MIT's AI reads an ECG and predicts heart failure worsening 12 months out

You know that feeling when a patient comes in for a routine follow-up, looks stable on paper, gets sent home — and is back in the ICU six months later in acute crisis? Before PULSE-HF, cardiologists had no validated, systematic way to distinguish heart failure patients who will stay stable from those heading toward severe deterioration within a year. Existing AI ECG tools detect current dysfunction — they answer 'does this patient have reduced ejection fraction right now?' not 'will this patient's ejection fraction crash within 12 months?' Meanwhile, the gold-standard for monitoring — echocardiograms — require a trained cardiac sonographer, expensive ultrasound equipment, and hospital infrastructure that rural clinics simply don't have. PULSE-HF takes an ECG input you already collect at every visit and outputs a 12-month risk forecast, with no additional equipment, no specialist, and no...

Your heart generates electrical signals with every beat — an ECG captures those signals as a waveform trace. PULSE-HF feeds that raw waveform into a deep learning model trained on ECG–echocardiogram pairs from three hospital systems: Massachusetts General Hospital, Brigham and Women's Hospital, and the public MIMIC-IV dataset. The model learns the subtle electrical signatures that precede a drop in left ventricular ejection fraction (LVEF) — the percentage of blood the heart pumps per beat — below 40%, which is the threshold for severe heart failure. During training, the team matched each ECG to a future echocardiogram taken within 12 months, so the model learns 'what does the ECG look like when the heart is heading toward failure, even before the failure is visible?' At inference, you feed in an ECG, and the model outputs a probability score: high probability means flag this patient for priority follow-up; low probability means they're likely stable and can reduce visit frequency.

• 12-month LVEF forecast — the only published model that predicts future ejection fraction decline among heart failure patients, not just current dysfunction; you get a forward-looking risk score from a routine ECG you're...
• AUROC 0.87–0.91 across three independent cohorts — validated on Massachusetts General Hospital, Brigham and Women's Hospital, and MIMIC-IV separately, so the performance isn't a single-institution artifact; this is the ...
• Single-lead version matches 12-lead accuracy — one electrode achieves the same AUROC as the full 10-electrode clinical setup, meaning PULSE-HF works on wearable ECG patches, Apple Watch-style single-lead devices, and ru...
• No echocardiogram required at inference — the model inputs only an ECG but forecasts an echocardiogram outcome, which means rural clinics and resource-limited settings without cardiac ultrasound infrastructure can still...
• Trained on messy real-world data — the team explicitly chose not to aggressively filter signal artifacts from restless patients and loose leads, so the model is robust to the imperfect data quality you'll encounter in a...
• Enables bidirectional resource optimization — high-risk patients get earlier, more intensive follow-up; low-risk patients safely reduce visit frequency; both outcomes improve care quality and reduce system-wide cost sim...

Who It Is For

If you build ML pipelines for clinical decision support, work in healthcare AI infrastructure, or are a researcher interested in time-series forecasting on physiological signals — this paper is a clean benchmark for ECG-based prognostic modeling at real-world scale. If you're a cardiologist or health system CTO evaluating AI triage tools, PULSE-HF is the first model in this specific forecasting n...

Worth Exploring?

Yes — this is one of the cleaner pieces of clinical AI research you'll see: specific clinical task, clear metric, multi-site external validation, published in a rigorous journal, with an honest acknowledgment that the prospective test hasn't happened yet. The single-lead result is the most practically significant finding — it's the detail that opens wearable integration and means the model doesn't require hospital-grade equipment to be useful. The honest limitation is that it's retrospective only; until a prospective study confirms the 0.87–0.91 AUROC holds in real clinical deployment (where ...