Artificial intelligence-based ECG reconstruction error as a continuous predictor of all-cause mortality: a multi-cohort retrospective validation study

Technical Summary: Artificial Intelligence-based ECG Reconstruction Error as a Continuous Predictor of All-Cause Mortality

Problem Statement
Current artificial intelligence (AI) models for electrocardiogram (ECG) risk stratification predominantly rely on supervised learning, where risk is defined by the discrepancy between a model's prediction and an external target (e.g., chronological age, sex, or specific clinical outcomes). This approach couples the risk score to the choice of the external label rather than the intrinsic cardiac signal, potentially limiting generalisability across diverse populations and clinical settings. Furthermore, these models may learn cohort-specific or outcome-specific artefacts. There is a need for a risk metric that reflects the inherent abnormality of the ECG signal itself, independent of external labels or demographic data.

Methodology
The authors propose a self-supervised learning framework using a masked autoencoder (MAE) to generate a continuous risk score based on ECG reconstruction error.

• Data Sources: The model was pre-trained on 85% of the CODE dataset (n = 7,212,109 ECGs) from the Telehealth Network of Minas Gerais, Brazil. Internal validation was performed on the remaining 15% (CODE-15%, n = 229,499). External validation was conducted across four independent cohorts representing diverse settings: MIMIC-IV-ECG (critical care, US), HEEDB (hospital, US), CHRIS (population-based, Italy), and Innsbruck (cardiology centre, Austria). A binary threshold analysis was also applied to the UK Biobank.
• Model Architecture: A Vision Transformer (ViT) adapted for ECG signals (ViT-Tiny) was employed.
  • Preprocessing: ECGs were resampled to 100 Hz, bandpass filtered (0.5–40 Hz), and lead-wise normalised to zero mean and unit standard deviation to harmonise amplitude and spectral distributions across cohorts.
  • Masking Strategy: During training, 80% of the ECG patches (36 total patches across 12 leads) were randomly masked. The model learned to reconstruct the original waveform from the visible segments using a mean squared error loss applied only to masked regions.
  • Inference: For risk scoring, each ECG was processed five times with different random crops and masking patterns. The final reconstruction error was the mean absolute error (MAE) across all leads, time points, and five inference runs. Crucially, the inference masking ensured no timepoint was completely unobserved (at least one lead was visible at every moment), preserving R-peak location information.
• Risk Score Derivation: The reconstruction error was normalised using the mean and standard deviation of the CODE-15% cohort (z-score). A binary high-risk threshold was defined as a normalised error >1.0 (equivalent to raw error > mean + 1 SD of CODE-15%).
• Statistical Analysis: Associations between reconstruction error and all-cause mortality were assessed using Cox proportional hazards models adjusted for age and sex. Discrimination was evaluated using Harrell's C-index and time-dependent ROC curves.

Key Findings

• Continuous Predictor: In all six cohorts, a higher reconstruction error was significantly associated with increased all-cause mortality (all p<0.001). The hazard ratios (HR) for a 1-SD increase in normalised error ranged from 1.23 (Innsbruck) to 1.41 (HEEDB).
• Short-Term vs. Long-Term: The prognostic value was strongest for short-term mortality. In the HEEDB cohort, the HR for 30-day mortality was 2.00, decreasing to 1.46 for 5-year mortality. Time-dependent AUCs followed a similar trend, being highest for 30-day predictions.
• Binary Threshold: A binary high-risk classification (error >1 SD above CODE-15% mean) identified a high-risk group with significantly elevated mortality across all cohorts, including the UK Biobank (HR 1.27, p=0.004).
• Model Training Dynamics: A counter-intuitive finding was that optimal prognostic performance was achieved by a model with intermediate reconstruction quality. The best risk stratification occurred at 2,600 training steps (less than one epoch). Models trained for longer (up to 100 epochs) or larger models (ViT-Base) achieved better signal reconstruction fidelity but yielded lower hazard ratios for mortality prediction.
• Robustness and Specificity:
  • The signal remained significant after adjusting for standard machine-read ECG parameters (e.g., QTc, QRS duration) and a broad panel of clinical comorbidities.
  • Subgroup analyses in MIMIC-IV-ECG showed the signal persisted in patients without premature ventricular contractions (PVCs) and those with good signal quality.
  • The signal was less discriminatory in patients with conduction defects, likely because these defects inherently alter QRS morphology, inflating reconstruction error regardless of mortality risk.
• Signal Localization: The reconstruction error was distributed across multiple leads, with leads I, II, aVR, V1, V4, V5, and V6 showing the largest differences between high- and low-risk groups. At the waveform level, the QRS complex and T wave contributed most to the risk signal, while the TP interval did not show elevated error in high-risk patients.

Significance and Claims
The paper claims that reconstruction error from a self-supervised MAE provides a generalisable, continuous, and label-free predictor of all-cause mortality.

• Intrinsic Risk: Unlike supervised approaches (e.g., ECG-age), this metric reflects the model's uncertainty about the cardiac electrical field itself, measuring how much a recording deviates from learned normal patterns without relying on external ground truth.
• Generalisability: The risk score, derived from a single threshold in a Brazilian primary care dataset, successfully stratified risk in critical care, hospital, and population-based cohorts across three continents, despite differences in acquisition hardware and patient demographics.
• Clinical Utility: The authors suggest this score could be computed automatically at the point of ECG acquisition to assist in early identification of deterioration in high-acuity settings or for opportunistic screening in primary care.
• Limitations: The study acknowledges its retrospective nature, the lack of cause-specific mortality data in some cohorts, and the need for prospective validation to establish clinical utility and optimal operating points for specific interventions. The binary threshold derived from CODE-15% may require recalibration for populations with vastly different baseline mortality rates.

The authors conclude that this approach successfully moves beyond the constraints of traditional, label-dependent algorithms by defining risk through the inherent abnormality of the ECG signal.