Hang Dong
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
Research Subjects
Hang Dong has a total of 130 co-authors. Here are the top co-authors by number of shared publications:
- Jiaoyan Chen
The University of Manchester, Manchester, UK
14 shared publications - Ian Horrocks
University of Oxford
13 shared publications - Yuan He
University of Oxford
9 shared publications - Ernesto Jiménez-Ruiz
City St George’s, University of London
4 shared publications Frans Coenen
4 shared publications- Honghan Wu
Institute of Health Informatics, University College London, London, UK
4 shared publications
Latest Publications
- Open Access
- 10 September 2026
- Published by JMIR Publications Inc. in JMIR Research Protocols
- Vol. 15, e91348
Exploring Bias in Medical Applications of Large Language Models: Protocol for a Systematic Review
Background: Large language models (LLMs) are increasingly applied in health care for clinical decision-making, education, and patient communication. However, bias in LLM outputs may exacerbate health care disparities and compromise trust. Despite rapid adoption, there is limited synthesis of how bias is identified, measured, and mitigated in medical applications of LLMs. Objective: This systematic review aims to evaluate how bias is detected, measured, and mitigated in health care applications of LLMs and to identify methodological trends and gaps in the current literature. Methods: We will conduct a systematic review of studies evaluating bias, fairness, or subgroup performance in medical applications of LLMs. Searches will be conducted across Embase, MEDLINE, PsycINFO, PubMed, ACL Anthology, ACM Digital Library, arXiv, medRxiv, and bioRxiv for studies published from 2017 onward. The review uses a staged design. In phase 1, records identified in the original June 2025 search were screened and assessed manually using predefined eligibility criteria. This fully manually reviewed dataset will serve as a reference standard for validating the review workflow. In phase 2, an updated June 2026 search will be screened using a prespecified LLM-assisted workflow. Two independent LLMs will apply the same eligibility criteria used in manual screening, with studies marked as potentially relevant or uncertain by either model progressing to further assessment. Final inclusion decisions will be made by at least 2 human reviewers. The performance of the LLM-assisted workflow will be evaluated against the manually reviewed reference dataset using sensitivity, specificity, precision, negative predictive value, inclusion agreement, and Cohen κ. Data will be extracted using a standardized framework covering study characteristics, model details, health care use case, bias type, bias assessment methods, mitigation strategies, transparency, generalizability, and ethical or regulatory framing. Findings will be synthesized narratively. Results: The phase 1 search, conducted on June 11, 2025, yielded 15,976 records after deduplication and has undergone complete manual screening. The phase 2 search, conducted on June 8, 2026, increased the total number of records to 41,143. Screening of the updated dataset using the LLM-assisted workflow is ongoing. The completed review will report included study characteristics, approaches to bias detection and mitigation, and validation metrics for the LLM-assisted workflow. The review is expected to be submitted for publication in late November 2026, with publication anticipated thereafter subject to the journal’s peer-review and editorial process. Conclusions: This review aims to provide a comprehensive synthesis of current approaches to bias detection and mitigation in medical LLMs, highlighting methodological strengths, limitations, and areas for future research. The findings aim to inform the development of more equitable and transparent AI systems in health care.
- 10 September 2026
- Published by JMIR Publications Inc. in JMIR Research Protocols
- Vol. 15, e91348
Open AccessExploring Bias in Medical Applications of Large Language Models: Protocol for a Systematic Review
Background: Large language models (LLMs) are increasingly applied in health care for clinical decision-making, education, and patient communication. However, bias in LLM outputs may exacerbate health care disparities and compromise trust. Despite rapid adoption, there is limited synthesis of how bias is identified, measured, and mitigated in medical applications of LLMs. Objective: This systematic review aims to evaluate how bias is detected, measured, and mitigated in health care applications of LLMs and to identify methodological trends and gaps in the current literature. Methods: We will conduct a systematic review of studies evaluating bias, fairness, or subgroup performance in medical applications of LLMs. Searches will be conducted across Embase, MEDLINE, PsycINFO, PubMed, ACL Anthology, ACM Digital Library, arXiv, medRxiv, and bioRxiv for studies published from 2017 onward. The review uses a staged design. In phase 1, records identified in the original June 2025 search were screened and assessed manually using predefined eligibility criteria. This fully manually reviewed dataset will serve as a reference standard for validating the review workflow. In phase 2, an updated June 2026 search will be screened using a prespecified LLM-assisted workflow. Two independent LLMs will apply the same eligibility criteria used in manual screening, with studies marked as potentially relevant or uncertain by either model progressing to further assessment. Final inclusion decisions will be made by at least 2 human reviewers. The performance of the LLM-assisted workflow will be evaluated against the manually reviewed reference dataset using sensitivity, specificity, precision, negative predictive value, inclusion agreement, and Cohen κ. Data will be extracted using a standardized framework covering study characteristics, model details, health care use case, bias type, bias assessment methods, mitigation strategies, transparency, generalizability, and ethical or regulatory framing. Findings will be synthesized narratively. Results: The phase 1 search, conducted on June 11, 2025, yielded 15,976 records after deduplication and has undergone complete manual screening. The phase 2 search, conducted on June 8, 2026, increased the total number of records to 41,143. Screening of the updated dataset using the LLM-assisted workflow is ongoing. The completed review will report included study characteristics, approaches to bias detection and mitigation, and validation metrics for the LLM-assisted workflow. The review is expected to be submitted for publication in late November 2026, with publication anticipated thereafter subject to the journal’s peer-review and editorial process. Conclusions: This review aims to provide a comprehensive synthesis of current approaches to bias detection and mitigation in medical LLMs, highlighting methodological strengths, limitations, and areas for future research. The findings aim to inform the development of more equitable and transparent AI systems in health care.
- Open Access
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 12207-12218
RxGuard: Knowledge-Guided Safety Guardrails for Medication Recommendation
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 12207-12218
Open AccessRxGuard: Knowledge-Guided Safety Guardrails for Medication Recommendation
- Open Access
Real-Time World Crafting: Generating Structured Game Behaviors from Natural Language with Large Language Models
We present a novel architecture for safely integrating Large Language Models (LLMs) into interactive game engines, allowing players to "program" new behaviors using natural language. Our framework mitigates risks by using an LLM to translate commands into a constrained Domain-Specific Language (DSL), which configures a custom Entity-Component-System (ECS) at runtime. We evaluated this system in a 2D spell-crafting game prototype by experimentally assessing models from the Gemini, GPT, and Claude families with various prompting strategies. A validated LLM judge qualitatively rated the outputs, showing that while larger models better captured creative intent, the optimal prompting strategy is task-dependent: Chain-of-Thought improved creative alignment, while few-shot examples were necessary to generate more complex DSL scripts. This work offers a validated LLM-ECS pattern for emergent gameplay and a quantitative performance comparison for developers.
Open AccessReal-Time World Crafting: Generating Structured Game Behaviors from Natural Language with Large Language Models
We present a novel architecture for safely integrating Large Language Models (LLMs) into interactive game engines, allowing players to "program" new behaviors using natural language. Our framework mitigates risks by using an LLM to translate commands into a constrained Domain-Specific Language (DSL), which configures a custom Entity-Component-System (ECS) at runtime. We evaluated this system in a 2D spell-crafting game prototype by experimentally assessing models from the Gemini, GPT, and Claude families with various prompting strategies. A validated LLM judge qualitatively rated the outputs, showing that while larger models better captured creative intent, the optimal prompting strategy is task-dependent: Chain-of-Thought improved creative alignment, while few-shot examples were necessary to generate more complex DSL scripts. This work offers a validated LLM-ECS pattern for emergent gameplay and a quantitative performance comparison for developers.
- Open Access
- 8 August 2025
- Published by SAGE Publications in Intelligenza Artificiale
- Vol. 19 (2) , 102-115
Evaluating Inductive Reasoning and Programming Capabilities of Large Language Models With The One-Dimensional Abstract Reasoning Corpus
We present an initial automated test to evaluate LLMs’ capacity to perform inductive reasoning tasks. We use the GPT-3.5 and GPT-4 models to create a system which generates Python code as hypotheses for inductive reasoning to transform sequences of the One Dimensional Abstract Reasoning Corpus (1D-ARC) challenge. We experiment with three prompting techniques, namely standard prompting, Chain of Thought (CoT), and direct feedback. We provide results and an analysis of cost-to-success rate and benefit-cost ratio. Our best result is an overall 25% success rate with our CoT prompting on GPT-4, significantly surpassing the standard prompting approach. We assess the programming capabilities of the LLM by analysing the execution rate and errors of the generated code for inductive reasoning. We discuss potential avenues to improve our experiments, testing other strategies, and combining deductive reasoning with LLM-based inductive reasoning.
- 8 August 2025
- Published by SAGE Publications in Intelligenza Artificiale
- Vol. 19 (2) , 102-115
Open AccessEvaluating Inductive Reasoning and Programming Capabilities of Large Language Models With The One-Dimensional Abstract Reasoning Corpus
We present an initial automated test to evaluate LLMs’ capacity to perform inductive reasoning tasks. We use the GPT-3.5 and GPT-4 models to create a system which generates Python code as hypotheses for inductive reasoning to transform sequences of the One Dimensional Abstract Reasoning Corpus (1D-ARC) challenge. We experiment with three prompting techniques, namely standard prompting, Chain of Thought (CoT), and direct feedback. We provide results and an analysis of cost-to-success rate and benefit-cost ratio. Our best result is an overall 25% success rate with our CoT prompting on GPT-4, significantly surpassing the standard prompting approach. We assess the programming capabilities of the LLM by analysing the execution rate and errors of the generated code for inductive reasoning. We discuss potential avenues to improve our experiments, testing other strategies, and combining deductive reasoning with LLM-based inductive reasoning.
- Open Access
- 1 January 2025
- Published by Association for Computational Linguistics (ACL)
- p. 126-141
Evaluating Retrieval Augmented Generation to Communicate UK Climate Change Information
Arjun Biswas, Hatim Chahout, Tristan Pigram, Hang Dong, Hywel T.p. Williams, Fai Fung, Hailun Xie. Proceedings of the 2nd Workshop on Natural Language Processing Meets Climate Change (ClimateNLP 2025). 2025.
- 1 January 2025
- Published by Association for Computational Linguistics (ACL)
- p. 126-141
Open AccessEvaluating Retrieval Augmented Generation to Communicate UK Climate Change Information
Arjun Biswas, Hatim Chahout, Tristan Pigram, Hang Dong, Hywel T.p. Williams, Fai Fung, Hailun Xie. Proceedings of the 2nd Workshop on Natural Language Processing Meets Climate Change (ClimateNLP 2025). 2025.
- Open Access
- 16 October 2024
- Published by IOS Press
- p. 1389-1396
Ontology Text Alignment: Aligning Textual Content to Terminological Axioms
Despite the impressive advancements in Large Language Models (LLMs), their ability to perform reasoning and provide explainable outcomes remains a challenge, underscoring the continued relevance of ontologies in certain areas, particularly due to the reasoning and validation capabilities of ontologies. Ontology modelling and semantic search, due to their inherent complexity, still demand considerable human effort and expertise. Addressing this gap, our paper introduces the problem of ontology text alignment, which involves finding the most relevant axioms with respect to the given reference text. We propose an advanced Retrieval Augmented Generation framework that leverages BERT models and generative LLMs, together with ontology semantic enhancement based on atomic decomposition. Additionally, we have developed benchmarks in geology and biomedical areas. Our evaluation demonstrates the positive impact of our framework.
- 16 October 2024
- Published by IOS Press
- p. 1389-1396
Open AccessOntology Text Alignment: Aligning Textual Content to Terminological Axioms
Despite the impressive advancements in Large Language Models (LLMs), their ability to perform reasoning and provide explainable outcomes remains a challenge, underscoring the continued relevance of ontologies in certain areas, particularly due to the reasoning and validation capabilities of ontologies. Ontology modelling and semantic search, due to their inherent complexity, still demand considerable human effort and expertise. Addressing this gap, our paper introduces the problem of ontology text alignment, which involves finding the most relevant axioms with respect to the given reference text. We propose an advanced Retrieval Augmented Generation framework that leverages BERT models and generative LLMs, together with ontology semantic enhancement based on atomic decomposition. Additionally, we have developed benchmarks in geology and biomedical areas. Our evaluation demonstrates the positive impact of our framework.
- Open Access
- 1 October 2024
- Published by Association for Computing Machinery (ACM) in ACM Computing Surveys
- Vol. 56 (12) , 1-41
A Unified Review of Deep Learning for Automated Medical Coding
Automated medical coding, an essential task for healthcare operation and delivery, makes unstructured data manageable by predicting medical codes from clinical documents. Recent advances in deep learning and natural language processing have been widely applied to this task. However, deep learning–based medical coding lacks a unified view of the design of neural network architectures. This review proposes a unified framework to provide a general understanding of the building blocks of medical coding models and summarizes recent advanced models under the proposed framework. Our unified framework decomposes medical coding into four main components, i.e., encoder modules for text feature extraction, mechanisms for building deep encoder architectures, decoder modules for transforming hidden representations into medical codes, and the usage of auxiliary information. Finally, we introduce the benchmarks and real-world usage and discuss key research challenges and future directions.
- 1 October 2024
- Published by Association for Computing Machinery (ACM) in ACM Computing Surveys
- Vol. 56 (12) , 1-41
Open AccessA Unified Review of Deep Learning for Automated Medical Coding
Automated medical coding, an essential task for healthcare operation and delivery, makes unstructured data manageable by predicting medical codes from clinical documents. Recent advances in deep learning and natural language processing have been widely applied to this task. However, deep learning–based medical coding lacks a unified view of the design of neural network architectures. This review proposes a unified framework to provide a general understanding of the building blocks of medical coding models and summarizes recent advanced models under the proposed framework. Our unified framework decomposes medical coding into four main components, i.e., encoder modules for text feature extraction, mechanisms for building deep encoder architectures, decoder modules for transforming hidden representations into medical codes, and the usage of auxiliary information. Finally, we introduce the benchmarks and real-world usage and discuss key research challenges and future directions.
- Open Access
- 13 September 2024
- Published by Oxford University Press (OUP) in Journal of the American Medical Informatics Association
- Vol. 31 (10) , 2284-2293
Can GPT-3.5 generate and code discharge summaries?
Objectives: The aim of this study was to investigate GPT-3.5 in generating and coding medical documents with International Classification of Diseases (ICD)-10 codes for data augmentation on low-resource labels. Materials and Methods: Employing GPT-3.5 we generated and coded 9606 discharge summaries based on lists of ICD-10 code descriptions of patients with infrequent (or generation) codes within the MIMIC-IV dataset. Combined with the baseline training set, this formed an augmented training set. Neural coding models were trained on baseline and augmented data and evaluated on an MIMIC-IV test set. We report micro- and macro-F1 scores on the full codeset, generation codes, and their families. Weak Hierarchical Confusion Matrices determined within-family and outside-of-family coding errors in the latter codesets. The coding performance of GPT-3.5 was evaluated on prompt-guided self-generated data and real MIMIC-IV data. Clinicians evaluated the clinical acceptability of the generated documents. Results: Data augmentation results in slightly lower overall model performance but improves performance for the generation candidate codes and their families, including 1 absent from the baseline training data. Augmented models display lower out-of-family error rates. GPT-3.5 identifies ICD-10 codes by their prompted descriptions but underperforms on real data. Evaluators highlight the correctness of generated concepts while suffering in variety, supporting information, and narrative. Discussion and Conclusion: While GPT-3.5 alone given our prompt setting is unsuitable for ICD-10 coding, it supports data augmentation for training neural models. Augmentation positively affects generation code families but mainly benefits codes with existing examples. Augmentation reduces out-of-family errors. Documents generated by GPT-3.5 state prompted concepts correctly but lack variety, and authenticity in narratives.
- 13 September 2024
- Published by Oxford University Press (OUP) in Journal of the American Medical Informatics Association
- Vol. 31 (10) , 2284-2293
Open AccessCan GPT-3.5 generate and code discharge summaries?
Objectives: The aim of this study was to investigate GPT-3.5 in generating and coding medical documents with International Classification of Diseases (ICD)-10 codes for data augmentation on low-resource labels. Materials and Methods: Employing GPT-3.5 we generated and coded 9606 discharge summaries based on lists of ICD-10 code descriptions of patients with infrequent (or generation) codes within the MIMIC-IV dataset. Combined with the baseline training set, this formed an augmented training set. Neural coding models were trained on baseline and augmented data and evaluated on an MIMIC-IV test set. We report micro- and macro-F1 scores on the full codeset, generation codes, and their families. Weak Hierarchical Confusion Matrices determined within-family and outside-of-family coding errors in the latter codesets. The coding performance of GPT-3.5 was evaluated on prompt-guided self-generated data and real MIMIC-IV data. Clinicians evaluated the clinical acceptability of the generated documents. Results: Data augmentation results in slightly lower overall model performance but improves performance for the generation candidate codes and their families, including 1 absent from the baseline training data. Augmented models display lower out-of-family error rates. GPT-3.5 identifies ICD-10 codes by their prompted descriptions but underperforms on real data. Evaluators highlight the correctness of generated concepts while suffering in variety, supporting information, and narrative. Discussion and Conclusion: While GPT-3.5 alone given our prompt setting is unsuitable for ICD-10 coding, it supports data augmentation for training neural models. Augmentation positively affects generation code families but mainly benefits codes with existing examples. Augmentation reduces out-of-family errors. Documents generated by GPT-3.5 state prompted concepts correctly but lack variety, and authenticity in narratives.
- 19 May 2024
- Published by Springer Nature
- p. 79-99
A Language Model Based Framework for New Concept Placement in Ontologies
- 19 May 2024
- Published by Springer Nature
- p. 79-99
A Language Model Based Framework for New Concept Placement in Ontologies