Hwee Kuan Lee
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Hwee Kuan Lee has a total of 257 co-authors. Here are the top co-authors by number of shared publications:
- Yutaka Okabe
Department of Physics Tokyo Metropolitan University 192-0397 Hachioji Tokyo Japan
7 shared publications - Connie Kou
Tencent, Singapore, Singapore
4 shared publications Danilo Medina Giron
Department of Pathology, Tan Tock Seng Hospital
4 shared publications- Malay Singh
Bioinformatics Institute, A*STAR, 30 Biopolis Street, Matrix, Singapore 138671, Singapore
4 shared publications - Andy M. Yip
Department of Mathematics, National University of Singapore, Singapore
3 shared publications Emarene Mationg Kalaw
UQ Centre for Clinical Research University of Queensland Brisbane Australia
3 shared publications
Latest Publications
- Open Access
- 29 July 2026
- Published by Springer Nature in npj Biomedical Innovations
- Vol. 3 (1) , 42
Rethinking pathology image analysis through shuffling
- 29 July 2026
- Published by Springer Nature in npj Biomedical Innovations
- Vol. 3 (1) , 42
Open AccessRethinking pathology image analysis through shuffling
- 28 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Medical Imaging
- Vol. PP (02780062) , 1
DiffBulk: Enhancing Spatial Transcriptomic Prediction with Diffusion-Based Training
Spatial Transcriptomics (ST) technology detects gene expression from tissue biopsies, playing an emerging role in cancer diagnosis and precision medicine. However, the high cost of ST technology limits its broader application. Recently, deep learning approaches have provided insight into predicting gene expression based on H&E-stained histopathology images. Nevertheless, the relationship between morphological features and gene expression is highly complex. To address these challenges, we propose DiffBulk, a novel two-stage framework that leverages conditional diffusion models to learn expressive image representations enriched with gene expression information. In the first stage, we introduce a gene-to-image conditional diffusion model equipped with a permutationinvariant open-embedding gene encoder, which enables unified training across diverse gene panels. In the second stage, diffusion-derived features are fused with representations from a pathology foundation model, effectively bridging the domain gap and improving downstream gene expression prediction. We evaluate DiffBulk on high-quality Xenium ST data curated from the HEST dataset and the CrunchDAO challenge, constructing tile-level pseudo-bulk datasets for training and evaluation. Extensive experiments demonstrate that DiffBulk consistently outperforms state-of-the-art baselines across all metrics for gene expression prediction. These findings highlight the potential of diffusion-based gene-image representation learning and suggest promising directions for future research.
- 28 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Medical Imaging
- Vol. PP (02780062) , 1
DiffBulk: Enhancing Spatial Transcriptomic Prediction with Diffusion-Based Training
Spatial Transcriptomics (ST) technology detects gene expression from tissue biopsies, playing an emerging role in cancer diagnosis and precision medicine. However, the high cost of ST technology limits its broader application. Recently, deep learning approaches have provided insight into predicting gene expression based on H&E-stained histopathology images. Nevertheless, the relationship between morphological features and gene expression is highly complex. To address these challenges, we propose DiffBulk, a novel two-stage framework that leverages conditional diffusion models to learn expressive image representations enriched with gene expression information. In the first stage, we introduce a gene-to-image conditional diffusion model equipped with a permutationinvariant open-embedding gene encoder, which enables unified training across diverse gene panels. In the second stage, diffusion-derived features are fused with representations from a pathology foundation model, effectively bridging the domain gap and improving downstream gene expression prediction. We evaluate DiffBulk on high-quality Xenium ST data curated from the HEST dataset and the CrunchDAO challenge, constructing tile-level pseudo-bulk datasets for training and evaluation. Extensive experiments demonstrate that DiffBulk consistently outperforms state-of-the-art baselines across all metrics for gene expression prediction. These findings highlight the potential of diffusion-based gene-image representation learning and suggest promising directions for future research.
- Open Access
- 21 January 2026
- Published by Elsevier in Computer Physics Communications
- Vol. 322
Scalable neural network driven molecular dynamics simulation
- 21 January 2026
- Published by Elsevier in Computer Physics Communications
- Vol. 322
Open AccessScalable neural network driven molecular dynamics simulation
- Open Access
Umbilical cord structure shapes feto-maternal heat transfer across mammals
The umbilical cord varies widely across placental mammals, yet the functional consequences of this structural diversity remain unclear. We compared cord anatomy across 130 mammal species, reconstructed ancestral states, and combined multimodal imaging with physics-based modelling of oxygen and heat transfer. Reconstructions indicate that the ancestral umbilical cord was uncoiled, with three vessels and an allantoic duct; the human combination of coiling with loss of the duct is rare. Across species, vessel number, duct persistence, and cord length increase with birth weight. Modelling predicts that oxygen delivery remains near maximal across architectures, but that placental heat loss depends more strongly on umbilical flow and on cord length and coiling, suggesting more efficient cooling in larger mammals. Microscopy and ultrasound imaging reveal geometric vessel layouts, including a human triple helix, consistent with configurations predicted to reduce inter-vessel heat shunting. As fetuses must dissipate heat through the mother, these results link cord morphology, evolution, and fetal thermal resilience.
Open AccessUmbilical cord structure shapes feto-maternal heat transfer across mammals
The umbilical cord varies widely across placental mammals, yet the functional consequences of this structural diversity remain unclear. We compared cord anatomy across 130 mammal species, reconstructed ancestral states, and combined multimodal imaging with physics-based modelling of oxygen and heat transfer. Reconstructions indicate that the ancestral umbilical cord was uncoiled, with three vessels and an allantoic duct; the human combination of coiling with loss of the duct is rare. Across species, vessel number, duct persistence, and cord length increase with birth weight. Modelling predicts that oxygen delivery remains near maximal across architectures, but that placental heat loss depends more strongly on umbilical flow and on cord length and coiling, suggesting more efficient cooling in larger mammals. Microscopy and ultrasound imaging reveal geometric vessel layouts, including a human triple helix, consistent with configurations predicted to reduce inter-vessel heat shunting. As fetuses must dissipate heat through the mother, these results link cord morphology, evolution, and fetal thermal resilience.
- Open Access
- 13 October 2025
- Published by Oxford University Press (OUP) in European Heart Journal - Digital Health
- Vol. 6 (6) , 1223-1233
Deep learning-based quantification of epicardial adipose tissue volume from non-contrast computed tomography images: a multi-centre study
Aims: Epicardial adipose tissue (EAT), located within the pericardial sac, has emerged as a biomarker for coronary artery disease (CAD) progression. This study aimed to develop and validate a deep learning-based system for automated EAT volume quantification using non-contrast computed tomography (NCCT) scans from a large, multi-centre, pan-Asian cohort. Methods and results: A total of 1243 NCCT patient scans from three centres were used to train and internally validate a deep learning model based on 3D UNet++ architecture for pericardium segmentation, followed by intensity thresholding to derive EAT volume. Epicardial adipose tissue quantification required ∼30 s per scan. The final model was evaluated on an external testing cohort of 160 patients, including 90 non-Asian individuals. In this cohort, AI-predicted EAT volumes showed excellent agreement with expert annotations (r = 0.975; P < 0.0001). The Bland–Altman analysis demonstrated a mean bias of −5.2 cm3with 95% limits of agreement from −25.1 to 14.7 cm3. Among the non-Asian subgroup, model performance remained strong (r = 0.970; bias, −3.2 cm3; limits of agreement, −25.1–18.7 cm3). AI-derived EAT volume was independently associated with obstructive CAD (odds ratio 1.11; 95% confidence interval, 1.04–1.19; P = 0.004), after adjusting for confounders. The global χ2 statistic increased from 81.7 with coronary calcium score alone to 93.3 when EAT volume was added (P = 0.001), indicating improved risk prediction. Conclusion: We developed and validated a deep learning system for automated EAT volume quantification from NCCT scans. The model demonstrated high accuracy and generalizability across ethnically diverse populations, supporting its potential for routine EAT assessment and CAD risk stratification. Trial Registration: ClinicalTrials.gov Identifier: NCT05509010.
- 13 October 2025
- Published by Oxford University Press (OUP) in European Heart Journal - Digital Health
- Vol. 6 (6) , 1223-1233
Open AccessDeep learning-based quantification of epicardial adipose tissue volume from non-contrast computed tomography images: a multi-centre study
Aims: Epicardial adipose tissue (EAT), located within the pericardial sac, has emerged as a biomarker for coronary artery disease (CAD) progression. This study aimed to develop and validate a deep learning-based system for automated EAT volume quantification using non-contrast computed tomography (NCCT) scans from a large, multi-centre, pan-Asian cohort. Methods and results: A total of 1243 NCCT patient scans from three centres were used to train and internally validate a deep learning model based on 3D UNet++ architecture for pericardium segmentation, followed by intensity thresholding to derive EAT volume. Epicardial adipose tissue quantification required ∼30 s per scan. The final model was evaluated on an external testing cohort of 160 patients, including 90 non-Asian individuals. In this cohort, AI-predicted EAT volumes showed excellent agreement with expert annotations (r = 0.975; P < 0.0001). The Bland–Altman analysis demonstrated a mean bias of −5.2 cm3with 95% limits of agreement from −25.1 to 14.7 cm3. Among the non-Asian subgroup, model performance remained strong (r = 0.970; bias, −3.2 cm3; limits of agreement, −25.1–18.7 cm3). AI-derived EAT volume was independently associated with obstructive CAD (odds ratio 1.11; 95% confidence interval, 1.04–1.19; P = 0.004), after adjusting for confounders. The global χ2 statistic increased from 81.7 with coronary calcium score alone to 93.3 when EAT volume was added (P = 0.001), indicating improved risk prediction. Conclusion: We developed and validated a deep learning system for automated EAT volume quantification from NCCT scans. The model demonstrated high accuracy and generalizability across ethnically diverse populations, supporting its potential for routine EAT assessment and CAD risk stratification. Trial Registration: ClinicalTrials.gov Identifier: NCT05509010.
- Open Access
- 28 August 2025
- Published by Springer Nature in Journal of Cheminformatics
- Vol. 17 (1) , 1-12
Systematic benchmarking of 13 AI methods for predicting cyclic peptide membrane permeability
Cyclic peptides are promising drug candidates due to their ability to modulate intracellular protein–protein interactions, a property often inaccessible to small molecules. However, their typically poor membrane permeability limits therapeutic applicability. Accurate computational prediction of permeability can accelerate the identification of cell-permeable candidates, reducing reliance on time-consuming and costly experimental screening. Although deep learning has shown potential in predicting molecular properties, its application in permeability prediction remains underexplored. A systematic evaluation of these models is important to assess current capabilities and guide future development. In this study, we conduct a comprehensive benchmark of 13 machine learning models for predicting cyclic peptide membrane permeability. These models cover four types of molecular representations: fingerprints, SMILES strings, molecular graphs, and 2D images. We use experimentally measured PAMPA permeability data from the CycPeptMPDB database, comprising nearly 6000 cyclic peptides, and evaluate performance across three prediction tasks: regression, binary classification, and soft-label classification. Two data-splitting strategies, random split and scaffold split, are used to assess the generalizability of trained models. Our results show that model performance depends strongly on molecular representation and model architecture. Graph-based models, particularly the Directed Message Passing Neural Network (DMPNN), consistently achieve top performance across tasks. Regression generally outperforms classification. Scaffold-based splitting, although intended to more rigorously assess generalization, yields substantially lower model generalizability compared to random splitting. Comparing prediction errors with experimental variability highlights the practical value of current models while also indicating room for further improvement. This study presents the first comprehensive benchmark of 13 machine learning models for predicting the membrane permeability of cyclic peptides. By evaluating multiple molecular representations and task formulations on a curated experimental dataset, the work reveals key performance trends, highlighting the superiority of graph-based models, the advantages of regression over classification, and the limitations of scaffold-based data splitting. These findings offer actionable insights for future model development and evaluation in cyclic peptide research.
- 28 August 2025
- Published by Springer Nature in Journal of Cheminformatics
- Vol. 17 (1) , 1-12
Open AccessSystematic benchmarking of 13 AI methods for predicting cyclic peptide membrane permeability
Cyclic peptides are promising drug candidates due to their ability to modulate intracellular protein–protein interactions, a property often inaccessible to small molecules. However, their typically poor membrane permeability limits therapeutic applicability. Accurate computational prediction of permeability can accelerate the identification of cell-permeable candidates, reducing reliance on time-consuming and costly experimental screening. Although deep learning has shown potential in predicting molecular properties, its application in permeability prediction remains underexplored. A systematic evaluation of these models is important to assess current capabilities and guide future development. In this study, we conduct a comprehensive benchmark of 13 machine learning models for predicting cyclic peptide membrane permeability. These models cover four types of molecular representations: fingerprints, SMILES strings, molecular graphs, and 2D images. We use experimentally measured PAMPA permeability data from the CycPeptMPDB database, comprising nearly 6000 cyclic peptides, and evaluate performance across three prediction tasks: regression, binary classification, and soft-label classification. Two data-splitting strategies, random split and scaffold split, are used to assess the generalizability of trained models. Our results show that model performance depends strongly on molecular representation and model architecture. Graph-based models, particularly the Directed Message Passing Neural Network (DMPNN), consistently achieve top performance across tasks. Regression generally outperforms classification. Scaffold-based splitting, although intended to more rigorously assess generalization, yields substantially lower model generalizability compared to random splitting. Comparing prediction errors with experimental variability highlights the practical value of current models while also indicating room for further improvement. This study presents the first comprehensive benchmark of 13 machine learning models for predicting the membrane permeability of cyclic peptides. By evaluating multiple molecular representations and task formulations on a curated experimental dataset, the work reveals key performance trends, highlighting the superiority of graph-based models, the advantages of regression over classification, and the limitations of scaffold-based data splitting. These findings offer actionable insights for future model development and evaluation in cyclic peptide research.
- Open Access
- 28 July 2025
- Published by Radcliffe Medical Media Ltd in Journal of Asian Pacific Society of Cardiology
- Vol. 4
Computational Methods for Analysing X-ray-guided Coronary Angiography: From Key Frame Selection to Vessel Segmentation, Stenosis Detection and Classification
Coronary artery disease (CAD) is a leading cause of mortality globally, highlighting the need for accurate diagnostic tools. X-ray-guided coronary angiography is the foundation for assessing CAD, involving three key tasks: key frame selection, vessel segmentation and stenosis detection. This review summarises conventional and advanced artificial intelligence (AI) methods applied to these tasks. AI’s integration is vital for enhancing efficiency and reproducibility, allowing real-time analysis and reducing manual labour. However, challenges remain, including the need for large, annotated datasets and difficulties in implementing AI systems in clinical trials. We discuss these challenges and the limitations of current AI approaches, exploring future directions for AI-driven solutions. Ultimately, we provide insights into AI’s potential to improve the precision and efficiency of coronary angiography analysis and CAD diagnosis, contributing to better clinical outcomes and advancing cardiovascular medicine.
- 28 July 2025
- Published by Radcliffe Medical Media Ltd in Journal of Asian Pacific Society of Cardiology
- Vol. 4
Open AccessComputational Methods for Analysing X-ray-guided Coronary Angiography: From Key Frame Selection to Vessel Segmentation, Stenosis Detection and Classification
Coronary artery disease (CAD) is a leading cause of mortality globally, highlighting the need for accurate diagnostic tools. X-ray-guided coronary angiography is the foundation for assessing CAD, involving three key tasks: key frame selection, vessel segmentation and stenosis detection. This review summarises conventional and advanced artificial intelligence (AI) methods applied to these tasks. AI’s integration is vital for enhancing efficiency and reproducibility, allowing real-time analysis and reducing manual labour. However, challenges remain, including the need for large, annotated datasets and difficulties in implementing AI systems in clinical trials. We discuss these challenges and the limitations of current AI approaches, exploring future directions for AI-driven solutions. Ultimately, we provide insights into AI’s potential to improve the precision and efficiency of coronary angiography analysis and CAD diagnosis, contributing to better clinical outcomes and advancing cardiovascular medicine.
- Open Access
- 26 February 2025
- Published by Oxford University Press (OUP) in Open Forum Infectious Diseases
- Vol. 12 (3) , ofaf095
Clinicodemographic and Radiological Features of Infective Ring-Enhancing Brain Lesions: A 4-Year Retrospective Study at a Tertiary Referral Center
Background: The diagnostic evaluation of ring-enhancing brain lesions (REBLs) is challenging, especially in immunocompromised patients. We conducted a retrospective study to describe the clinicodemographic and radiological features among patients presenting with REBLs to a tertiary referral center. Methods: Radiological reports of all patients who underwent brain computed tomography or magnetic resonance (MR) imaging between 1 November 2013 and 31 October 2017 were filtered for terms indicative of REBLs. Infectious diseases physicians reviewed the medical records to confirm the diagnosis. Results: Over the 4-year study period, there were 42 patients with infective REBLs and 249 with neoplastic REBLs. Pyogenic brain abscesses (PBAs) (20 of 42 [47.6%]) were the most common cause of infective REBLs, followed by tuberculous brain abscesses (TBAs) (9 of 42 [21.4%]) and Nocardia brain abscesses (NBAs) (6 of 42 [14.3%]). The patients were predominantly male, with a mean age of 55.2 years. Fewer than half were febrile at presentation. Cerebrospinal fluid investigations established the microbiological diagnosis in fewer than half of those who underwent lumbar puncture or extraventricular drain insertion. Conversely, brain biopsy yielded the microbiological diagnosis in almost all patients (16 of 17) who underwent the operation. Median white blood cell counts and C-reactive protein were higher in those with PBAs or NBAs than in those with TBAs. All with PBAs and NBAs who underwent MR imaging had diffusion-weighted imaging–hyperintense lesions, compared with only about half of those with TBAs. Conclusions: Our study has revealed important distinguishing features between infective REBLs and neoplastic REBLs and between PBAs, TBAs, and NBAs.
- 26 February 2025
- Published by Oxford University Press (OUP) in Open Forum Infectious Diseases
- Vol. 12 (3) , ofaf095
Open AccessClinicodemographic and Radiological Features of Infective Ring-Enhancing Brain Lesions: A 4-Year Retrospective Study at a Tertiary Referral Center
Background: The diagnostic evaluation of ring-enhancing brain lesions (REBLs) is challenging, especially in immunocompromised patients. We conducted a retrospective study to describe the clinicodemographic and radiological features among patients presenting with REBLs to a tertiary referral center. Methods: Radiological reports of all patients who underwent brain computed tomography or magnetic resonance (MR) imaging between 1 November 2013 and 31 October 2017 were filtered for terms indicative of REBLs. Infectious diseases physicians reviewed the medical records to confirm the diagnosis. Results: Over the 4-year study period, there were 42 patients with infective REBLs and 249 with neoplastic REBLs. Pyogenic brain abscesses (PBAs) (20 of 42 [47.6%]) were the most common cause of infective REBLs, followed by tuberculous brain abscesses (TBAs) (9 of 42 [21.4%]) and Nocardia brain abscesses (NBAs) (6 of 42 [14.3%]). The patients were predominantly male, with a mean age of 55.2 years. Fewer than half were febrile at presentation. Cerebrospinal fluid investigations established the microbiological diagnosis in fewer than half of those who underwent lumbar puncture or extraventricular drain insertion. Conversely, brain biopsy yielded the microbiological diagnosis in almost all patients (16 of 17) who underwent the operation. Median white blood cell counts and C-reactive protein were higher in those with PBAs or NBAs than in those with TBAs. All with PBAs and NBAs who underwent MR imaging had diffusion-weighted imaging–hyperintense lesions, compared with only about half of those with TBAs. Conclusions: Our study has revealed important distinguishing features between infective REBLs and neoplastic REBLs and between PBAs, TBAs, and NBAs.
- 3 December 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Systems, Man, and Cybernetics: Systems
- Vol. 55 (2) , 1374-1387
D-LORD for Motion Stylization
This article introduces a novel framework named double-latent optimization for representation disentanglement (D-LORD), which is designed for motion stylization (motion style transfer and motion retargeting). The primary objective of this framework is to separate the class and content information from a given motion sequence using a data-driven latent optimization approach. Here, class refers to person-specific style, such as a particular emotion or an individual’s identity, while content relates to the style-agnostic aspect of an action, such as walking or jumping, as universally understood concepts. The key advantage of D-LORD is its ability to perform style transfer without needing paired motion data. Instead, it utilizes class and content labels during the latent optimization process. By disentangling the representation, the framework enables the transformation of one motion sequence’s style to another’s style using adaptive instance normalization. The proposed D-LORD framework is designed with a focus on generalization, allowing it to handle different class and content labels for various applications. In addition, it can generate diverse motion sequences when specific class and content labels are provided. The framework’s efficacy is demonstrated through experimentation on three datasets: 1) the CMU XIA dataset for motion style transfer; 2) the multimodal human action database dataset; and 3) the RRIS Ability dataset for motion retargeting. Notably, this article presents the first generalized framework for motion style transfer and motion retargeting, showcasing its potential contributions in this area.
- 3 December 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Systems, Man, and Cybernetics: Systems
- Vol. 55 (2) , 1374-1387
D-LORD for Motion Stylization
This article introduces a novel framework named double-latent optimization for representation disentanglement (D-LORD), which is designed for motion stylization (motion style transfer and motion retargeting). The primary objective of this framework is to separate the class and content information from a given motion sequence using a data-driven latent optimization approach. Here, class refers to person-specific style, such as a particular emotion or an individual’s identity, while content relates to the style-agnostic aspect of an action, such as walking or jumping, as universally understood concepts. The key advantage of D-LORD is its ability to perform style transfer without needing paired motion data. Instead, it utilizes class and content labels during the latent optimization process. By disentangling the representation, the framework enables the transformation of one motion sequence’s style to another’s style using adaptive instance normalization. The proposed D-LORD framework is designed with a focus on generalization, allowing it to handle different class and content labels for various applications. In addition, it can generate diverse motion sequences when specific class and content labels are provided. The framework’s efficacy is demonstrated through experimentation on three datasets: 1) the CMU XIA dataset for motion style transfer; 2) the multimodal human action database dataset; and 3) the RRIS Ability dataset for motion retargeting. Notably, this article presents the first generalized framework for motion style transfer and motion retargeting, showcasing its potential contributions in this area.
- Open Access
Cohort profile: AI-driven national Platform for CCTA for clinicaL and industriaL applicatiOns (APOLLO)
Purpose Coronary CT angiography (CCTA) is well established for the diagnostic evaluation and prognostication of coronary artery disease (CAD). The growing burden of CAD in Asia and the emergence of novel CT-based risk markers highlight the need for an automated platform that integrates patient data with CCTA findings to provide tailored, accurate cardiovascular risk assessments. This study aims to develop an artificial intelligence (AI)-driven platform for CAD assessment using CCTA in Singapore’s multiethnic population. We will conduct a hybrid retrospective-prospective recruitment of patients who have undergone CCTA as part of the diagnostic workup for CAD, along with prospective follow-up for clinical endpoints. CCTA images will be analysed locally and by a core lab for coronary stenosis grading, Agatston scoring, epicardial adipose tissue evaluation and plaque analysis. The images and analyses will also be uploaded to an AI platform for deidentification, integration and automated reporting, generating precision AI toolkits for each parameter. Participants CCTA images and baseline characteristics have been collected and verified for 4196 recruited patients, comprising 75% Chinese, 6% Malay, 10% Indian and 9% from other ethnic groups. Among the participants, 41% are female, with a mean age of 55±11 years. Additionally, 41% have hypertension, 51% have dyslipidaemia, 15% have diabetes and 22% have a history of smoking. Findings to date The cohort data have been used to develop four AI modules for training, testing and validation. During the development process, data preprocessing standardised the format, resolution and other relevant attributes of the images. Future plans We will conduct prospective follow-up on the cohort to track clinical endpoints, including cardiovascular events, hospitalisations and mortality. Additionally, we will monitor the long-term impact of the AI-driven platform on patient outcomes and healthcare delivery. Trial registration number NCT05509010.
Open AccessCohort profile: AI-driven national Platform for CCTA for clinicaL and industriaL applicatiOns (APOLLO)
Purpose Coronary CT angiography (CCTA) is well established for the diagnostic evaluation and prognostication of coronary artery disease (CAD). The growing burden of CAD in Asia and the emergence of novel CT-based risk markers highlight the need for an automated platform that integrates patient data with CCTA findings to provide tailored, accurate cardiovascular risk assessments. This study aims to develop an artificial intelligence (AI)-driven platform for CAD assessment using CCTA in Singapore’s multiethnic population. We will conduct a hybrid retrospective-prospective recruitment of patients who have undergone CCTA as part of the diagnostic workup for CAD, along with prospective follow-up for clinical endpoints. CCTA images will be analysed locally and by a core lab for coronary stenosis grading, Agatston scoring, epicardial adipose tissue evaluation and plaque analysis. The images and analyses will also be uploaded to an AI platform for deidentification, integration and automated reporting, generating precision AI toolkits for each parameter. Participants CCTA images and baseline characteristics have been collected and verified for 4196 recruited patients, comprising 75% Chinese, 6% Malay, 10% Indian and 9% from other ethnic groups. Among the participants, 41% are female, with a mean age of 55±11 years. Additionally, 41% have hypertension, 51% have dyslipidaemia, 15% have diabetes and 22% have a history of smoking. Findings to date The cohort data have been used to develop four AI modules for training, testing and validation. During the development process, data preprocessing standardised the format, resolution and other relevant attributes of the images. Future plans We will conduct prospective follow-up on the cohort to track clinical endpoints, including cardiovascular events, hospitalisations and mortality. Additionally, we will monitor the long-term impact of the AI-driven platform on patient outcomes and healthcare delivery. Trial registration number NCT05509010.