Artificial intelligence-enabled electrocardiographic detection of acute occlusive myocardial infarction

Technical Summary: Artificial Intelligence-Enabled Electrocardiographic Detection of Acute Occlusive Myocardial Infarction

Problem Statement
Acute occluded myocardial infarction (OMI), defined by complete or near-complete coronary occlusion causing transmural ischemia, requires timely revascularization to prevent irreversible injury. While coronary angiography (CAG) is the definitive diagnostic tool, it is invasive and resource-intensive. Current clinical pathways rely heavily on ST-elevation myocardial infarction (STEMI) criteria to trigger emergent CAG; however, only 65–75% of patients with angiographically confirmed OMI present with diagnostic ST-elevations. The remaining 25–35% (often classified as NSTEMI) possess acutely occluded arteries but lack specific ECG markers, leading to delayed recognition and treatment. Furthermore, visual interpretation of subtle ECG patterns in non-ST-elevation cases is prone to inter- and intra-observer variability. Existing deep learning (DL) models for myocardial infarction detection have largely been trained on small, publicly available datasets lacking systematic angiographic confirmation, potentially introducing spectrum bias and failing to capture the diagnostic continuum seen in real-world emergency settings.

Methodology
This retrospective case-control study utilized a large-scale, registry-linked dataset to develop and validate a foundation model for OMI detection.

• Data Collection and Curation: The study analyzed 17,165 standard 12-lead ECGs from 11,338 patients recorded between 2016 and 2022 at Akershus University Hospital and affiliated ambulance services. ECGs were linked to the nationwide Norwegian Registry of Invasive Cardiology (NORIC) to obtain angiographic outcomes.

  • OMI Definition: Cases were labeled as OMI if they involved an acute culprit lesion with Thrombolysis In Myocardial Infarction (TIMI) flow grade 0–2 treated with percutaneous coronary intervention (PCI).
  • Control Groups: The control population included patients without PCI, those with normal angiographic findings, and those with non-occlusive lesions (TIMI flow >2) or old lesions. To ensure robustness, the control group was stratified into "healthy controls," "Angio <24 hours" (patients undergoing angiography for non-urgent reasons), and "urgent clinical evaluation" (patients requiring close monitoring but not immediate reperfusion).
  • Exclusion Criteria: Elective procedures, secondary interventions (unless specific criteria were met), and ECGs recorded outside the 0–72 hour window relative to the procedure were excluded. Patients with missing troponin data or troponin levels ≥14 ng/L in the control group were also excluded to minimize false negatives.

• Model Architecture and Training:

  • Foundation Model: The study employed the ECGFounder model, a pretrained deep learning architecture trained on over 10 million ECG recordings. This was fine-tuned on the study dataset.
  • Preprocessing: Raw digital waveforms were resampled to 500 Hz, truncated to 10 seconds, and processed with baseline wander removal (0.5 Hz high-pass), notch filtering (50 Hz), and min-max scaling.
  • Training Strategy: The model utilized a supervised contrastive pre-training phase followed by ordinal classification. This approach treated the four clinical outcome groups (Emergency/OMI, Urgent, Angio <24h, Healthy) as an ordered hierarchy of clinical urgency rather than independent categories.
  • Loss Functions: To address severe class imbalance, the study implemented asymmetric focal loss for binary classification and a focal ordinal loss for the multi-class task, emphasizing hard-to-classify samples.
  • Validation: The dataset was split into an 80% development set (used for 5-fold cross-validation) and a 20% held-out test set, with patient-level separation to prevent data leakage.

• Evaluation Metrics: Performance was assessed using the Area Under the Receiver Operating Characteristic Curve (AUROC) and the Area Under the Precision-Recall Curve (AUPRC). The model was also benchmarked against current clinical practice using "door-to-procedure" time (<30 minutes) as a pragmatic surrogate for OMI recognition.

Key Results

• Overall Performance: The fine-tuned ECGFounder model achieved an AUROC of 0.93 (95% CI, 0.90–0.95) and an AUPRC of 0.72 (95% CI, 0.65–0.79) for distinguishing OMI from non-OMI cases and controls. This performance significantly exceeded the prevalence-based no-skill baseline (AUPRC ~0.10).
• ST-Elevation vs. Non-ST-Elevation:
  • STEMI (with ST-elevation): The model demonstrated exceptional performance with an AUROC of 0.98 and AUPRC of 0.82.
  • NSTEMI (without ST-elevation): Performance was moderate but clinically relevant, with an AUROC of 0.84 and AUPRC of 0.20. The lower AUPRC reflects the inherent difficulty of detecting non-ST-elevation occlusions in an imbalanced dataset.
• TIMI Flow Stratification: The model maintained strong discrimination across TIMI flow grades (0, 0–1, 0–2), with AUROCs ranging from 0.93 to 0.95 for all OMI cases. Performance improved when restricting the non-ST-elevation subgroup to patients with troponin levels ≥500 ng/L, suggesting the model captures hemodynamically relevant ischemia.
• Comparison with Clinical Practice: When compared to the "door-to-procedure <30 minutes" benchmark:
  • For STEMI cases, the model identified 93% of cases (vs. 83% by clinical timing).
  • For NSTEMI cases, the model identified 43% of cases (vs. 11% by clinical timing), highlighting its potential to detect high-risk occlusions currently missed by standard triage.
• Demographic and Biomarker Correlations:
  • The model showed slightly higher performance in males than females, potentially linked to higher median troponin levels in males.
  • Predicted probabilities correlated positively with peak troponin levels and inversely with TIMI flow grades, indicating the model captures the severity of myocardial injury and flow reduction.
  • Misclassifications were more frequent in Left Circumflex (LCx) occlusions, a known diagnostic challenge on ECG.

Significance and Claims
The authors claim that this study demonstrates the feasibility of using a fine-tuned ECG foundation model to detect acute OMI directly from standard 12-lead ECGs with high discriminative accuracy. Key claims include:

1. Detection of "Hidden" Occlusions: The model significantly outperforms current clinical workflows in identifying OMI in patients without ST-elevation (NSTEMI), a group where diagnostic delays are common.
2. Foundation Model Utility: The study validates the utility of large-scale foundation models (ECGFounder) fine-tuned on real-world, registry-confirmed data, overcoming the limitations of smaller, publicly available datasets.
3. Clinical Integration: The model is designed as a decision-support tool to complement, not replace, clinician judgment. It operates on standard 10-second ECGs without requiring additional technical expertise.
4. Biomarker Correlation: The model's predicted probabilities align with biological markers of injury (troponin) and angiographic severity (TIMI flow), suggesting it learns clinically relevant pathophysiological features rather than spurious artifacts.

The authors remain modest regarding immediate clinical deployment, noting the need for prospective external validation across different healthcare systems and ambulance workflows before implementation. They emphasize that while the model improves detection rates, particularly for NSTEMI, it does not eliminate the need for angiographic confirmation.