Florian Leiser
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Florian Leiser has a total of 39 co-authors. Here are the top co-authors by number of shared publications:
- Ali Sunyaev
Technische Universität München, Heilbronn, Deutschland
12 shared publications - Scott Thiebes
School of Economics and Management, Tongji University
6 shared publications - Alexander Maedche
Human-Centered Systems Lab (h-lab), Karlsruhe Institute of Technology (KIT), Karlsruhe, Germany
3 shared publications - Gerhard Schwabe
University of Zurich, Zurich, Switzerland
3 shared publications - Konstantin D. Pandl
Institute of Applied Informatics and Formal Description Methods, Karlsruhe Institute of Technology, Karlsruhe, Germany
3 shared publications - Sven Eckhardt
University of Zürich
3 shared publications
Latest Publications
- Open Access
- 26 August 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Computing for Healthcare
Reward Systems for Trustworthy Medical Federated Learning
Federated learning (FL) is a popular technique to train machine learning (ML) models for healthcare where institutions collaborate by sharing ML model gradients. An undesired phenomenon in ML models is bias, which may cause unfairness against specific subgroups. In FL research, it remains unclear to which extent bias occurs, and how to best incentivize institutions contributing toward trustworthy FL models. Besides bias, trustworthiness aspects include high predictive performance. Existing reward systems, incentivizing predictive performance only, can transfer model bias against patients to an institutional level. Therefore, we evaluate how to measure the contributions of institutions toward either predictive performance or bias in FL and design corresponding reward systems, before we propose a combined reward system that incentivizes both. We evaluate our work using multiple chest X-ray datasets focusing on sex- and age-related patient subgroups. Our results show that we can successfully measure contributions toward bias, and an integrated reward system successfully incentivizes contributions toward a well-performing model with low bias. Thereby, institutions with data predominantly from one subgroup introduce a favorable bias for this subgroup. In a label flip experiment, we furthermore show that reward systems can incentivize institutions to ensure a high dataset quality.
- 26 August 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Computing for Healthcare
Open AccessReward Systems for Trustworthy Medical Federated Learning
Federated learning (FL) is a popular technique to train machine learning (ML) models for healthcare where institutions collaborate by sharing ML model gradients. An undesired phenomenon in ML models is bias, which may cause unfairness against specific subgroups. In FL research, it remains unclear to which extent bias occurs, and how to best incentivize institutions contributing toward trustworthy FL models. Besides bias, trustworthiness aspects include high predictive performance. Existing reward systems, incentivizing predictive performance only, can transfer model bias against patients to an institutional level. Therefore, we evaluate how to measure the contributions of institutions toward either predictive performance or bias in FL and design corresponding reward systems, before we propose a combined reward system that incentivizes both. We evaluate our work using multiple chest X-ray datasets focusing on sex- and age-related patient subgroups. Our results show that we can successfully measure contributions toward bias, and an integrated reward system successfully incentivizes contributions toward a well-performing model with low bias. Thereby, institutions with data predominantly from one subgroup introduce a favorable bias for this subgroup. In a label flip experiment, we furthermore show that reward systems can incentivize institutions to ensure a high dataset quality.
- Open Access
- 19 June 2025
- Published by JMIR Publications Inc. in Journal of Medical Internet Research
- Vol. 27, e70315
Large Language Model Architectures in Health Care: Scoping Review of Research Perspectives
Background: Large language models (LLMs) can support health care professionals in their daily work, for example, when writing and filing reports or communicating diagnoses. With the rise of LLMs, current research investigates how LLMs could be applied in medical practice and their benefits for physicians in clinical workflows. However, most studies neglect the importance of selecting suitable LLM architectures. Objective: In this literature review, we aim to provide insights on the different LLM model architecture families (ie, Bidirectional Encoder Representations from Transformers [BERT]–based or generative pretrained transformer [GPT]–based models) used in previous research. We report on the suitability and benefits of different LLM model architecture families for various research foci. Methods: To this end, we conduct a scoping review to identify which LLMs are used in health care. Our search included manuscripts from PubMed, arXiv, and medRxiv. We used open and selective coding to assess the 114 identified manuscripts regarding 11 dimensions related to usage and technical facets and the research focus of the manuscripts. Results: We identified 4 research foci that emerged previously in manuscripts, with LLM performance being the main focus. We found that GPT-based models are used for communicative purposes such as examination preparation or patient interaction. In contrast, BERT-based models are used for medical tasks such as knowledge discovery and model improvements. Conclusions: Our study suggests that GPT-based models are better suited for communicative purposes such as report generation or patient interaction. BERT-based models seem to be better suited for innovative applications such as classification or knowledge discovery. This could be due to the architectural differences where GPT processes language unidirectionally and BERT bidirectionally, allowing more in-depth understanding of the text. In addition, BERT-based models seem to allow more straightforward extensions of their models for domain-specific tasks that generally lead to better results. In summary, health care professionals should consider the benefits and differences of the LLM architecture families when selecting a suitable model for their intended purpose.
- 19 June 2025
- Published by JMIR Publications Inc. in Journal of Medical Internet Research
- Vol. 27, e70315
Open AccessLarge Language Model Architectures in Health Care: Scoping Review of Research Perspectives
Background: Large language models (LLMs) can support health care professionals in their daily work, for example, when writing and filing reports or communicating diagnoses. With the rise of LLMs, current research investigates how LLMs could be applied in medical practice and their benefits for physicians in clinical workflows. However, most studies neglect the importance of selecting suitable LLM architectures. Objective: In this literature review, we aim to provide insights on the different LLM model architecture families (ie, Bidirectional Encoder Representations from Transformers [BERT]–based or generative pretrained transformer [GPT]–based models) used in previous research. We report on the suitability and benefits of different LLM model architecture families for various research foci. Methods: To this end, we conduct a scoping review to identify which LLMs are used in health care. Our search included manuscripts from PubMed, arXiv, and medRxiv. We used open and selective coding to assess the 114 identified manuscripts regarding 11 dimensions related to usage and technical facets and the research focus of the manuscripts. Results: We identified 4 research foci that emerged previously in manuscripts, with LLM performance being the main focus. We found that GPT-based models are used for communicative purposes such as examination preparation or patient interaction. In contrast, BERT-based models are used for medical tasks such as knowledge discovery and model improvements. Conclusions: Our study suggests that GPT-based models are better suited for communicative purposes such as report generation or patient interaction. BERT-based models seem to be better suited for innovative applications such as classification or knowledge discovery. This could be due to the architectural differences where GPT processes language unidirectionally and BERT bidirectionally, allowing more in-depth understanding of the text. In addition, BERT-based models seem to allow more straightforward extensions of their models for domain-specific tasks that generally lead to better results. In summary, health care professionals should consider the benefits and differences of the LLM architecture families when selecting a suitable model for their intended purpose.
- 13 May 2025
- Published by Springer Nature
- p. 399-415
Applying Random Forests in Federated Learning: A Synthesis of Aggregation Techniques
- 13 May 2025
- Published by Springer Nature
- p. 399-415
Applying Random Forests in Federated Learning: A Synthesis of Aggregation Techniques
- Open Access
- 25 April 2025
- Published by Association for Computing Machinery (ACM)
- p. 1-18
TELL-ME: Toward Personalized Explanations of Large Language Models
- 25 April 2025
- Published by Association for Computing Machinery (ACM)
- p. 1-18
Open AccessTELL-ME: Toward Personalized Explanations of Large Language Models
- Open Access
- 19 December 2024
- Published by JMIR Publications Inc. in JMIR Preprints
Large Language Model Architectures in Health Care: Scoping Review of Research Perspectives (Preprint)
BACKGROUND Large language models (LLMs) can support health care professionals in their daily work, for example, when writing and filing reports or communicating diagnoses. With the rise of LLMs, current research investigates how LLMs could be applied in medical practice and their benefits for physicians in clinical workflows. However, most studies neglect the importance of selecting suitable LLM architectures. OBJECTIVE In this literature review, we aim to provide insights on the different LLM model architecture families (ie, Bidirectional Encoder Representations from Transformers [BERT]–based or generative pretrained transformer [GPT]–based models) used in previous research. We report on the suitability and benefits of different LLM model architecture families for various research foci. METHODS To this end, we conduct a scoping review to identify which LLMs are used in health care. Our search included manuscripts from PubMed, arXiv, and medRxiv. We used open and selective coding to assess the 114 identified manuscripts regarding 11 dimensions related to usage and technical facets and the research focus of the manuscripts. RESULTS We identified 4 research foci that emerged previously in manuscripts, with LLM performance being the main focus. We found that GPT-based models are used for communicative purposes such as examination preparation or patient interaction. In contrast, BERT-based models are used for medical tasks such as knowledge discovery and model improvements. CONCLUSIONS Our study suggests that GPT-based models are better suited for communicative purposes such as report generation or patient interaction. BERT-based models seem to be better suited for innovative applications such as classification or knowledge discovery. This could be due to the architectural differences where GPT processes language unidirectionally and BERT bidirectionally, allowing more in-depth understanding of the text. In addition, BERT-based models seem to allow more straightforward extensions of their models for domain-specific tasks that generally lead to better results. In summary, health care professionals should consider the benefits and differences of the LLM architecture families when selecting a suitable model for their intended purpose. CLINICALTRIAL
- 19 December 2024
- Published by JMIR Publications Inc. in JMIR Preprints
Open AccessLarge Language Model Architectures in Health Care: Scoping Review of Research Perspectives (Preprint)
BACKGROUND Large language models (LLMs) can support health care professionals in their daily work, for example, when writing and filing reports or communicating diagnoses. With the rise of LLMs, current research investigates how LLMs could be applied in medical practice and their benefits for physicians in clinical workflows. However, most studies neglect the importance of selecting suitable LLM architectures. OBJECTIVE In this literature review, we aim to provide insights on the different LLM model architecture families (ie, Bidirectional Encoder Representations from Transformers [BERT]–based or generative pretrained transformer [GPT]–based models) used in previous research. We report on the suitability and benefits of different LLM model architecture families for various research foci. METHODS To this end, we conduct a scoping review to identify which LLMs are used in health care. Our search included manuscripts from PubMed, arXiv, and medRxiv. We used open and selective coding to assess the 114 identified manuscripts regarding 11 dimensions related to usage and technical facets and the research focus of the manuscripts. RESULTS We identified 4 research foci that emerged previously in manuscripts, with LLM performance being the main focus. We found that GPT-based models are used for communicative purposes such as examination preparation or patient interaction. In contrast, BERT-based models are used for medical tasks such as knowledge discovery and model improvements. CONCLUSIONS Our study suggests that GPT-based models are better suited for communicative purposes such as report generation or patient interaction. BERT-based models seem to be better suited for innovative applications such as classification or knowledge discovery. This could be due to the architectural differences where GPT processes language unidirectionally and BERT bidirectionally, allowing more in-depth understanding of the text. In addition, BERT-based models seem to allow more straightforward extensions of their models for domain-specific tasks that generally lead to better results. In summary, health care professionals should consider the benefits and differences of the LLM architecture families when selecting a suitable model for their intended purpose. CLINICALTRIAL
- 26 July 2024
- Published by Springer Nature
- p. 253-261
How Experts Rely on Intuition in Medical Image Annotation—A Study Proposal
- 26 July 2024
- Published by Springer Nature
- p. 253-261
How Experts Rely on Intuition in Medical Image Annotation—A Study Proposal
- Open Access
- 31 May 2024
- Published by Association for Computing Machinery (ACM)
- p. 279-290
Knowledge-Guided Learning of Temporal Dynamics and its Application to Gas Turbines
Modeling dynamical systems is a fundamental task in scientific and engineering fields, often accomplished by applying theory-based models with mathematical equations. Yet, in cases where these equations cannot be established or parameterized properly, theory-based models are not applicable. Instead, a viable alternative is to learn the system dynamics directly from data, for example with deep learning models. However, traditional deep learning models often produce physically inconsistent results and struggle to generalize to unseen data, especially when training data is limited. One solution to this shortcoming is knowledge-guided deep learning, leveraging prior knowledge about the expected behavior of a dynamical system. In this work, we identify and formalize permissible system states, a novel type of prior knowledge that is often available for systems in the context of temporal dynamics modeling. This prior knowledge describes dynamic states that the system is allowed to take during its operation. We propose a knowledge-guided multi-state constraint to encode this type of prior knowledge through a loss function, making it applicable to any deep learning model. This approach allows to create an accurate data-driven model with minimal effort and data requirements. We validate the effectiveness of our method by applying it to model the temporal behavior of a gas turbine in response to an input control signal. Our results indicate that the proposed method reduces the prediction error by up to 40%. In addition to reducing the dependency on extensive training data, our method mitigates training randomness and enhances the consistency of predictions with the expected behavior.
- 31 May 2024
- Published by Association for Computing Machinery (ACM)
- p. 279-290
Open AccessKnowledge-Guided Learning of Temporal Dynamics and its Application to Gas Turbines
Modeling dynamical systems is a fundamental task in scientific and engineering fields, often accomplished by applying theory-based models with mathematical equations. Yet, in cases where these equations cannot be established or parameterized properly, theory-based models are not applicable. Instead, a viable alternative is to learn the system dynamics directly from data, for example with deep learning models. However, traditional deep learning models often produce physically inconsistent results and struggle to generalize to unseen data, especially when training data is limited. One solution to this shortcoming is knowledge-guided deep learning, leveraging prior knowledge about the expected behavior of a dynamical system. In this work, we identify and formalize permissible system states, a novel type of prior knowledge that is often available for systems in the context of temporal dynamics modeling. This prior knowledge describes dynamic states that the system is allowed to take during its operation. We propose a knowledge-guided multi-state constraint to encode this type of prior knowledge through a loss function, making it applicable to any deep learning model. This approach allows to create an accurate data-driven model with minimal effort and data requirements. We validate the effectiveness of our method by applying it to model the temporal behavior of a gas turbine in response to an input control signal. Our results indicate that the proposed method reduces the prediction error by up to 40%. In addition to reducing the dependency on extensive training data, our method mitigates training randomness and enhances the consistency of predictions with the expected behavior.
- Top Cited PapersOpen Access
- 11 May 2024
- Published by Association for Computing Machinery (ACM)
- p. 1-13
HILL: A Hallucination Identifier for Large Language Models
Large language models (LLMs) are prone to hallucinations, i.e., nonsensical, unfaithful, and undesirable text. Users tend to overrely on LLMs and corresponding hallucinations which can lead to misinterpretations and errors. To tackle the problem of overreliance, we propose HILL, the "Hallucination Identifier for Large Language Models". First, we identified design features for HILL with a Wizard of Oz approach with nine participants. Subsequently, we implemented HILL based on the identified design features and evaluated HILL’s interface design by surveying 17 participants. Further, we investigated HILL’s functionality to identify hallucinations based on an existing question-answering dataset and five user interviews. We find that HILL can correctly identify and highlight hallucinations in LLM responses which enables users to handle LLM responses with more caution. With that, we propose an easy-to-implement adaptation to existing LLMs and demonstrate the relevance of user-centered designs of AI artifacts.
- 11 May 2024
- Published by Association for Computing Machinery (ACM)
- p. 1-13
Top Cited PapersOpen AccessHILL: A Hallucination Identifier for Large Language Models
Large language models (LLMs) are prone to hallucinations, i.e., nonsensical, unfaithful, and undesirable text. Users tend to overrely on LLMs and corresponding hallucinations which can lead to misinterpretations and errors. To tackle the problem of overreliance, we propose HILL, the "Hallucination Identifier for Large Language Models". First, we identified design features for HILL with a Wizard of Oz approach with nine participants. Subsequently, we implemented HILL based on the identified design features and evaluated HILL’s interface design by surveying 17 participants. Further, we investigated HILL’s functionality to identify hallucinations based on an existing question-answering dataset and five user interviews. We find that HILL can correctly identify and highlight hallucinations in LLM responses which enables users to handle LLM responses with more caution. With that, we propose an easy-to-implement adaptation to existing LLMs and demonstrate the relevance of user-centered designs of AI artifacts.
- Open Access
HILL: A Hallucination Identifier for Large Language Models
Large language models (LLMs) are prone to hallucinations, i.e., nonsensical, unfaithful, and undesirable text. Users tend to overrely on LLMs and corresponding hallucinations which can lead to misinterpretations and errors. To tackle the problem of overreliance, we propose HILL, the "Hallucination Identifier for Large Language Models". First, we identified design features for HILL with a Wizard of Oz approach with nine participants. Subsequently, we implemented HILL based on the identified design features and evaluated HILL's interface design by surveying 17 participants. Further, we investigated HILL's functionality to identify hallucinations based on an existing question-answering dataset and five user interviews. We find that HILL can correctly identify and highlight hallucinations in LLM responses which enables users to handle LLM responses with more caution. With that, we propose an easy-to-implement adaptation to existing LLMs and demonstrate the relevance of user-centered designs of AI artifacts.
Open AccessHILL: A Hallucination Identifier for Large Language Models
Large language models (LLMs) are prone to hallucinations, i.e., nonsensical, unfaithful, and undesirable text. Users tend to overrely on LLMs and corresponding hallucinations which can lead to misinterpretations and errors. To tackle the problem of overreliance, we propose HILL, the "Hallucination Identifier for Large Language Models". First, we identified design features for HILL with a Wizard of Oz approach with nine participants. Subsequently, we implemented HILL based on the identified design features and evaluated HILL's interface design by surveying 17 participants. Further, we investigated HILL's functionality to identify hallucinations based on an existing question-answering dataset and five user interviews. We find that HILL can correctly identify and highlight hallucinations in LLM responses which enables users to handle LLM responses with more caution. With that, we propose an easy-to-implement adaptation to existing LLMs and demonstrate the relevance of user-centered designs of AI artifacts.
- Open Access
- 22 December 2023
- Published by JMIR Publications Inc. in JMIR Research Protocols
- Vol. 12 (1) , e48892
NephroCAGE—German-Canadian Consortium on AI for Improved Kidney Transplantation Outcome: Protocol for an Algorithm Development and Validation Study
Background: Recent advances in hardware and software enabled the use of artificial intelligence (AI) algorithms for analysis of complex data in a wide range of daily-life use cases. We aim to explore the benefits of applying AI to a specific use case in transplant nephrology: risk prediction for severe posttransplant events. For the first time, we combine multinational real-world transplant data, which require specific legal and technical protection measures. Objective: The German-Canadian NephroCAGE consortium aims to develop and evaluate specific processes, software tools, and methods to (1) combine transplant data of more than 8000 cases over the past decades from leading transplant centers in Germany and Canada, (2) implement specific measures to protect sensitive transplant data, and (3) use multinational data as a foundation for developing high-quality prognostic AI models. Methods: To protect sensitive transplant data addressing the first and second objectives, we aim to implement a decentralized NephroCAGE federated learning infrastructure upon a private blockchain. Our NephroCAGE federated learning infrastructure enables a switch of paradigms: instead of pooling sensitive data into a central database for analysis, it enables the transfer of clinical prediction models (CPMs) to clinical sites for local data analyses. Thus, sensitive transplant data reside protected in their original sites while the comparable small algorithms are exchanged instead. For our third objective, we will compare the performance of selected AI algorithms, for example, random forest and extreme gradient boosting, as foundation for CPMs to predict severe short- and long-term posttransplant risks, for example, graft failure or mortality. The CPMs will be trained on donor and recipient data from retrospective cohorts of kidney transplant patients. Results: We have received initial funding for NephroCAGE in February 2021. All clinical partners have applied for and received ethics approval as of 2022. The process of exploration of clinical transplant database for variable extraction has started at all the centers in 2022. In total, 8120 patient records have been retrieved as of August 2023. The development and validation of CPMs is ongoing as of 2023. Conclusions: For the first time, we will (1) combine kidney transplant data from nephrology centers in Germany and Canada, (2) implement federated learning as a foundation to use such real-world transplant data as a basis for the training of CPMs in a privacy-preserving way, and (3) develop a learning software system to investigate population specifics, for example, to understand population heterogeneity, treatment specificities, and individual impact on selected posttransplant outcomes. International Registered Report Identifier (IRRID): DERR1-10.2196/48892
- 22 December 2023
- Published by JMIR Publications Inc. in JMIR Research Protocols
- Vol. 12 (1) , e48892
Open AccessNephroCAGE—German-Canadian Consortium on AI for Improved Kidney Transplantation Outcome: Protocol for an Algorithm Development and Validation Study
Background: Recent advances in hardware and software enabled the use of artificial intelligence (AI) algorithms for analysis of complex data in a wide range of daily-life use cases. We aim to explore the benefits of applying AI to a specific use case in transplant nephrology: risk prediction for severe posttransplant events. For the first time, we combine multinational real-world transplant data, which require specific legal and technical protection measures. Objective: The German-Canadian NephroCAGE consortium aims to develop and evaluate specific processes, software tools, and methods to (1) combine transplant data of more than 8000 cases over the past decades from leading transplant centers in Germany and Canada, (2) implement specific measures to protect sensitive transplant data, and (3) use multinational data as a foundation for developing high-quality prognostic AI models. Methods: To protect sensitive transplant data addressing the first and second objectives, we aim to implement a decentralized NephroCAGE federated learning infrastructure upon a private blockchain. Our NephroCAGE federated learning infrastructure enables a switch of paradigms: instead of pooling sensitive data into a central database for analysis, it enables the transfer of clinical prediction models (CPMs) to clinical sites for local data analyses. Thus, sensitive transplant data reside protected in their original sites while the comparable small algorithms are exchanged instead. For our third objective, we will compare the performance of selected AI algorithms, for example, random forest and extreme gradient boosting, as foundation for CPMs to predict severe short- and long-term posttransplant risks, for example, graft failure or mortality. The CPMs will be trained on donor and recipient data from retrospective cohorts of kidney transplant patients. Results: We have received initial funding for NephroCAGE in February 2021. All clinical partners have applied for and received ethics approval as of 2022. The process of exploration of clinical transplant database for variable extraction has started at all the centers in 2022. In total, 8120 patient records have been retrieved as of August 2023. The development and validation of CPMs is ongoing as of 2023. Conclusions: For the first time, we will (1) combine kidney transplant data from nephrology centers in Germany and Canada, (2) implement federated learning as a foundation to use such real-world transplant data as a basis for the training of CPMs in a privacy-preserving way, and (3) develop a learning software system to investigate population specifics, for example, to understand population heterogeneity, treatment specificities, and individual impact on selected posttransplant outcomes. International Registered Report Identifier (IRRID): DERR1-10.2196/48892