Duke-NUS Deep Learning Breakthrough Shows Prognostic Value in Heart Disease Prognosis

Singapore's AI Revolution in Cardiovascular Imaging from NHCS and Duke-NUS

  • research-publication-news
  • duke-nus
  • singapore-research
  • deep-learning
  • ai-cardiology

Be the first to comment on this article!

You

Please keep comments respectful and on-topic.

a tall building with a sign on top of it
Photo by Chunjiang on Unsplash

Promote Your Research… Share it Worldwide

Have a story or a research paper to share? Become a contributor and publish your work on AcademicJobs.com.

Submit your Research - Make it Global News

The Dawn of AI-Driven Precision in Heart Attack Prognosis

A groundbreaking collaboration between Duke-NUS Medical School and the National Heart Centre Singapore (NHCS) has introduced LGE-CMRnet, an innovative end-to-end deep learning framework that automates the assessment of myocardial scar and microvascular obstruction (MVO) using late gadolinium enhancement cardiovascular magnetic resonance (LGE CMR) imaging. This advancement promises to transform post-acute myocardial infarction (AMI) risk stratification by providing rapid, accurate prognostic insights previously limited by manual analysis.

Acute myocardial infarction, commonly known as a heart attack, remains a leading cause of mortality worldwide, claiming 23 lives daily in Singapore alone according to recent health ministry data. Traditional LGE CMR, the gold standard for visualizing myocardial scar tissue and MVO—key indicators of adverse outcomes—relies on labor-intensive manual segmentation, prone to variability and delays. LGE-CMRnet addresses these challenges head-on, processing images in mere 0.05 seconds with expert-level prognostic accuracy.

Understanding the Clinical Challenge: Scar and MVO in AMI

Myocardial scar forms when heart muscle tissue dies following an AMI due to prolonged ischemia, while MVO represents areas of microvascular injury preventing reperfusion. Both are strong predictors of major adverse cardiac events (MACE), including recurrent infarction, heart failure, and mortality. Quantifying their extent via LGE CMR involves identifying hyperenhanced regions on images after gadolinium contrast administration.

Step-by-step, the process traditionally requires: (1) localizing the left ventricle, (2) segmenting healthy myocardium, (3) delineating scar (persistent enhancement), and (4) isolating MVO (hypointense cores within scar). Inter-observer variability can exceed 20%, and analysis times span 20-30 minutes per patient, hindering routine use in busy clinics like NHCS, Singapore's national referral center for complex cardiac cases.

In Singapore, where cardiovascular disease accounts for over 30% of deaths and incidence is projected to surge 150% in Southeast Asia by 2050 per NUS studies, scalable tools are critical. Duke-NUS, Singapore's premier graduate medical school, and NHCS have long championed translational research to combat this burden.

LGE-CMRnet: Engineering an End-to-End Solution

Developed by a multidisciplinary team led by researchers from NHCS's Cardiovascular Systems Imaging and Artificial Intelligence (CVS.AI) core and Duke-NUS's Cardiovascular & Metabolic Disorders Programme, LGE-CMRnet integrates YOLOv8 for automated heart localization and nnU-Net for simultaneous segmentation of myocardium, scar, and MVO.

Trained on 3,874 LGE images from 567 AMI patients (409 internal, 158 external validation), the model leverages vast NHCS datasets, reflecting Singapore's multi-ethnic population. This end-to-end pipeline eliminates manual preprocessing, delivering outputs as percentages of left ventricular mass (%Scar, %MVO)—clinically intuitive metrics for prognosis.

Diagram of LGE-CMRnet end-to-end deep learning pipeline for myocardial scar and MVO segmentation

Unmatched Accuracy and Speed: Benchmarking Against Experts

External validation yielded Dice similarity coefficients (DSC)—a measure of segmentation overlap—of 0.83±0.11 for scar and 0.88±0.11 for MVO, with volumetric correlations r=0.90 (scar) and r=0.98 (MVO) to expert tracings. Bland-Altman plots confirmed minimal bias (scar: 2.5±8.9 cm³; MVO: 0.20±0.89 cm³), positioning LGE-CMRnet as non-inferior for clinical deployment.

Processing at 0.05 seconds per image versus 20+ minutes manually scales effortlessly for high-volume centers like NHCS, which performs thousands of CMRs annually. This efficiency could reduce reporting backlogs, enabling same-day prognostication.

Prognostic Power: Predicting MACE with AI Precision

In the 158-patient external cohort (median follow-up 24.4 months), 22.2% suffered MACE. Cox regression revealed LGE-CMRnet-derived %MVO (HR 1.06, 95% CI 1.02-1.09, P=0.003) and %Scar (HR 1.05, 95% CI 1.02-1.08, P=0.001) as independent predictors, outperforming traditional models after adjusting for age, ejection fraction, and TIMI risk score.

C-index comparisons showed equivalence to experts (Δ0.02 for %MVO, Δ0.01 for %Scar), affirming reliability. The full study in JACC: Cardiovascular Imaging underscores AI's role in personalized post-AMI management.

Duke-NUS and NHCS: Pillars of Singapore's Biomedical Innovation

Duke-NUS Medical School, a National University of Singapore-Duke University partnership since 2005, excels in clinician-scientist training, producing leaders like principal investigators Derek Hausenloy and Liang Zhong. NHCS, Asia's largest heart center, hosts the CVS.AI lab, pioneering tools like SENSE for coronary artery disease detection—now trialed across SingHealth polyclinics.

Their synergy exemplifies Singapore's Biomedical Research Council vision, investing S$25 billion by 2030 in AI-health hubs like DAISI at Duke-NUS. This LGE-CMRnet work builds on prior feats, including AI-triage for chest pain and APOLLO imaging platforms.

Real-World Impact on Singapore's Healthcare Landscape

Singapore's ageing population (20% over 65 by 2030) faces rising AMI burdens, with ischaemic heart disease topping mortalities. LGE-CMRnet could integrate into NHCS workflows, aiding risk-tailored therapies like intensified statins for high %Scar patients or closer monitoring for MVO-dominant cases.

Stakeholders praise: NHCS Director Prof. Nicholas Chan notes "AI augments expertise, not replaces it"; Duke-NUS Dean Prof. Dean Ho highlights translational speed. Economically, reducing MACE by 10% could save S$100 million annually in HF admissions.

Challenges, Limitations, and Ethical Considerations

While robust, LGE-CMRnet requires diverse datasets to mitigate ethnic biases—Singapore's multi-ethnic validation helps. Limitations include dependency on LGE quality and generalizability beyond AMI. Regulatory hurdles via Singapore's Health Sciences Authority loom, demanding prospective trials.

  • Pros: Speed, scalability, prognostic equivalence
  • Cons: Training data needs, explainability gaps
  • Risks: Over-reliance, data privacy

Singapore's AI governance framework ensures ethical deployment.

Singapore's AI Vanguard in Cardiovascular Research

Beyond LGE-CMRnet, Duke-NUS/NHCS advances include AI for HFpEF biomarkers and quantum-enhanced imaging. Nationally, A*STAR's PREVENT-AMI predicts events pre-symptomatically. NHCS CVS.AI fosters clinician-AI synergy, training fellows in deep learning.

Collaboration between Duke-NUS Medical School and National Heart Centre Singapore in AI cardiovascular research

Future Outlook: From Bench to Bedside

Prospective RCTs at NHCS will validate clinical utility, potentially FDA/CE-marking by 2028. Integration with wearables for dynamic prognostication looms, aligning with Singapore's Smart Nation 2.0. For higher ed, Duke-NUS expands AI curricula, preparing clinician-scientists.

Actionable insights: Clinicians adopt hybrid AI-manual workflows; researchers pursue multi-modal fusion (CMR+echo+ECG).

green trees near white building during daytime

Photo by David Kubovsky on Unsplash

Career Pathways in Singapore's AI-Cardiology Frontier

Duke-NUS offers PhD fellowships in DAISI; NHCS recruits AI specialists via SingHealth. With 1,000+ biomedical jobs yearly, opportunities abound in research assistantships, postdocs, and faculty roles.

Portrait of Dr. Elena Ramirez

Dr. Elena RamirezView full profile

Contributing Writer

Advancing higher education excellence through expert policy reforms and equity initiatives.

Discussion

Sort by:

Be the first to comment on this article!

You

Please keep comments respectful and on-topic.

New0 comments

Join the conversation!

Add your comments now!

Have your say

Engagement level

Frequently Asked Questions

🧠What is LGE-CMRnet?

LGE-CMRnet is an end-to-end deep learning pipeline developed by Duke-NUS and NHCS for automated segmentation of myocardial scar and microvascular obstruction on late gadolinium enhancement CMR images.

📊How accurate is LGE-CMRnet compared to experts?

It achieves DSC of 0.83 for scar and 0.88 for MVO, with correlations r=0.90/0.98 to experts and non-inferior prognostic C-indexes. PubMed study.

❤️What prognostic value does it offer?

%MVO (HR 1.06, P=0.003) and %Scar (HR 1.05, P=0.001) independently predict MACE in AMI patients over 24 months.

🔬What is the dataset size?

3,874 LGE images from 567 AMI patients at NHCS, including multi-ethnic Singaporeans.

How fast does it process images?

0.05 seconds per image, versus 20-30 minutes manually—ideal for high-volume centers.

🎓Role of Duke-NUS in this research?

Duke-NUS's CVMD programme leads, leveraging DAISI for AI integration in clinician training.

🏥NHCS contributions?

Provides datasets and CVS.AI lab; national hub for advanced cardiac imaging.

⚠️Limitations of LGE-CMRnet?

Relies on LGE quality; needs prospective trials for full validation.

🚀Future applications?

Integration in NHCS workflows, multi-modal AI, wearables for dynamic prognosis.

🇸🇬Impact on Singapore healthcare?

Scales AMI risk assessment amid rising CVD, supporting Smart Nation AI goals.

💼Career opportunities?

Duke-NUS PhDs, NHCS AI roles via AcademicJobs research jobs.