Duke-NUS Adapts Advanced AI Model to Predict Neurological Recovery After Cardiac Arrest

Duke-NUS Pioneers AI Adaptation for Cardiac Arrest Recovery Insights

In a significant advancement for emergency medicine, researchers at Duke-NUS Medical School in Singapore have successfully adapted an advanced artificial intelligence (AI) model to predict neurological recovery following out-of-hospital cardiac arrest (OHCA). This breakthrough, detailed in the prestigious npj Digital Medicine journal and highlighted in the 2026 Issue 2 of MEDICUS, Duke-NUS's quarterly magazine, addresses a critical challenge in resource-limited healthcare settings worldwide, including parts of Asia.

Out-of-hospital cardiac arrest remains a leading cause of death globally, with survival rates hovering around 3-4% in Singapore despite world-class emergency services. Among survivors, predicting who will achieve favorable neurological outcomes—defined as Cerebral Performance Category (CPC) scores of 1 or 2 at 30 days or hospital discharge—is notoriously difficult early on. This AI adaptation offers clinicians a powerful tool to guide decisions on intensive care, family counseling, and resource allocation.

The work stems from the Pan-Asian Resuscitation Outcomes Study (PAROS), a multinational registry that has amassed vast data on OHCA across Asia since 2010. Duke-NUS scientists leveraged this to refine a model originally developed in Japan, demonstrating the power of transfer learning—a technique where a pre-trained AI model is fine-tuned on new data with minimal additional training.

The Burden of Cardiac Arrest in Singapore

Singapore faces a rising tide of OHCA cases, driven by its aging population. According to the Singapore Heart Foundation's Out-of-Hospital Cardiac Arrest Data Report (2011-2021), annual incidents reached 3,637 in 2021, up from 3,432 in 2020, with an age-standardized incidence of 45.8 per 100,000 population. Projections suggest further increases, potentially to 132.1 per 100,000 crude rate by 2030 due to demographic shifts.

Key chain-of-survival metrics show progress: bystander cardiopulmonary resuscitation (CPR) rates climbed to 59.4% in 2021, and automated external defibrillator (AED) use hit 9.5%. Yet, overall survival to discharge stands at 3.8%, with Utstein survival (witnessed shockable rhythm, cardiac etiology) at 19.9%. Crucially, 79.3% of survivors achieve good-to-moderate neurological function, underscoring the need for better prognostic tools to optimize post-arrest care.

• Over 3,000 cases annually, 80% outside hospitals.
• Males and older adults (>65) disproportionately affected.
• Return of spontaneous circulation (ROSC) at scene: 8.5%.

These figures highlight why precise, early prediction of neurological recovery is vital for Singapore's healthcare system, which handles ~16,000 OHCA cases in PAROS data alone from 2017-2021.

Duke-NUS Medical School: A Hub for Translational Research

Established in 2005 as a collaboration between Duke University and the National University of Singapore (NUS), Duke-NUS Medical School is Singapore's only graduate-entry medical school, emphasizing clinician-scientist training through its MD-PhD program. With a focus on precision medicine, neuroscience, and population health, it has produced over 1,000 alumni who lead in research and clinical practice.

The school's Pre-hospital & Emergency Research Centre (PERC) and Duke-NUS AI + Medical Sciences Initiative (DAISI), directed by Associate Professor Nan Liu, spearhead innovations in resuscitation science. DAISI bridges AI experts and clinicians to tackle real-world problems, from triage tools to outcome prediction.

PAROS, coordinated in Singapore, exemplifies Duke-NUS's regional leadership, collecting standardized OHCA data from seven Asian sites to benchmark and improve outcomes.

From Japan to Singapore: The Original Model

The base model, developed by Japanese researchers (Nishioka et al.), used Lasso regression on 46,918 OHCA patients, achieving an area under the receiver operating characteristic curve (AUROC) of 0.943. Key predictors included age, initial rhythm, no-flow/low-flow times, bystander AED use, prehospital defibrillation, epinephrine doses, and emergency department (ED) rhythm.

This model excels in high-resource Japan but falters in diverse settings due to data and demographic shifts.

Photo by Bekky Bekks on Unsplash

Transfer Learning: The Key Innovation

Transfer learning (TL) allows AI models to 'transfer' knowledge from large datasets to smaller ones, ideal for low-data environments. Duke-NUS employed Trans-Lasso, initializing with Japanese parameters and fine-tuning on PAROS data.

For Singapore (15,916 patients, 2017-2021), data preprocessing involved missForest imputation and standardization. The TL model boosted AUROC to 0.955 (95% CI: 0.940–0.967) from 0.945, and area under precision-recall curve (AUPRC) to 0.885 from 0.527.

In Vietnam (243 patients), gains were dramatic: AUROC rose to 0.807 (95% CI: 0.626–0.948) from 0.467, proving TL's value in data-scarce areas.

Clinical Performance and Validation

The adapted model distinguishes high- from low-risk patients with ~80% accuracy, per Duke-NUS reports. It retains critical features like shockable rhythm while adapting to local nuances.

Validated on held-out test sets, TL preserved interpretability via Lasso, aiding clinician trust. In Singapore's PAROS cohort, favorable outcomes occurred in 3-4%—rare events where TL shines by improving precision for imbalanced classes.

Implications for Singapore Healthcare

This tool could transform post-OHCA care at institutions like Singapore General Hospital (SGH), where Prof Marcus Ong practices. Early identification enables targeted hypothermia, counseling, and organ donation discussions.

In Singapore's integrated system, integrating such AI into electronic records aligns with the Healthier SG agenda, enhancing equity amid rising cases. Dr Jasmine Ong noted, “AI bridges gaps where resources are stretched.” Assoc Prof Nan Liu emphasized equitable global adoption.Read the full study in npj Digital Medicine.

Global Health Equity and Future Horizons

TL democratizes AI, extending high-resource models to low-resource Asia. Duke-NUS plans prospective validation and expansion to in-hospital arrests.

DAISI's work complements Singapore's AI Singapore initiative, fostering clinician-AI synergy. Future integrations may include real-time ECG analysis for rearrest prediction.

• Potential 20-30% better risk stratification.
• Supports personalized post-resuscitation care.
• Paves way for AI in Singapore's Smart Nation health pillar.

Challenges remain: data privacy, bias mitigation, regulatory approval via Singapore's Health Sciences Authority.

Photo by Milad Fakurian on Unsplash

Stakeholder Perspectives and Broader Impact

Clinicians praise interpretability; ethicists highlight fairness. Patients/families gain prognostic clarity, reducing uncertainty.

For Singapore universities, this exemplifies clinician-scientist synergy, inspiring MD-PhD trainees. Duke-NUS's PAROS leadership positions it as Asia's resuscitation research vanguard.Singapore Heart Foundation OHCA stats.

Path Forward: Integrating AI into Practice

Pilots at SGH/National Heart Centre could deploy via apps. Training via Duke-NUS programs ensures uptake.

As Singapore eyes 2030 OHCA surge, AI-driven precision promises better outcomes, aligning with global Sustainable Development Goals.