PULSE-HF: How an MIT Deep Learning Model Reads Your Heartbeat to Predict Heart Failure Decline a Year in Advance

Heart failure is one of those diagnoses that rewrites a patient's entire future. Unlike a broken bone or a treatable infection, it is chronic, progressive, and — as of 2026 — still incurable. The heart muscle weakens over months and years, losing its ability to pump blood efficiently. Fluid slowly accumulates in the lungs, legs, and abdomen. Eventually, for many patients, the trajectory ends in arrhythmia or sudden cardiac arrest. Globally, an estimated 64 million people live with heart failure, and approximately half of them will die within five years of diagnosis [3][5]. The economic toll is staggering: the World Health Organization estimates that cardiovascular diseases collectively claimed 19.8 million lives in 2022 alone, while global expenditure on heart failure treatment reaches an estimated $346 billion annually [4][5].

Against this backdrop, a team of researchers at MIT, Mass General Brigham, and Harvard Medical School has introduced a tool that could fundamentally change how clinicians manage heart failure patients. Their deep learning model, called PULSE-HF — short for "Predict changes in left ventricULar Systolic function from ECGs of patients who have Heart Failure" — does something no other clinical AI tool has done before: it forecasts whether a patient's heart function will dangerously deteriorate within the next twelve months, using nothing more than a standard electrocardiogram [1][2].

The Problem: Detection Is Not Prediction

To understand why PULSE-HF matters, you need to understand the metric that defines heart failure severity: left ventricular ejection fraction, or LVEF. A healthy heart pumps roughly 50 to 70 percent of the blood in its left ventricle with each beat. When that number drops below 40 percent, the patient enters the most severe category of heart failure — heart failure with reduced ejection fraction, or HFrEF. This is the cohort with the highest mortality risk, the most hospitalizations, and the greatest need for aggressive pharmacological and device-based interventions [3].

The clinical challenge is straightforward but profoundly difficult to solve: which patients currently sitting at an LVEF of, say, 45 or 50 percent will cross that critical 40-percent threshold in the coming year? Until now, the answer has largely been educated guesswork. Cardiologists monitor patients through periodic echocardiograms — ultrasound imaging of the heart — but these are expensive, require specialized sonographers, and are typically only available at larger medical centers. More importantly, existing AI tools in cardiology have focused almost exclusively on detection: they can identify patients who already have reduced ejection fraction, but they cannot predict who will develop it [1].

"The biggest thing that distinguishes PULSE-HF from other heart failure ECG methods is instead of detection, it does forecasting," says Tiffany Yau, an MIT PhD student in the lab of Professor Collin Stultz and co-first author of the paper. The study notes that to date, no other methods exist for predicting future LVEF decline among patients with heart failure [1].

How PULSE-HF Works: Behind the Architecture

PULSE-HF's approach is elegantly intuitive, even if the underlying deep learning architecture is anything but simple. The model takes two inputs: a standard 12-lead electrocardiogram recording and the patient's history of prior LVEF measurements. It then outputs a single, clinically actionable prediction — the probability that the patient's LVEF will fall below 40 percent within the next year [1][2].

What makes this particularly significant is the choice of input data. ECGs are among the most ubiquitous diagnostic tools in medicine. A 12-lead ECG involves placing 10 electrodes on the patient's chest and limbs and recording the electrical activity of the heart for about 10 seconds. The test costs a fraction of an echocardiogram, requires no specialized imaging equipment, and can be performed by any trained clinical staff member in virtually any setting — from a major academic medical center to a rural primary care office [1].

flowchart LR
    A["12-Lead ECG\nRecording"] --> C["PULSE-HF\nDeep Learning Model"]
    B["Patient LVEF\nHistory"] --> C
    C --> D{"LVEF < 40%\nWithin 1 Year?"}
    D -->|High Risk| E["Prioritize for\nFollow-up Care"]
    D -->|Low Risk| F["Reduce Hospital\nVisit Frequency"]

The model was developed and retrospectively validated across three distinct patient cohorts drawn from Massachusetts General Hospital and Brigham and Women's Hospital — two of the largest and most respected academic medical centers in the United States — as well as the MIMIC-IV dataset, a publicly available critical care database maintained by the MIT Laboratory for Computational Physiology. This multi-cohort validation is critical: it demonstrates that PULSE-HF's predictions generalize across different patient populations, clinical workflows, and data collection practices [1][2].

Performance: AUROC of 0.87 to 0.91

The researchers measured PULSE-HF's performance using the area under the receiver operating characteristic curve, or AUROC — a standard metric in clinical machine learning that quantifies a model's ability to correctly distinguish between patients who will and will not experience a given outcome. An AUROC of 0.5 represents random guessing; 1.0 represents perfect discrimination.

Across all three patient cohorts, PULSE-HF achieved AUROCs ranging from 0.87 to 0.91 — performance that places it comfortably in the range considered clinically useful for risk stratification [1][2]. To put this in perspective: many widely used clinical prediction tools in cardiology operate in the 0.70–0.80 AUROC range. PULSE-HF's negative predictive values exceeded 97 percent for underlying LVEF worsening of 10 percentage points per year, assuming 80 percent sensitivity [2]. In practical terms, this means that when PULSE-HF says a patient is low-risk, clinicians can be highly confident that the patient's heart function will remain stable.

The Single-Lead Breakthrough

Perhaps the most surprising — and clinically consequential — finding of the PULSE-HF study is that the researchers also built a version of the model that works with single-lead ECGs. A single-lead ECG requires just one electrode placed on the body, as opposed to the ten electrodes needed for a full 12-lead recording. Single-lead recordings can be obtained not only from dedicated medical devices but also from consumer wearables such as the Apple Watch, Samsung Galaxy Watch, and other smartwatch-class devices equipped with ECG sensors [1].

The performance of the single-lead version of PULSE-HF was, according to the researchers, comparable to the 12-lead version [1][2]. This finding has profound implications for healthcare delivery. It means the model could potentially be deployed in rural clinics that lack the equipment and trained staff for full 12-lead ECGs. It could be integrated into telemedicine workflows. And it opens the possibility of continuous or periodic heart failure monitoring through devices that patients already wear on their wrists.

"The model can be deployed in low-resource clinical settings, including doctors' offices in rural areas that don't typically have a cardiac sonographer employed to run ultrasounds on a daily basis," Yau explains [1]. In a world where heart failure disproportionately affects populations with limited access to specialist care — including elderly patients, rural communities, and healthcare systems in low- and middle-income countries — this portability is not a minor technical detail. It is a potential paradigm shift.

The Scale of the Heart Failure Crisis

To fully appreciate PULSE-HF's potential impact, it helps to step back and examine the sheer scale of the heart failure epidemic. The numbers are sobering. According to the Heart Failure Society of America's HF Stats 2025 report, the lifetime risk of developing heart failure has now risen to 24 percent — roughly one in four people [5]. In the United States alone, approximately 6.7 million people currently live with heart failure, a number projected to climb to 8.7 million by 2030 and 11.4 million by 2050. Hospitalizations for heart failure exceeded 1.2 million in the United States in 2021 [5].

The economic burden is equally staggering. Direct medical costs for heart failure in the United States were approximately $32 billion in 2020, with indirect costs adding another $14 billion. Projections suggest that U.S. heart failure costs could surge anywhere from $142 billion to $858 billion by 2050, depending on the modeling assumptions [5]. Globally, annual expenditure on heart failure treatment is estimated at $346 billion, projected to reach approximately $398 billion by 2030 [4][5].

Against this backdrop, any tool that can help clinicians identify high-risk patients earlier and allocate resources more efficiently has enormous potential value. "Understanding how a patient will fare after hospitalization is really important in allocating finite resources," says Teya Bergamaschi, co-first author of the study [1].

The Data Challenge: Why Clinical AI Is Harder Than It Looks

One of the most illuminating aspects of the PULSE-HF story is the candid account its creators give of the painstaking data work required to build a clinical AI model. Despite the elegant simplicity of the concept — feed an ECG into a model, get a prediction out — the actual execution took years of effort, with much of that time spent on the unglamorous but essential task of collecting, processing, and cleaning medical data [1].

The labels that PULSE-HF needs to learn from — future LVEF values — had to be extracted from echocardiogram reports. These reports typically come as PDF files, not clean structured data. Converting PDFs to machine-readable text is notoriously unreliable: line breaks, formatting artifacts, and inconsistent layouts make automated extraction difficult. "When PDFs are converted to TXT files, the text becomes difficult for the model to read," Bergamaschi explains [1].

The ECG data presented its own challenges. Real-world ECGs are often far from the clean signals shown in textbooks. Restless patients produce motion artifacts. Loose electrodes create noise. Electrode placement can vary between technicians. "There are a lot of signal artifacts that need to be cleaned," says Bergamaschi. "It's kind of a never-ending rabbit hole" [1].

The team ultimately made a pragmatic decision about data quality — a decision that may, paradoxically, make PULSE-HF more useful in real clinical settings. Rather than over-engineering the data preprocessing pipeline, they trained the model on slightly noisy data, reasoning that real-world deployment would inevitably involve imperfect recordings. "You have to think about the use case," says Yau. "Is it easiest to have this model that works on data that is slightly messy? Because it probably will be" [1].

The Competitive Landscape: Where AI Meets Cardiology

PULSE-HF enters a rapidly maturing field of AI-powered cardiac diagnostics. Over the past three years, researchers worldwide have demonstrated that deep learning models can extract remarkable amounts of information from the seemingly simple electrical traces of an ECG. These models have shown the ability to detect conditions ranging from atrial fibrillation and myocardial infarction to hypertrophic cardiomyopathy and even conditions seemingly unrelated to the heart, such as thyroid dysfunction and low blood potassium.

In the specific domain of heart failure, several notable AI-ECG models have emerged. A 2025 study published in Frontiers demonstrated a transformer-based AI-ECG model that could identify patients with mildly reduced ejection fraction (HFmrEF) with an AUC of 0.824. A separate study, published in JAMA in 2025, demonstrated a noise-adapted AI-ECG model that could estimate heart failure risk using only lead I ECGs across diverse multinational cohorts — suggesting the potential for integration with wearable devices. Meanwhile, a 2024 study in PLOS ONE used machine learning with circadian ECG features to classify heart failure patients by ejection fraction category, achieving AUROCs as high as 0.99.

What distinguishes PULSE-HF from all of these models is its fundamentally different objective. Every other model in this landscape performs detection or classification — identifying patients who currently have reduced ejection fraction. PULSE-HF is the first to perform forecasting: predicting who will develop reduced ejection fraction in the future. This distinction is not merely academic. Detection tells a clinician what is already happening; forecasting tells them what is about to happen, providing a window of opportunity for early intervention that could prevent or slow the progression of disease [1][2].

Clinical Implications: From Research to Bedside

If PULSE-HF successfully transitions from retrospective validation to prospective clinical use, its impact could manifest in several concrete ways. First, it could enable a more intelligent triage system for heart failure patients. Rather than scheduling all patients for the same frequency of follow-up visits and echocardiograms, clinicians could use PULSE-HF to identify the patients most likely to deteriorate and concentrate monitoring resources on them. "If PULSE-HF predicts that a patient's ejection fraction is likely to worsen within a year, the clinician can prioritize the patient for follow-up," the researchers explain. "Subsequently, lower-risk patients can reduce their number of hospital visits" [1].

Second, PULSE-HF could influence the timing of therapeutic interventions. Heart failure treatment guidelines include a range of medications — including beta-blockers, ACE inhibitors, ARBs, ARNI (angiotensin receptor-neprilysin inhibitor), SGLT2 inhibitors, and mineralocorticoid receptor antagonists — whose efficacy depends partly on when they are initiated. Identifying patients headed toward HFrEF before they cross the 40-percent threshold could allow earlier initiation of guideline-directed medical therapy, potentially slowing or reversing the decline in cardiac function.

Third, there are significant healthcare cost implications. Heart failure hospitalizations are among the most expensive in medicine, with each admission costing an average of $15,000–$25,000 in the United States. If PULSE-HF can help prevent even a fraction of these hospitalizations by enabling earlier outpatient intervention, the savings could be substantial — both for healthcare systems and for patients who avoid the physical and emotional toll of hospitalization.

The Road Ahead: Prospective Validation and Regulatory Pathway

The researchers are clear-eyed about what PULSE-HF has demonstrated and what remains to be proven. The model has been validated retrospectively — meaning it was tested on historical patient data where the outcomes are already known. The next critical step is prospective validation: deploying PULSE-HF in a clinical setting and testing its predictions on patients whose future LVEF values are not yet known [1].

Prospective validation is the gold standard for clinical AI tools, and the path from a successful retrospective study to a clinically deployed tool is often long and expensive. The model will need to be tested across more diverse patient populations — the current validation cohorts, while multi-institutional, are drawn from the Greater Boston area and may not fully represent the demographic, genetic, and comorbidity diversity of the broader population. FDA clearance, whether through the 510(k), De Novo, or PMA pathway, would likely be required before PULSE-HF could be used for clinical decision-making in the United States.

There is also the question of integration. Even the most accurate AI model is only useful if it fits into clinical workflows. Hospitals run on electronic health record systems — primarily Epic and Cerner in the United States — and any new AI tool must integrate seamlessly with these platforms. The model will need to produce predictions in a format that is actionable for clinicians, ideally embedded within the same interfaces they use to review ECG results.

The Human Story Behind the Science

Behind the technical achievement of PULSE-HF is a deeply human story about two PhD students who spent years working through the messy, unglamorous realities of clinical AI research. Bergamaschi and Yau acknowledge that the project took longer than expected — years longer, in fact — and involved countless iterations as they worked through data quality issues, model architecture decisions, and the inherent unpredictability of biological data [1].

I think things are rewarding partially because they're challenging. A friend said to me, 'If you think you will find your calling after graduation, if your calling is truly calling, it will be there in the one additional year it takes you to graduate.'

Yau, who joined Stultz's lab after a personal health event that made her realize the importance of machine learning in healthcare, frames the work in terms that transcend academic metrics. "There's too much suffering in the world," she says. "Anything that tries to ease suffering is something that I would consider a valuable use of my time" [1].

The paper was published in Lancet eClinicalMedicine in February 2026 and was conducted in the lab of Collin Stultz, the Nina T. and Robert H. Rubin Professor at MIT, whose research group is affiliated with the MIT Abdul Latif Jameel Clinic for Machine Learning in Health (Jameel Clinic). The collaboration with Mass General Brigham and Harvard Medical School reflects the increasingly interdisciplinary nature of clinical AI research, where computer scientists, cardiologists, and data engineers must work together across institutional boundaries [1][2].

Looking Forward

PULSE-HF arrives at a moment when the cardiology community is grappling simultaneously with a growing patient population, constrained healthcare resources, and an explosion of new AI capabilities. The model does not claim to solve the heart failure crisis. It is, by design, a single tool that addresses a single clinical question: will this patient get worse in the next year? But it answers that question with unprecedented accuracy, using data that is already routinely collected in virtually every clinical setting on the planet.

If the prospective validation studies confirm what the retrospective data suggests, PULSE-HF could become one of the first AI tools to move the cardiology paradigm from reactive to predictive — from asking "how is this patient doing now?" to asking "how will this patient be doing a year from now?" For the 64 million people worldwide living with heart failure, that shift from detection to forecasting could mean the difference between a crisis managed and a crisis prevented.