Deep Learning-Based Approach to Emergency Department Chest Pain Risk Stratification by ECG

Shyon Parsa, J. Weston Hughes, Guson Kang, James Zou, Albert J. Rogers, Sanjiv M. Narayan, Euan A. Ashley, Marco PerezEMJ 2025 · DOI

TL;DR

Researchers at Stanford have developed a deep learning model that analyzes raw ECG waveforms to predict short-term mortality in emergency department patients presenting with chest pain. The resulting "ECG-Risk" score significantly outperforms traditional clinical models and, crucially, provides prognostic value that is independent of troponin levels, offering a potential new "rule-out" tool for low-risk patients.

What Problem Does This Paper Address?

Chest pain is one of the most common reasons people visit the Emergency Department (ED), but it presents a high-stakes triage challenge. Clinicians must quickly distinguish between benign causes and life-threatening cardiac events.

Currently, triage relies on clinical variables (like age and blood pressure) and troponin assays—a blood test that detects proteins released when the heart muscle is damaged. However, troponin levels often take time to rise after the onset of symptoms, meaning early tests can be deceptively normal. While AI has been used to detect specific arrhythmias or chronic heart disease from ECGs, its application for acute, near-term risk stratification in the ED has been less explored. The authors sought to fill this gap by creating an automated, waveform-based risk score available immediately upon arrival.

Methods

The researchers employed a Convolutional Neural Network (CNN) to analyze raw waveform data from standard 12-lead ECGs.

  • Beyond Human Interpretation: Unlike traditional ECG analysis, which focuses on discrete intervals (like the PR interval or QRS duration), the CNN processes the entire raw signal. This allows the model to detect subtle, non-linear patterns in cardiac electrical activity that are invisible to the human eye.
  • Integration: The model generates a continuous "ECG-Risk" score. To test its utility, the researchers integrated this score into a multivariable logistic regression model that included standard clinical factors: age, sex, race, blood pressure, smoking status, and—most importantly—troponin levels.
  • Evaluation: Model performance was measured using the Area Under the Receiver-Operating Characteristic curve (AUC), with improvements validated via DeLong testing and reclassification indices.

Datasets

The study utilized a massive retrospective dataset from Stanford University:

  • Initial Pool: Over 1.5 million ECGs recorded between 2005 and 2019.
  • Final Cohort: 22,590 adults who presented specifically with chest pain.
  • Exclusions: The team filtered out low-quality tracings, patients with missing troponin data, and non-chest pain presentations to ensure the model was tested on a clinically relevant population.
  • Ground Truth: Outcomes (mortality and revascularization) were tracked through Electronic Health Records (EHR) and the Social Security Death Index (SSDI).

Key Insights and Ideas

  • Independence from Biomarkers: Perhaps the most significant "aha moment" was that the ECG-Risk score did not correlate with troponin levels ($p > 0.05$). This suggests the AI is capturing physiological signals of cardiac distress that blood tests miss, making it a complementary tool rather than a redundant one.
  • A Reliable "Rule-Out" Threshold: The study identified a specific threshold (ECG-Risk < -1.25) below which zero deaths occurred within 30 days. In a crowded ED, a tool that can safely identify "zero-risk" patients could drastically reduce unnecessary admissions and resource strain.
  • Short-term vs. Long-term: While designed for 30-day risk, the score remained a strong predictor for 365-day mortality, suggesting it captures fundamental markers of cardiovascular health.

Key Results

  • Predictive Power: The ECG-Risk score was strongly associated with 30-day mortality, with an adjusted Odds Ratio (OR) of 3.47 (95% CI: 2.66–4.54).
  • Improved Discrimination: Adding the ECG-Risk score to the baseline clinical model (which included troponin) increased the AUC from 0.75 to 0.80 ($p < 0.05$).
  • Secondary Endpoints: The score also successfully predicted a composite of 30-day cardiovascular mortality or revascularization (OR: 2.43) and 365-day mortality (OR: 2.03).
  • Cohort Demographics: The study population had a median age of 55, was 52% female, and had an overall 30-day mortality rate of 1.4%.

Limitations and Future Work

As a retrospective study, these results are based on historical data. The authors acknowledge that prospective validation—testing the model in real-time in an active ED—is the necessary next step before clinical adoption.

Future research will likely focus on how to best integrate this score into existing clinical workflows. For instance, could an automated "high risk" flag on an ECG machine prompt earlier intervention, or could a "low risk" score allow for faster discharge? Additionally, exploring how the model performs across different hospital systems and more diverse populations will be critical for ensuring the tool's generalizability.

Summary updated

BibTeX
@inproceedings{parsa2025deep,
  author = {Shyon Parsa and J. Weston Hughes and Guson Kang and James Zou and Albert J. Rogers and Sanjiv M. Narayan and Euan A. Ashley and Marco Perez},
  title = {{Deep Learning-Based Approach to Emergency Department Chest Pain Risk Stratification by ECG}},
  booktitle = {EMJ},
  year = {2025},
  doi = {10.33590/emjcardiol/kkid8224},
}
Figure