth clinical decision support application in diagnosing hypertension to concurrent physician evaluation. Methods: We performed a prospective diagnostic accuracy study in which adults from rural Guatemalan communities were assessed independently by a CHW aided by a mHealth application and a physician. Assessment included medical history; measurement of blood pressure, height and weight; and determination of hypertension status. CHW-physician agreement on hypertension status and past medical history elements was assessed by Kappa analysis and proportional agreement, with a priori thresholds of Kappa = 0.61 and agreement of 90%. Agreement on patient measurements was evaluated using Bland-Altman and regression analyses. Results: Of 359 participants enrolled, 47 (13%) were confirmed to have hypertension and another 11 (3%) had possible hypertension. CHW-physician agreement was high for hypertension diagnosis, with Kappa = 0.8 (95% CI = 0.72, 0.88) and overall agreement 92.8% (95% CI = 90.1%, 95.4%). Bland-Altman analysis showed small biases toward lower systolic blood pressure, higher height, and lower BMI measurements by CHWs. Most patient history characteristics showed moderate to almost perfect (Kappa: 0.41–1) agreement between physicians and CHWs. Conclusions: In this study based in rural Guatemala, CHWs using a mHealth clinical decision support application were found to screen adult patients for hypertension with similar accuracy to a physician. This approach could be adapted to other low-resource settings to reduce the burden of undiagnosed and untreated hypertension.
- Open Access
- 26 December 2024
- Published by Oxford University Press (OUP) in JAMIA Open
- Vol. 8 (1) , ooae154
Uncertainty estimation in diagnosis generation from large language models: next-word probability is not pre-test probability
Objective: To evaluate large language models (LLMs) for pre-test diagnostic probability estimation and compare their uncertainty estimation performance with a traditional machine learning classifier. Materials and Methods: We assessed 2 instruction-tuned LLMs, Mistral-7B-Instruct and Llama3-70B-chat-hf, on predicting binary outcomes for Sepsis, Arrhythmia, and Congestive Heart Failure (CHF) using electronic health record (EHR) data from 660 patients. Three uncertainty estimation methods—Verbalized Confidence, Token Logits, and LLM Embedding+XGB—were compared against an eXtreme Gradient Boosting (XGB) classifier trained on raw EHR data. Performance metrics included AUROC and Pearson correlation between predicted probabilities. Results: The XGB classifier outperformed the LLM-based methods across all tasks. LLM Embedding+XGB showed the closest performance to the XGB baseline, while Verbalized Confidence and Token Logits underperformed. Discussion: These findings, consistent across multiple models and demographic groups, highlight the limitations of current LLMs in providing reliable pre-test probability estimations and underscore the need for improved calibration and bias mitigation strategies. Future work should explore hybrid approaches that integrate LLMs with numerical reasoning modules and calibrated embeddings to enhance diagnostic accuracy and ensure fairer predictions across diverse populations. Conclusions: LLMs demonstrate potential but currently fall short in estimating diagnostic probabilities compared to traditional machine learning classifiers trained on structured EHR data. Further improvements are needed for reliable clinical use.
- 26 December 2024
- Published by Oxford University Press (OUP) in JAMIA Open
- Vol. 8 (1) , ooae154
Open AccessUncertainty estimation in diagnosis generation from large language models: next-word probability is not pre-test probability
Objective: To evaluate large language models (LLMs) for pre-test diagnostic probability estimation and compare their uncertainty estimation performance with a traditional machine learning classifier. Materials and Methods: We assessed 2 instruction-tuned LLMs, Mistral-7B-Instruct and Llama3-70B-chat-hf, on predicting binary outcomes for Sepsis, Arrhythmia, and Congestive Heart Failure (CHF) using electronic health record (EHR) data from 660 patients. Three uncertainty estimation methods—Verbalized Confidence, Token Logits, and LLM Embedding+XGB—were compared against an eXtreme Gradient Boosting (XGB) classifier trained on raw EHR data. Performance metrics included AUROC and Pearson correlation between predicted probabilities. Results: The XGB classifier outperformed the LLM-based methods across all tasks. LLM Embedding+XGB showed the closest performance to the XGB baseline, while Verbalized Confidence and Token Logits underperformed. Discussion: These findings, consistent across multiple models and demographic groups, highlight the limitations of current LLMs in providing reliable pre-test probability estimations and underscore the need for improved calibration and bias mitigation strategies. Future work should explore hybrid approaches that integrate LLMs with numerical reasoning modules and calibrated embeddings to enhance diagnostic accuracy and ensure fairer predictions across diverse populations. Conclusions: LLMs demonstrate potential but currently fall short in estimating diagnostic probabilities compared to traditional machine learning classifiers trained on structured EHR data. Further improvements are needed for reliable clinical use.
- Open Access
- 3 October 2024
- Published by JMIR Publications Inc. in JMIR Preprints
Comparing Deep Learning Approaches for Predicting Clinical Deterioration Using Chest Radiographs (Preprint)
BACKGROUND Early detection of clinical deterioration and timely intervention for hospitalized patients can improve patient outcomes. Existing early warning systems rely on variables from structured data, such as vital signs and laboratory values, and do not incorporate other potentially predictive data modalities. Because respiratory failure is a common cause of deterioration, chest radiographs are often acquired in deteriorating patients, which may be informative for predicting their risk of intensive care unit (ICU) transfer. OBJECTIVE To compare and validate different computer vision models and data augmentation approaches with chest radiographs for predicting clinical deterioration. METHODS This retrospective observational study included adult patients hospitalized at the University of Wisconsin Health System between 2009 and 2020 with an elevated eCART score, a validated clinical deterioration early warning score, on the medical-surgical wards. Patients with a chest radiograph within 48 hours prior to the elevated score were included in this study. Three computer vision model architectures (VGG16, Densenet121, Vision Transformer) and four data augmentation methods (Histogram Normalization, Random Flip, Random Gaussian Noise, and Random Rotate) were compared using the area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPRC) for predicting clinical deterioration (i.e., intensive care unit transfer or ward death in the following 24 hours). RESULTS The study included 21,817 patient admissions, of which 1,655 (7.6%) experienced the outcome. Densenet121 model pre-trained on chest radiograph datasets with histogram normalization and random Gaussian noise augmentation had the highest discrimination (AUROC 0.734 and AUPRC 0.414), while vision transformer having 24 transformer blocks with random rotate augmentation had the lowest discrimination (AUROC=0.598). CONCLUSIONS The Densenet121 architecture pretrained with chest radiographs performed better than other architectures in most experiments, and the addition of histogram normalization with random Gaussian noise data augmentation may enhance performance for Densenet121 and pre-trained VGG16 architectures. CLINICALTRIAL
- 3 October 2024
- Published by JMIR Publications Inc. in JMIR Preprints
Open AccessComparing Deep Learning Approaches for Predicting Clinical Deterioration Using Chest Radiographs (Preprint)
BACKGROUND Early detection of clinical deterioration and timely intervention for hospitalized patients can improve patient outcomes. Existing early warning systems rely on variables from structured data, such as vital signs and laboratory values, and do not incorporate other potentially predictive data modalities. Because respiratory failure is a common cause of deterioration, chest radiographs are often acquired in deteriorating patients, which may be informative for predicting their risk of intensive care unit (ICU) transfer. OBJECTIVE To compare and validate different computer vision models and data augmentation approaches with chest radiographs for predicting clinical deterioration. METHODS This retrospective observational study included adult patients hospitalized at the University of Wisconsin Health System between 2009 and 2020 with an elevated eCART score, a validated clinical deterioration early warning score, on the medical-surgical wards. Patients with a chest radiograph within 48 hours prior to the elevated score were included in this study. Three computer vision model architectures (VGG16, Densenet121, Vision Transformer) and four data augmentation methods (Histogram Normalization, Random Flip, Random Gaussian Noise, and Random Rotate) were compared using the area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPRC) for predicting clinical deterioration (i.e., intensive care unit transfer or ward death in the following 24 hours). RESULTS The study included 21,817 patient admissions, of which 1,655 (7.6%) experienced the outcome. Densenet121 model pre-trained on chest radiograph datasets with histogram normalization and random Gaussian noise augmentation had the highest discrimination (AUROC 0.734 and AUPRC 0.414), while vision transformer having 24 transformer blocks with random rotate augmentation had the lowest discrimination (AUROC=0.598). CONCLUSIONS The Densenet121 architecture pretrained with chest radiographs performed better than other architectures in most experiments, and the addition of histogram normalization with random Gaussian noise data augmentation may enhance performance for Densenet121 and pre-trained VGG16 architectures. CLINICALTRIAL