Livie Yumeng Li
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Livie Yumeng Li has a total of 52 co-authors. Here are the top co-authors by number of shared publications:
- Adam Hulman
Department of Public Health, Aarhus University, Aarhus, Denmark
6 shared publications Benjamin Lebiecka-Johansen
Steno Diabetes Center Aarhus, Aarhus University Hospital, Aarhus, Denmark
3 shared publicationsStine Byberg
Clinical Epidemiological Research, Copenhagen University Hospital - Steno Diabetes Center Copenhagen, Herlev, Denmark; Steno Diabetes Center Greenland, Nuuk, Greenland
3 shared publicationsAnders Aasted Isaksen
Affiliation Steno Diabetes Center Aarhus, Aarhus, Denmark
2 shared publications- Anders Aasted Isaksen
Steno Diabetes Center Aarhus, Aarhus University Hospital, Aarhus, Denmark
2 shared publications - Benjamin Lebiecka-Johansen
Steno Diabetes Center Aarhus, Palle Juul-Jensens Boulevard 11, Aarhus, Central Denmark Region, 8200, Denmark, +45 23 70 74 81
2 shared publications
Latest Publications
- Open Access
- 21 May 2026
- Published by Elsevier in International Journal of Medical Informatics
- Vol. 217, 106503
Assessing the generalisability of foundation models to ultra-wide field retinal imaging for diabetic retinopathy screening in Denmark and Greenland
- 21 May 2026
- Published by Elsevier in International Journal of Medical Informatics
- Vol. 217, 106503
Open AccessAssessing the generalisability of foundation models to ultra-wide field retinal imaging for diabetic retinopathy screening in Denmark and Greenland
- Open Access
Comparison of foundation models and transfer learning strategies for diabetic retinopathy classification
Diabetic retinopathy (DR) is a leading cause of vision impairment, requiring accurate and scalable diagnostic tools. Foundation models are increasingly applied to clinical imaging, but concerns remain about their calibration. We evaluated DINOv3, RETFound, and VisionFM for DR classification using different transfer learning strategies in BRSET (n = 16,266) and mBRSET (n = 5,164). Models achieved high discrimination in binary classification (normal vs retinopathy) in BRSET (AUROC 0.90-0.98), with DINOv3 achieving the best under full fine-tuning (AUROC 0.98 [95% CI: 0.97-0.99]). External validation on mBRSET showed decreased performance for all models regardless of the fine-tuning strategy (AUROC 0.70–0.85), though fine-tuning improved performance. Foundation models achieved strong discrimination but poor calibration, generally overestimating DR risk. While the generalist model, DINOv3, benefited from deeper fine-tuning, miscalibration remained evident. These findings underscore the need to improve calibration and the comprehensive evaluation of foundation models, which are essential in clinical settings. Author summary: Artificial intelligence is increasingly being used to detect eye diseases such as diabetic retinopathy from retinal images. Recent advances have introduced “foundation models,” which are trained on large datasets and can be adapted to new tasks. We aimed to evaluate how well these models perform in a clinical prediction context, with a focus not only on accuracy but also on how reliably they estimate disease risk. In this study, we compared different types of foundation models using two independent datasets from Brazil. We found that while these models were generally good at distinguishing between healthy and diseased eyes, their predicted risks were often poorly calibrated. In other words, the estimated probabilities did not consistently reflect the true likelihood of disease. We also examined whether adapting the models to the target population could improve performance. Although this approach led to improvements, calibration issues remained. However, post-training correction improved the agreement between predicted risks and observed outcomes. Our findings highlight an important gap between model performance and clinical usefulness. We suggest that improving the reliability of risk estimates is essential before such systems can be safely used in healthcare.
Open AccessComparison of foundation models and transfer learning strategies for diabetic retinopathy classification
Diabetic retinopathy (DR) is a leading cause of vision impairment, requiring accurate and scalable diagnostic tools. Foundation models are increasingly applied to clinical imaging, but concerns remain about their calibration. We evaluated DINOv3, RETFound, and VisionFM for DR classification using different transfer learning strategies in BRSET (n = 16,266) and mBRSET (n = 5,164). Models achieved high discrimination in binary classification (normal vs retinopathy) in BRSET (AUROC 0.90-0.98), with DINOv3 achieving the best under full fine-tuning (AUROC 0.98 [95% CI: 0.97-0.99]). External validation on mBRSET showed decreased performance for all models regardless of the fine-tuning strategy (AUROC 0.70–0.85), though fine-tuning improved performance. Foundation models achieved strong discrimination but poor calibration, generally overestimating DR risk. While the generalist model, DINOv3, benefited from deeper fine-tuning, miscalibration remained evident. These findings underscore the need to improve calibration and the comprehensive evaluation of foundation models, which are essential in clinical settings. Author summary: Artificial intelligence is increasingly being used to detect eye diseases such as diabetic retinopathy from retinal images. Recent advances have introduced “foundation models,” which are trained on large datasets and can be adapted to new tasks. We aimed to evaluate how well these models perform in a clinical prediction context, with a focus not only on accuracy but also on how reliably they estimate disease risk. In this study, we compared different types of foundation models using two independent datasets from Brazil. We found that while these models were generally good at distinguishing between healthy and diseased eyes, their predicted risks were often poorly calibrated. In other words, the estimated probabilities did not consistently reflect the true likelihood of disease. We also examined whether adapting the models to the target population could improve performance. Although this approach led to improvements, calibration issues remained. However, post-training correction improved the agreement between predicted risks and observed outcomes. Our findings highlight an important gap between model performance and clinical usefulness. We suggest that improving the reliability of risk estimates is essential before such systems can be safely used in healthcare.
- Open Access
- 30 June 2025
- Published by Public Library of Science (PLoS) in PLOS Digital Health
- Vol. 4 (6) , e0000918
Racial disparities in continuous glucose monitoring-based 60-min glucose predictions among people with type 1 diabetes
Non-Hispanic white (White) populations are overrepresented in medical studies. Potential healthcare disparities can happen when machine learning models, used in diabetes technologies, are trained on data from primarily White patients. We aimed to evaluate algorithmic fairness in glucose predictions. This study utilized continuous glucose monitoring (CGM) data from 101 White and 104 Black participants with type 1 diabetes collected by the JAEB Center for Health Research, US. Long short-term memory (LSTM) deep learning models were trained on 11 datasets of different proportions of White and Black participants and tailored to each individual using transfer learning to predict glucose 60 minutes ahead based on 60-minute windows. Root mean squared errors (RMSE) were calculated for each participant. Linear mixed-effect models were used to investigate the association between racial composition and RMSE while accounting for age, sex, and training data size. A median of 9 weeks (IQR: 7, 10) of CGM data was available per participant. The divergence in performance (RMSE slope by proportion) was not statistically significant for either group. However, the slope difference (from 0% White and 100% Black to 100% White and 0% Black) between groups was statistically significant (p = 0.02), meaning the RMSE increased 0.04 [0.01, 0.08] mmol/L more for Black participants compared to White participants when the proportion of White participants increased from 0 to 100% in the training data. This difference was attenuated in the transfer learned models (RMSE: 0.02 [-0.01, 0.05] mmol/L, p = 0.20). The racial composition of training data created a small statistically significant difference in the performance of the models, which was not present after using transfer learning. This demonstrates the importance of diversity in datasets and the potential value of transfer learning for developing more fair prediction models.
- 30 June 2025
- Published by Public Library of Science (PLoS) in PLOS Digital Health
- Vol. 4 (6) , e0000918
Open AccessRacial disparities in continuous glucose monitoring-based 60-min glucose predictions among people with type 1 diabetes
Non-Hispanic white (White) populations are overrepresented in medical studies. Potential healthcare disparities can happen when machine learning models, used in diabetes technologies, are trained on data from primarily White patients. We aimed to evaluate algorithmic fairness in glucose predictions. This study utilized continuous glucose monitoring (CGM) data from 101 White and 104 Black participants with type 1 diabetes collected by the JAEB Center for Health Research, US. Long short-term memory (LSTM) deep learning models were trained on 11 datasets of different proportions of White and Black participants and tailored to each individual using transfer learning to predict glucose 60 minutes ahead based on 60-minute windows. Root mean squared errors (RMSE) were calculated for each participant. Linear mixed-effect models were used to investigate the association between racial composition and RMSE while accounting for age, sex, and training data size. A median of 9 weeks (IQR: 7, 10) of CGM data was available per participant. The divergence in performance (RMSE slope by proportion) was not statistically significant for either group. However, the slope difference (from 0% White and 100% Black to 100% White and 0% Black) between groups was statistically significant (p = 0.02), meaning the RMSE increased 0.04 [0.01, 0.08] mmol/L more for Black participants compared to White participants when the proportion of White participants increased from 0 to 100% in the training data. This difference was attenuated in the transfer learned models (RMSE: 0.02 [-0.01, 0.05] mmol/L, p = 0.20). The racial composition of training data created a small statistically significant difference in the performance of the models, which was not present after using transfer learning. This demonstrates the importance of diversity in datasets and the potential value of transfer learning for developing more fair prediction models.
- Open Access
- 23 June 2025
- Published by Association for Computing Machinery (ACM)
- p. 511-531
In the Picture: Medical Imaging Datasets, Artifacts, and their Living Review
- 23 June 2025
- Published by Association for Computing Machinery (ACM)
- p. 511-531
Open AccessIn the Picture: Medical Imaging Datasets, Artifacts, and their Living Review
- Open Access
Racial disparities in continuous glucose monitoring-based 60-min glucose predictions among people with type 1 diabetes
Non-Hispanic white (White) populations are overrepresented in medical studies. Potential healthcare disparities can happen when machine learning models, used in diabetes technologies, are trained on data from primarily White patients. We aimed to evaluate algorithmic fairness in glucose predictions. This study utilized continuous glucose monitoring (CGM) data from 101 White and 104 Black participants with type 1 diabetes collected by the JAEB Center for Health Research, US. Long short-term memory (LSTM) deep learning models were trained on 11 datasets of different proportions of White and Black participants and tailored to each individual using transfer learning to predict glucose 60 minutes ahead based on 60-minute windows. Root mean squared errors (RMSE) were calculated for each participant. Linear mixed-effect models were used to investigate the association between racial composition and RMSE while accounting for age, sex, and training data size. A median of 9 weeks (IQR: 7, 10) of CGM data was available per participant. The divergence in performance (RMSE slope by proportion) was not statistically significant for either group. However, the slope difference (from 0% White and 100% Black to 100% White and 0% Black) between groups was statistically significant (p=0.02), meaning the RMSE increased 0.04 [0.01, 0.08] mmol/L more for Black participants compared to White participants when the proportion of White participants increased from 0 to 100% in the training data. This difference was attenuated in the transfer learned models (RMSE: 0.02 [-0.01, 0.05] mmol/L, p=0.20). The racial composition of training data created a small statistically significant difference in the performance of the models, which was not present after using transfer learning. This demonstrates the importance of diversity in datasets and the potential value of transfer learning for developing more fair prediction models.
Open AccessRacial disparities in continuous glucose monitoring-based 60-min glucose predictions among people with type 1 diabetes
Non-Hispanic white (White) populations are overrepresented in medical studies. Potential healthcare disparities can happen when machine learning models, used in diabetes technologies, are trained on data from primarily White patients. We aimed to evaluate algorithmic fairness in glucose predictions. This study utilized continuous glucose monitoring (CGM) data from 101 White and 104 Black participants with type 1 diabetes collected by the JAEB Center for Health Research, US. Long short-term memory (LSTM) deep learning models were trained on 11 datasets of different proportions of White and Black participants and tailored to each individual using transfer learning to predict glucose 60 minutes ahead based on 60-minute windows. Root mean squared errors (RMSE) were calculated for each participant. Linear mixed-effect models were used to investigate the association between racial composition and RMSE while accounting for age, sex, and training data size. A median of 9 weeks (IQR: 7, 10) of CGM data was available per participant. The divergence in performance (RMSE slope by proportion) was not statistically significant for either group. However, the slope difference (from 0% White and 100% Black to 100% White and 0% Black) between groups was statistically significant (p=0.02), meaning the RMSE increased 0.04 [0.01, 0.08] mmol/L more for Black participants compared to White participants when the proportion of White participants increased from 0 to 100% in the training data. This difference was attenuated in the transfer learned models (RMSE: 0.02 [-0.01, 0.05] mmol/L, p=0.20). The racial composition of training data created a small statistically significant difference in the performance of the models, which was not present after using transfer learning. This demonstrates the importance of diversity in datasets and the potential value of transfer learning for developing more fair prediction models.
- Open Access
- 10 September 2024
- Published by Oxford University Press (OUP) in European Heart Journal - Digital Health
- Vol. 5 (6) , 660-669
Prediction of cardiovascular markers and diseases using retinal fundus images and deep learning: a systematic scoping review
Rapid development in deep learning for image analysis inspired studies to focus on predicting cardiovascular risk using retinal fundus images. This scoping review aimed to identify and describe studies using retinal fundus images and deep learning to predict cardiovascular risk markers and diseases. We searched MEDLINE and Embase on 17 November 2023. Abstracts and relevant full-text articles were independently screened by two reviewers. We included studies that used deep learning for the analysis of retinal fundus images to predict cardiovascular risk markers or cardiovascular diseases (CVDs) and excluded studies only using predefined characteristics of retinal fundus images. Study characteristics were presented using descriptive statistics. We included 24 articles published between 2018 and 2023. Among these, 23 (96%) were cross-sectional studies and eight (33%) were follow-up studies with clinical CVD outcomes. Seven studies included a combination of both designs. Most studies (96%) used convolutional neural networks to process images. We found nine (38%) studies that incorporated clinical risk factors in the prediction and four (17%) that compared the results to commonly used clinical risk scores in a prospective setting. Three of these reported improved discriminative performance. External validation of models was rare (21%). There is increasing interest in using retinal fundus images in cardiovascular risk assessment with some studies demonstrating some improvements in prediction. However, more prospective studies, comparisons of results to clinical risk scores, and models augmented with traditional risk factors can strengthen further research in the field.
- 10 September 2024
- Published by Oxford University Press (OUP) in European Heart Journal - Digital Health
- Vol. 5 (6) , 660-669
Open AccessPrediction of cardiovascular markers and diseases using retinal fundus images and deep learning: a systematic scoping review
Rapid development in deep learning for image analysis inspired studies to focus on predicting cardiovascular risk using retinal fundus images. This scoping review aimed to identify and describe studies using retinal fundus images and deep learning to predict cardiovascular risk markers and diseases. We searched MEDLINE and Embase on 17 November 2023. Abstracts and relevant full-text articles were independently screened by two reviewers. We included studies that used deep learning for the analysis of retinal fundus images to predict cardiovascular risk markers or cardiovascular diseases (CVDs) and excluded studies only using predefined characteristics of retinal fundus images. Study characteristics were presented using descriptive statistics. We included 24 articles published between 2018 and 2023. Among these, 23 (96%) were cross-sectional studies and eight (33%) were follow-up studies with clinical CVD outcomes. Seven studies included a combination of both designs. Most studies (96%) used convolutional neural networks to process images. We found nine (38%) studies that incorporated clinical risk factors in the prediction and four (17%) that compared the results to commonly used clinical risk scores in a prospective setting. Three of these reported improved discriminative performance. External validation of models was rare (21%). There is increasing interest in using retinal fundus images in cardiovascular risk assessment with some studies demonstrating some improvements in prediction. However, more prospective studies, comparisons of results to clinical risk scores, and models augmented with traditional risk factors can strengthen further research in the field.
- Open Access
Prediction of Cardiovascular Markers and Diseases Using Retinal Fundus Images and Deep Learning: A Systematic Scoping Review
Background Cardiovascular risk prediction models based on sociodemographic factors and traditional clinical measurements have received significant attention. With rapid development in deep learning for image analysis in the last decade and the well-known association between micro- and macrovascular complications, some recent studies focused on the prediction of cardiovascular risk using retinal fundus images. The objective of this scoping review is to identify and describe studies using retinal fundus images and deep learning to predict cardiovascular risk markers and diseases. Methods We searched MEDLINE and Embase for peer-reviewed articles on 17 November 2023. Abstracts and relevant full-text articles were independently screened by two reviewers. We included studies that used deep learning for the analysis of retinal fundus images to predict cardiovascular risk markers (e.g. blood pressure, coronary artery calcification, intima-media thickness) or cardiovascular diseases (prevalent or incident). Studies that used only predefined characteristics of retinal fundus images (e.g. tortuosity, fractal dimension) were not considered. Study characteristics were extracted by the first author and verified by the senior author. Results are presented using descriptive statistics. Results We included 24 articles in the review, published between 2018 and 2023. Among these, 21 (88%) were cross-sectional studies and eight (33%) were follow-up studies with outcome of clinical CVD. Five studies included a combination of both designs. Most studies (n=23, 96%) used convolutional neural networks to process images. We found nine (38%) studies that incorporated clinical risk factors in the prediction and four (17%) that compared the results to commonly used clinical risk scores in a prospective setting. Three of these reported improved discriminative performance. External validation of models was rare (n=5, 21%). Only four (17%) studies made their code publicly available. Conclusions There is an increasing interest in using retinal fundus images in cardiovascular risk assessment. However, there is a need for more prospective studies, comparisons of results to clinical risk scores and models augmented with traditional risk factors. Moreover, more extensive code sharing is necessary to make findings reproducible and more impactful beyond a specific study.
Open AccessPrediction of Cardiovascular Markers and Diseases Using Retinal Fundus Images and Deep Learning: A Systematic Scoping Review
Background Cardiovascular risk prediction models based on sociodemographic factors and traditional clinical measurements have received significant attention. With rapid development in deep learning for image analysis in the last decade and the well-known association between micro- and macrovascular complications, some recent studies focused on the prediction of cardiovascular risk using retinal fundus images. The objective of this scoping review is to identify and describe studies using retinal fundus images and deep learning to predict cardiovascular risk markers and diseases. Methods We searched MEDLINE and Embase for peer-reviewed articles on 17 November 2023. Abstracts and relevant full-text articles were independently screened by two reviewers. We included studies that used deep learning for the analysis of retinal fundus images to predict cardiovascular risk markers (e.g. blood pressure, coronary artery calcification, intima-media thickness) or cardiovascular diseases (prevalent or incident). Studies that used only predefined characteristics of retinal fundus images (e.g. tortuosity, fractal dimension) were not considered. Study characteristics were extracted by the first author and verified by the senior author. Results are presented using descriptive statistics. Results We included 24 articles in the review, published between 2018 and 2023. Among these, 21 (88%) were cross-sectional studies and eight (33%) were follow-up studies with outcome of clinical CVD. Five studies included a combination of both designs. Most studies (n=23, 96%) used convolutional neural networks to process images. We found nine (38%) studies that incorporated clinical risk factors in the prediction and four (17%) that compared the results to commonly used clinical risk scores in a prospective setting. Three of these reported improved discriminative performance. External validation of models was rare (n=5, 21%). Only four (17%) studies made their code publicly available. Conclusions There is an increasing interest in using retinal fundus images in cardiovascular risk assessment. However, there is a need for more prospective studies, comparisons of results to clinical risk scores and models augmented with traditional risk factors. Moreover, more extensive code sharing is necessary to make findings reproducible and more impactful beyond a specific study.