AI-ECG LV Dysfunction (Attia 2019) Calculator

AI-ECG Deep Learning Model for Left Ventricular Dysfunction — Explanation and Clinical Context This calculator implements the clinical screening logic described by Attia et al. (Nature Medicine 2019). The original model is a convolutional neural network trained on paired 12-lead ECG and echocardiogram pairs from tens of thousands of patients. The model was trained to detect reduced systolic function defined as left ventricular ejection fraction (LVEF) < 35% using the ECG signal alone.

On an independent test set the investigators reported excellent discrimination (AUC ~0.93) with sensitivity approximately 86.3% and specificity approximately 85.7% for that outcome. The published model returns a continuous probability score; clinical use requires choosing an operating threshold to balance sensitivity and specificity for the intended screening setting. Because deploying the original DNN weights and reproducing inference requires the trained model files and runtime, this calculator instead accepts an AI-ECG probability score as input (0–100%) and uses the published sensitivity and specificity to estimate the positive predictive value (PPV) and negative predictive value (NPV) for a user-supplied prevalence (Bayes' theorem).

PPV and NPV depend strongly on underlying prevalence: in low-prevalence screening populations even a strong model will have modest PPV, while NPV remains high. Thus interpret results in the context of local prevalence and clinical assessment; a positive AI-ECG screen should prompt confirmatory echocardiography or cardiology referral rather than immediate treatment decisions. If you can supply locally estimated sensitivity/specificity (from local validation) these constants should replace the published values used here to improve precision.

If you want an integrated inference (model running on the server), provide the original model artifacts and runtime environment and we can discuss embedding real-time inference — note this requires compliance with licenses, data governance, and compute resources.