A Multimodal and Explainable Machine Learning Approach to Diagnosing Multi-Class Ejection Fraction from Electrocardiograms

1. Problem Statement

Clinical Context: Left Ventricular Ejection Fraction (LVEF) is the gold standard for diagnosing heart failure and stratifying risk for sudden cardiac death. However, current assessment relies on echocardiography or cardiac MRI, which are resource-intensive, costly, time-consuming, and often inaccessible in primary care or low-resource settings.
The Gap: While the 12-lead Electrocardiogram (ECG) is ubiquitous and low-cost, it has historically been unable to provide LVEF assessments because the relevant signals are too subtle for human interpretation.
Limitations of Existing AI: Previous AI approaches have largely focused on:

• Binary classification: Distinguishing only between "normal" and "reduced" EF (often using a single cutoff like <35%).
• Unimodal data: Relying solely on ECG waveforms or solely on Electronic Health Record (EHR) data, rather than integrating both.
• Black-box models: Using deep learning architectures that lack interpretability and impose high computational barriers for clinical deployment.
• Lack of granularity: Failing to distinguish between mild, moderate, and severe systolic dysfunction, which are critical for therapeutic decision-making.

Objective: To develop a multimodal, explainable, and computationally efficient machine learning framework that classifies LVEF into four clinically distinct strata using standard 12-lead ECGs and structured EHR data.

---

2. Methodology

A. Data Cohort

• Source: Retrospective data from Hartford HealthCare (HHC), a large network in Connecticut.
• Inclusion: Adult outpatients with a standard 12-lead ECG and a physician-reviewed echocardiogram within a 14-day window.
• Development Cohort: 36,784 ECG–echocardiogram pairs from 30,952 unique patients.
  • Class Distribution: Normal (≥50%): 88.35%; Mildly Reduced (40–50%): 5.93%; Moderately Reduced (30–40%): 3.47%; Severely Reduced (<30%): 2.26%.
• Temporal Validation Cohort: 19,966 ECGs from a subsequent period (Jan 2024 – July 2025) to test generalizability over time.
• Exclusion: ECGs flagged as "poor quality" or "artifact."

B. Feature Engineering

The study utilizes a multimodal approach combining two distinct data types:

1. ECG Time-Series Features (Structured):
   • Preprocessing: 0.5-Hz high-pass Butterworth filtering, powerline interference removal, and z-score standardization.
   • Clinical Features: 127 manually engineered features capturing RR-interval statistics, heart rate metrics, and PQRST morphology (using established ECG feature sets).
   • Automated Features: 777 time-series descriptors generated via the tsfresh library, including statistical moments, stationarity, autocorrelation, information-theoretic measures, and frequency-domain characteristics (e.g., spectral entropy).
   • Total ECG Features: ~900 features per lead, concatenated across all 12 leads.
2. EHR Structured Variables:
   • Extracted from the 6 months preceding the echocardiogram.
   • Includes demographics, body measurements, medication history, and comorbidities (ICD-10 groupings).
   • Selection: Retained the 50 most frequent diagnosis groups and 50 most frequent medication groups to reduce sparsity.
   • Leakage Prevention: Explicitly excluded LVEF-proximal codes (e.g., specific heart failure codes) to prevent label leakage.

C. Model Architecture

• Algorithm: XGBoost (Extreme Gradient Boosting).
• Rationale: Chosen over deep learning for computational efficiency, ease of deployment in hospital systems (CPU-compatible), and superior performance on tabular data.
• Task: Multi-class classification (4 classes) using a One-vs-Rest strategy.
• Training: 80% training, 10% validation, 10% held-out test (patient-level split to prevent data leakage).
• Hyperparameters: Learning rate 0.08, max depth 7, min child weight 5, objective multi:softprob.

D. Explainability

• Method: SHAP (SHapley Additive exPlanations).
• Implementation: Computed class-specific SHAP values on the held-out test set to identify which features drive predictions for each LVEF stratum.
• Stability Analysis: Performed bootstrap resampling (B=20) to verify that top feature rankings remained consistent across samples.

---

3. Key Results

A. Performance Metrics (Internal Test Set)

The Multimodal (ECG + EHR) model outperformed unimodal baselines (ECG-only and EHR-only) across all four classes.

• AUROC (One-vs-Rest):
  • Severe (<30%): 0.95 (Multimodal) vs. 0.94 (ECG-only) vs. 0.87 (EHR-only).
  • Moderate (30–40%): 0.92 (Multimodal) vs. 0.88 (ECG-only) vs. 0.86 (EHR-only). Largest gain observed here.
  • Mild (40–50%): 0.82 (Multimodal) vs. 0.80 (ECG-only) vs. 0.76 (EHR-only).
  • Normal (≥50%): 0.91 (Multimodal) vs. 0.89 (ECG-only) vs. 0.84 (EHR-only).

B. Temporal External Validation

The model maintained robust performance on the future temporal cohort (Jan 2024–July 2025), demonstrating stability over time.

• AUROC: 0.93 (Severe), 0.89 (Moderate), 0.81 (Mild), 0.89 (Normal).
• The multimodal approach consistently provided incremental gains over ECG-only models, confirming the value of integrating routine EHR data.

C. Class Imbalance and Thresholds

• Due to the high prevalence of "Normal" LVEF (88%), threshold-dependent metrics (F1, Sensitivity) varied.
• Normal Class: Achieved high F1 (0.95) and sensitivity (0.99) but low specificity (0.26) at the F1-maximizing threshold, as the model is permissive in classifying the dominant class.
• Reduced Classes: Showed modest F1 scores (0.27–0.35) but high specificity (0.82–0.95), indicating the model is effective at ruling out reduced EF in healthy patients but requires careful thresholding for screening purposes.

D. Feature Attribution (Explainability)

SHAP analysis identified consistent, clinically plausible drivers:

• ECG Features: Lead I and V6 QR-interval amplitude statistics, spectral entropy (frequency domain complexity), and specific voltage/morphology descriptors in lateral leads.
• EHR Features: History of cardiomyopathy (I42.9), ischemic cardiomyopathy (I25.5), systolic blood pressure, and sex.
• Stability: Top 10 features appeared in 95–100% of bootstrap resamples, confirming robustness.

---

4. Key Contributions

1. Multi-Class Granularity: Moves beyond binary "normal vs. reduced" classification to a clinically actionable 4-tier stratification (Normal, Mild, Moderate, Severe), aligning with therapeutic guidelines.
2. Multimodal Fusion: Demonstrates that combining engineered ECG time-series features with structured EHR data yields superior performance compared to using either modality alone, particularly for moderate and mild reduction cases.
3. Explainability & Trust: Utilizes SHAP to map model decisions back to interpretable clinical variables (e.g., specific ECG amplitudes and ICD-10 codes), addressing the "black box" concern of deep learning.
4. Deployment Feasibility: Employs an XGBoost-based tabular pipeline that is computationally efficient (training time ~3.3 minutes on a single GPU), open-source, and compatible with standard hospital IT infrastructure, unlike resource-heavy deep learning models.
5. Temporal Validation: Validates the model on a distinct future cohort, proving that the learned signals are stable over time and not merely artifacts of a specific training period.

---

5. Significance and Clinical Implications

• Screening and Triage: The model serves as a practical decision-support tool for primary care. It can identify patients with high-risk reduced LVEF who might otherwise go unrecognized, prioritizing them for confirmatory echocardiography.
• Resource Optimization: In settings with limited access to echocardiography, this AI tool can reduce unnecessary imaging for low-risk patients while flagging high-risk individuals for urgent care.
• Global Health: The reliance on ubiquitous ECGs and standard EHR data makes this approach scalable to low- and middle-income countries where advanced imaging is scarce.
• Future Work: The authors note that while discrimination is high, prospective clinical trials are needed to determine how these predictions influence referral patterns and patient outcomes in real-world workflows.

Conclusion: This study presents a robust, explainable, and deployable AI framework that unlocks latent diagnostic information in routine ECGs, offering a scalable solution for heart failure risk stratification in resource-constrained environments.