Zhenyu Gong
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
Zhenyu Gong has a total of 56 co-authors. Here are the top co-authors by number of shared publications:
- Benedikt Wiestler
Technical University of Munich
2 shared publications - Bailiang Jian
Technical University of Munich
1 shared publication - Chao Li
School of Marine Science and Engineering, Qingdao Agricultural University, Qingdao 266109, China
1 shared publication Che Liu
1 shared publicationChen Chen
1 shared publicationChen Niu
PET/CT Center, The First Affiliated Hospital of Xi’an Jiaotong University, Xi’an, China
1 shared publication
Latest Publications
- Open Access
- 15 July 2026
- Published by Springer Nature in Nature Health
- Vol. 1 (11) , 1254-1270
Addressing benchmarking gaps in large language models for health and medicine with dynamic red-teaming
Large language models (LLMs) are increasingly used to answer health-related questions and support healthcare workflows, yet evidence for their safety still relies heavily on static benchmarks that can rapidly become obsolete or be optimized against. Here we introduce a Dynamic, Automatic and Systematic (DAS) red-teaming audit framework that continuously stress-tests LLMs for health across four safety-critical axes: robustness, privacy, bias and hallucination. Validated against board-certified clinicians, a suite of adversarial agents autonomously mutates health-related test cases to uncover vulnerabilities in real time. Applying DAS to 15 state-of-the-art LLMs revealed a profound gap between high static benchmark performance and low dynamic reliability—the ‘benchmarking gap’. Despite median MedQA accuracy exceeding 80%, 94% of previously correct answers failed under dynamic robustness testing. This brittleness generalized to the realistic, open-ended HealthBench dataset, where top-tier models exhibited failure rates exceeding 70%, suggesting that high scores on established static benchmarks may reflect superficial memorization. We observed similarly high failure rates across other domains: privacy leaks were elicited in 86% of scenarios, cognitive bias priming altered recommendations in 81% of fairness tests and hallucination rates exceeded 74% in widely used models. By converting LLM safety evaluation for health from a static checklist into a living adversarial audit, DAS provides a scalable framework for surfacing latent risks before such systems are deployed in consumer-facing health assistants and broader clinical workflows.
- 15 July 2026
- Published by Springer Nature in Nature Health
- Vol. 1 (11) , 1254-1270
Open AccessAddressing benchmarking gaps in large language models for health and medicine with dynamic red-teaming
Large language models (LLMs) are increasingly used to answer health-related questions and support healthcare workflows, yet evidence for their safety still relies heavily on static benchmarks that can rapidly become obsolete or be optimized against. Here we introduce a Dynamic, Automatic and Systematic (DAS) red-teaming audit framework that continuously stress-tests LLMs for health across four safety-critical axes: robustness, privacy, bias and hallucination. Validated against board-certified clinicians, a suite of adversarial agents autonomously mutates health-related test cases to uncover vulnerabilities in real time. Applying DAS to 15 state-of-the-art LLMs revealed a profound gap between high static benchmark performance and low dynamic reliability—the ‘benchmarking gap’. Despite median MedQA accuracy exceeding 80%, 94% of previously correct answers failed under dynamic robustness testing. This brittleness generalized to the realistic, open-ended HealthBench dataset, where top-tier models exhibited failure rates exceeding 70%, suggesting that high scores on established static benchmarks may reflect superficial memorization. We observed similarly high failure rates across other domains: privacy leaks were elicited in 86% of scenarios, cognitive bias priming altered recommendations in 81% of fairness tests and hallucination rates exceeded 74% in widely used models. By converting LLM safety evaluation for health from a static checklist into a living adversarial audit, DAS provides a scalable framework for surfacing latent risks before such systems are deployed in consumer-facing health assistants and broader clinical workflows.
- 13 February 2025
- Published by Elsevier in Biomaterials
- Vol. 319, 123180
Challenges and material innovations in drug delivery to central nervous system tumors
- 13 February 2025
- Published by Elsevier in Biomaterials
- Vol. 319, 123180
Challenges and material innovations in drug delivery to central nervous system tumors
- Open Access
- 17 July 2024
- Published by Springer Nature in Scientific Data
- Vol. 11 (1) , 1-6
A Multi-Center, Multi-Parametric MRI Dataset of Primary and Secondary Brain Tumors
Brain metastases (BMs) and high-grade gliomas (HGGs) are the most common and aggressive types of malignant brain tumors in adults, with often poor prognosis and short survival. As their clinical symptoms and image appearances on conventional magnetic resonance imaging (MRI) can be astonishingly similar, their accurate differentiation based solely on clinical and radiological information can be very challenging, particularly for “cancer of unknown primary”, where no systemic malignancy is known or found. Non-invasive multiparametric MRI and radiomics offer the potential to identify these distinct biological properties, aiding in the characterization and differentiation of HGGs and BMs. However, there is a scarcity of publicly available multi-origin brain tumor imaging data for tumor characterization. In this paper, we introduce a multi-center, multi-origin brain tumor MRI (MOTUM) imaging dataset obtained from 67 patients: 29 with high-grade gliomas, 20 with lung metastases, 10 with breast metastases, 2 with gastric metastasis, 4 with ovarian metastasis, and 2 with melanoma metastasis. This dataset includes anonymized DICOM files alongside processed FLAIR, T1-weighted, contrast-enhanced T1-weighted, T2-weighted sequences images, segmentation masks of two tumor regions, and clinical data. Our data-sharing initiative is to support the benchmarking of automated tumor segmentation, multi-modal machine learning, and disease differentiation of multi-origin brain tumors in a multi-center setting.
- 17 July 2024
- Published by Springer Nature in Scientific Data
- Vol. 11 (1) , 1-6
Open AccessA Multi-Center, Multi-Parametric MRI Dataset of Primary and Secondary Brain Tumors
Brain metastases (BMs) and high-grade gliomas (HGGs) are the most common and aggressive types of malignant brain tumors in adults, with often poor prognosis and short survival. As their clinical symptoms and image appearances on conventional magnetic resonance imaging (MRI) can be astonishingly similar, their accurate differentiation based solely on clinical and radiological information can be very challenging, particularly for “cancer of unknown primary”, where no systemic malignancy is known or found. Non-invasive multiparametric MRI and radiomics offer the potential to identify these distinct biological properties, aiding in the characterization and differentiation of HGGs and BMs. However, there is a scarcity of publicly available multi-origin brain tumor imaging data for tumor characterization. In this paper, we introduce a multi-center, multi-origin brain tumor MRI (MOTUM) imaging dataset obtained from 67 patients: 29 with high-grade gliomas, 20 with lung metastases, 10 with breast metastases, 2 with gastric metastasis, 4 with ovarian metastasis, and 2 with melanoma metastasis. This dataset includes anonymized DICOM files alongside processed FLAIR, T1-weighted, contrast-enhanced T1-weighted, T2-weighted sequences images, segmentation masks of two tumor regions, and clinical data. Our data-sharing initiative is to support the benchmarking of automated tumor segmentation, multi-modal machine learning, and disease differentiation of multi-origin brain tumors in a multi-center setting.
- Open Access
Universal Urban Spreading Pattern of COVID-19 and Its Underlying Mechanism
Currently, the global situation of COVID-19 is aggravating, pressingly calling for efficient control and prevention measures. Understanding spreading pattern of COVID-19 has been widely recognized as a vital step for implementing non-pharmaceutical measures. Previous studies investigated such an issue in large-scale (e.g., inter-country or inter-state) scenarios while urban spreading pattern still remains an open issue. Here, we fill this gap by leveraging the trajectory data of 197,808 smartphone users (including 17,808 anonymous confirmed cases) in 9 cities in China. We find a universal spreading pattern in all cities: the spatial distribution of confirmed cases follows a power-law-like model and the spreading centroid is time-invariant. Moreover, we reveal that human mobility in a city drives the spatialtemporal spreading process: long average travelling distance results in a high growth rate of spreading radius and wide spatial diffusion of confirmed cases. With such insight, we adopt Kendall model to simulate urban spreading of COVID-19 that can well fit the real spreading process. Our results unveil the underlying mechanism behind the spatial-temporal urban evolution of COVID-19, and can be used to evaluate the performance of mobility restriction policies implemented by many governments and to estimate the evolving spreading situation of COVID-19.
Open AccessUniversal Urban Spreading Pattern of COVID-19 and Its Underlying Mechanism
Currently, the global situation of COVID-19 is aggravating, pressingly calling for efficient control and prevention measures. Understanding spreading pattern of COVID-19 has been widely recognized as a vital step for implementing non-pharmaceutical measures. Previous studies investigated such an issue in large-scale (e.g., inter-country or inter-state) scenarios while urban spreading pattern still remains an open issue. Here, we fill this gap by leveraging the trajectory data of 197,808 smartphone users (including 17,808 anonymous confirmed cases) in 9 cities in China. We find a universal spreading pattern in all cities: the spatial distribution of confirmed cases follows a power-law-like model and the spreading centroid is time-invariant. Moreover, we reveal that human mobility in a city drives the spatialtemporal spreading process: long average travelling distance results in a high growth rate of spreading radius and wide spatial diffusion of confirmed cases. With such insight, we adopt Kendall model to simulate urban spreading of COVID-19 that can well fit the real spreading process. Our results unveil the underlying mechanism behind the spatial-temporal urban evolution of COVID-19, and can be used to evaluate the performance of mobility restriction policies implemented by many governments and to estimate the evolving spreading situation of COVID-19.
- Open Access
- 19 March 2020
- Published by Springer Nature in BMC Medical Genetics
- Vol. 21 (1) , 1-13
An eight-mRNA signature outperforms the lncRNA-based signature in predicting prognosis of patients with glioblastoma
Background: The prognosis of the glioblastoma (GBM) is dismal. This study aims to select an optimal RNA signature for prognostic prediction of GBM patients. Methods: For the training set, the long non-coding RNA (lncRNA) and mRNA expression profiles of 151 patients were downloaded from the TCGA. Differentially expressed mRNAs (DEGs) and lncRNAs (DE-lncRNAs) were identified between good prognosis and bad prognosis patients. Optimal prognostic mRNAs and lncRNAs were selected respectively, by using univariate Cox proportional-hazards (PH) regression model and LASSO Cox-PH model. Subsequently, four prognostic scoring models were built based on expression levels or expression status of the selected prognostic lncRNAs or mRNAs, separately. Each prognostic model was applied to the training set and an independent validation set. Function analysis was used to uncover the biological roles of these prognostic DEGs between different risk groups classified by the mRNA-based signature. Results: We obtained 261 DEGs and 33 DE-lncRNAs between good prognosis and bad prognosis patients. A panel of eight mRNAs and a combination of ten lncRNAs were determined as predictive RNAs by LASSO Cox-PH model. Among the four prognostic scoring models using the eight-mRNA signature or the ten-lncRNA signature, the one based on the expression levels of the eight mRNAs showed the greatest predictive power. The DEGs between different risk groups using the eight prognostic mRNAs were functionally involved in calcium signaling pathway, neuroactive ligand-receptor interaction pathway, and Wnt signaling pathway. Conclusion: The eight-mRNA signature has greater prognostic value than the ten-lncRNA-based signature for GBM patients based on bioinformatics analysis.
- 19 March 2020
- Published by Springer Nature in BMC Medical Genetics
- Vol. 21 (1) , 1-13
Open AccessAn eight-mRNA signature outperforms the lncRNA-based signature in predicting prognosis of patients with glioblastoma
Background: The prognosis of the glioblastoma (GBM) is dismal. This study aims to select an optimal RNA signature for prognostic prediction of GBM patients. Methods: For the training set, the long non-coding RNA (lncRNA) and mRNA expression profiles of 151 patients were downloaded from the TCGA. Differentially expressed mRNAs (DEGs) and lncRNAs (DE-lncRNAs) were identified between good prognosis and bad prognosis patients. Optimal prognostic mRNAs and lncRNAs were selected respectively, by using univariate Cox proportional-hazards (PH) regression model and LASSO Cox-PH model. Subsequently, four prognostic scoring models were built based on expression levels or expression status of the selected prognostic lncRNAs or mRNAs, separately. Each prognostic model was applied to the training set and an independent validation set. Function analysis was used to uncover the biological roles of these prognostic DEGs between different risk groups classified by the mRNA-based signature. Results: We obtained 261 DEGs and 33 DE-lncRNAs between good prognosis and bad prognosis patients. A panel of eight mRNAs and a combination of ten lncRNAs were determined as predictive RNAs by LASSO Cox-PH model. Among the four prognostic scoring models using the eight-mRNA signature or the ten-lncRNA signature, the one based on the expression levels of the eight mRNAs showed the greatest predictive power. The DEGs between different risk groups using the eight prognostic mRNAs were functionally involved in calcium signaling pathway, neuroactive ligand-receptor interaction pathway, and Wnt signaling pathway. Conclusion: The eight-mRNA signature has greater prognostic value than the ten-lncRNA-based signature for GBM patients based on bioinformatics analysis.