publications

2025

1. TVST

   Advancing question-answering in ophthalmology with retrieval-Augmented Generation: Benchmarking open-source and proprietary large language models

   Quang Nguyen, Duy-Anh Nguyen, Khang Dang, and 9 more authors

   Transl. Vis. Sci. Technol., Sep 2025

   Bib

   @article{Nguyen2025-sa,
     title = {Advancing question-answering in ophthalmology with
                    retrieval-Augmented Generation: Benchmarking open-source and
                    proprietary large language models},
     author = {Nguyen, Quang and Nguyen, Duy-Anh and Dang, Khang and Liu, Siyin and Wang, Sophia Y and Woof, William A and Thomas, Peter B M and Patel, Praveen J and Balaskas, Konstantinos and Thygesen, Johan H and Wu, Honghan and Pontikos, Nikolas},
     journal = {Transl. Vis. Sci. Technol.},
     publisher = {The Association for Research in Vision and Ophthalmology},
     volume = {14},
     number = {9},
     pages = {18},
     month = sep,
     year = {2025},
     keywords = {benchmarking; ophthalmology; large language models;
                    retrieval-augmented generation; llamas; datasets},
     language = {en}
   }

2. Comput. Biol. Med.

   Infusing clinical knowledge into language models by subword optimisation and embedding initialisation

   Abul Hasan, Jinge Wu, Quang Ngoc Nguyen, and 7 more authors

   Comput. Biol. Med., Sep 2025

   Abs

   OBJECTIVE: This study introduces a novel tokenisation methodology, K-Tokeniser, to infuse clinical knowledge into language models for clinical text processing. METHODS: Technically, at initialisation stage, K-Tokeniser populates global representations of tokens based on semantic types of domain concepts (such as drugs or diseases) from either a domain ontology like Unified Medical Language System or the training data of the task related corpus. At training or inference stage, sentence level localised context will be utilised for choosing the optimal global token representation to realise the semantic-based tokenisation. To avoid pretraining using the new tokeniser, an embedding initialisation approach is proposed to generate representations for new tokens. RESULTS: Using three transformer-based language models, a comprehensive set of experiments are conducted on four real-world datasets for evaluating K-Tokeniser in a wide range of clinical text analytics tasks including clinical concept and relation extraction, automated clinical coding, clinical phenotype identification, and clinical research article classification. Overall, our models demonstrate consistent improvements over their counterparts in all tasks. In particular, substantial improvements are observed in the automated clinical coding task with 13% increase on Micro F1 score. Furthermore, K-Tokeniser also shows significant capacities in facilitating quicker convergence of language models. CONCLUSION: Models built using K-Tokeniser have shown faster convergence. Specifically,the language models would only require 50% of the training data to achieve the best performance of the baseline tokeniser using all training data in the concept extraction task and less than 20% of the data for the automated coding task. It is worth mentioning that all these improvements require no pre-training process, making the approach generalisable. Code availability: Our full implementation is openly available at https://github.com/abulhasanbbk/K-Tokenizer.

3. arXiv

   LEME: Open Large Language Models for Ophthalmology with Advanced Reasoning and Clinical Validation

   Hyunjae Kim, Xuguang Ai, Sahana Srinivasan, and 27 more authors

   2025

2024

1. AMIA

   Addressing generalizability in clinical Named Entity Recognition: Federated Learning or Large Language Models?: A case study on Visual Acuity extraction from US and UK eye institutes

   Quang N Nguyen, Honghan Wu, Nikolas Pontikos, and 1 more author

   AMIA Annu. Symp. Proc., 2024

   Abs Bib

   Clinical Named Entity Recognition (NER) is vital for extracting structured data from clinical text, but ensuring model generalizability across institutions remains challenging. This study compares two approaches: (1) Federated Learning (FL), a privacy-preserving decentralized method, and (2) Large Language Models (LLMs) trained on diverse corpora. We evaluate Visual Acuity (VA) extraction from ophthalmology notes at Stanford (USA) and Moorfields Eye Hospital (UK), using BERT-based models, FL strategies (FedAvg, STWT), and LLMs (Llama-3-70B, Mixtral-8x7B). Results show that FL significantly improves generalization, with STWT outperforming FedAvg in stability and accuracy. LLMs demonstrate strong performance on MEH data but struggle with structured Stanford notes. These findings highlight FL’s effectiveness for cross-institutional learning while revealing domain-specific limitations of LLMs, underscoring the need for tailored approaches to clinical NER.

   @article{Nguyen2024-sc,
     title = {Addressing generalizability in clinical Named Entity Recognition:
                   Federated Learning or Large Language Models?: A case study on
                   Visual Acuity extraction from {US} and {UK} eye institutes},
     author = {Nguyen, Quang N and Wu, Honghan and Pontikos, Nikolas and Wang, Sophia Y},
     journal = {AMIA Annu. Symp. Proc.},
     volume = {2024},
     pages = {949--958},
     year = {2024},
     language = {en}
   }

2. ACL

   End-to-End Relation Extraction of Pharmacokinetic Estimates from the Scientific Literature

   Ferran Gonzalez Hernandez, Victoria Smith, Quang Nguyen, and 9 more authors

   In Proceedings of the 23rd Workshop on Biomedical Natural Language Processing, Aug 2024

   Abs

   The lack of comprehensive and standardised databases containing Pharmacokinetic (PK) parameters presents a challenge in the drug development pipeline. Efficiently managing the increasing volume of published PK Parameters requires automated approaches that centralise information from diverse studies. In this work, we present the Pharmacokinetic Relation Extraction Dataset (PRED), a novel, manually curated corpus developed by pharmacometricians and NLP specialists, covering multiple types of PK parameters and numerical expressions reported in open-access scientific articles. PRED covers annotations for various entities and relations involved in PK parameter measurements from 3,600 sentences. We also introduce an end-to-end relation extraction model based on BioBERT, which is trained with joint named entity recognition (NER) and relation extraction objectives. The optimal pipeline achieved a micro-average F1-score of 94% for NER and over 85% F1-score across all relation types. This work represents the first resource for training and evaluating models for PK end-to-end extraction across multiple parameters and study types. We make our corpus and model openly available to accelerate the construction of large PK databases and to support similar endeavours in other scientific disciplines.

3. Sci. Rep.

   Named entity recognition of pharmacokinetic parameters in the scientific literature

   Ferran Gonzalez Hernandez, Quang Nguyen, Victoria C Smith, and 10 more authors

   Sci. Rep., Oct 2024

   Abs

   The development of accurate predictions for a new drug’s absorption, distribution, metabolism, and excretion profiles in the early stages of drug development is crucial due to high candidate failure rates. The absence of comprehensive, standardised, and updated pharmacokinetic (PK) repositories limits pre-clinical predictions and often requires searching through the scientific literature for PK parameter estimates from similar compounds. While text mining offers promising advancements in automatic PK parameter extraction, accurate Named Entity Recognition (NER) of PK terms remains a bottleneck due to limited resources. This work addresses this gap by introducing novel corpora and language models specifically designed for effective NER of PK parameters. Leveraging active learning approaches, we developed an annotated corpus containing over 4000 entity mentions found across the PK literature on PubMed. To identify the most effective model for PK NER, we fine-tuned and evaluated different NER architectures on our corpus. Fine-tuning BioBERT exhibited the best results, achieving a strict F 1 score of 90.37% in recognising PK parameter mentions, significantly outperforming heuristic approaches and models trained on existing corpora. To accelerate the development of end-to-end PK information extraction pipelines and improve pre-clinical PK predictions, the PK NER models and the labelled corpus were released open source at https://github.com/PKPDAI/PKNER .

4. ESCRS

   Posterior Capsule Opacification: Analysing 130,000 Cataract Surgeries With Natural Language Processing AI To Determine Incidence And Risk Factors

   Ken Kawamoto, Joshua Luis, Quang Nguyen, Laxmi Raja, Alex Ionides

   2024

2023

1. BMJ Open

   Can artificial intelligence accelerate the diagnosis of inherited retinal diseases? Protocol for a data-only retrospective cohort study (Eye2Gene)

   Quang Nguyen, William Woof, Nathaniel Kabiri, and 25 more authors

   BMJ Open, Mar 2023

   Abs

   INTRODUCTION: Inherited retinal diseases (IRD) are a leading cause of visual impairment and blindness in the working age population. Mutations in over 300 genes have been found to be associated with IRDs and identifying the affected gene in patients by molecular genetic testing is the first step towards effective care and patient management. However, genetic diagnosis is currently slow, expensive and not widely accessible. The aim of the current project is to address the evidence gap in IRD diagnosis with an AI algorithm, Eye2Gene, to accelerate and democratise the IRD diagnosis service. METHODS AND ANALYSIS: The data-only retrospective cohort study involves a target sample size of 10 000 participants, which has been derived based on the number of participants with IRD at three leading UK eye hospitals: Moorfields Eye Hospital (MEH), Oxford University Hospital (OUH) and Liverpool University Hospital (LUH), as well as a Japanese hospital, the Tokyo Medical Centre (TMC). Eye2Gene aims to predict causative genes from retinal images of patients with a diagnosis of IRD. For this purpose, 36 most common causative IRD genes have been selected to develop a training dataset for the software to have enough examples for training and validation for detection of each gene. The Eye2Gene algorithm is composed of multiple deep convolutional neural networks, which will be trained on MEH IRD datasets, and externally validated on OUH, LUH and TMC. ETHICS AND DISSEMINATION: This research was approved by the IRB and the UK Health Research Authority (Research Ethics Committee reference 22/WA/0049) ’Eye2Gene: accelerating the diagnosis of IRDs’ Integrated Research Application System (IRAS) project ID: 242050. All research adhered to the tenets of the Declaration of Helsinki. Findings will be reported in an open-access format.

2022

1. IFMBE Proceedings

   Automatic foveal avascular zone segmentation using hessian-based filter and U-net deep learning network

   Quang Ngoc Nguyen, Vinh Tuong-Quang Nguyen, Tammy Hsu, and 2 more authors

   In IFMBE Proceedings, 2022

   Abs

   Accurate segmentation of Foveal Avascular Zone (FAZ) in Optical Coherence Tomography Angiography (OCTA) images is important for OCTA images analysis. In this work, we developed an algorithm for automatic segmentation of FAZ in OCTA images using a Hessian-based filter and an U-Net deep learning network. A total of 260 OCTA images were used to train and test the algorithm. The images were first enhanced by a Hessian-based filter and then fed into a U-Net deep learning network. Eighty percent of the dataset was used for training and twenty percent was used for testing. Our method achieved 87.8% Jaccard Index (Intersection over Union metric) with 6% false-negative error and 5% false-positive error. The results showed that U-Net deep learning network could achieve good accuracy in automatically segmenting FAZ in OCTA images despite a small training set. The study also showed that image preprocessing techniques such as Hessian-based filtering helped to improve accuracy of U-Net deep learning network.

2015

1. EMBC

   Fractals properties of EEG during event-related desynchronization of motor imagery

   Ngoc Quang Nguyen, Quang Dang Khoa Truong, and Toshiyuki Kondo

   Annu. Int. Conf. IEEE Eng. Med. Biol. Soc., 2015

   Abs Bib

   Chaos and fractal dimension are emerging modalities for the research of electroencephalogram (EEG) signal processing. The capability of measuring non-linear characteristics of the fractal dimension enables new methodologies to identify distinct brain activities. Recent studies on the topic focus on utilizing various types of fractals as features in order to design better brain state classification system. However, we have little insight about the EEG signals projected in fractal dimension. In this paper, we investigate the relationship between the non-linear characteristics of ongoing EEG signals and event-related desynchronization (ERD) during motor imagery. We observed a considerable synchronization between ERD and fractal dimension. This finding suggests further usage of chaos and fractal theory in investigating brain activities.

   @article{Nguyen2015-nl,
     title = {Fractals properties of {EEG} during event-related
                    desynchronization of motor imagery},
     author = {Nguyen, Ngoc Quang and Truong, Quang Dang Khoa and Kondo, Toshiyuki},
     journal = {Annu. Int. Conf. IEEE Eng. Med. Biol. Soc.},
     publisher = {IEEE},
     volume = {2015},
     pages = {4146--4149},
     year = {2015},
     language = {en},
     projects = {FractalsEEG}
   }